Adaptive Sensor Fusion

Online modification of process and measurement covariance, measurement-quality-dependent weighting, adaptive noise estimation, innovation-based adaptation, changing sensor conditions, and robust spacecraft navigation.

4.2.27 Adaptive Sensor Fusion

Spacecraft sensor-fusion systems are often introduced using fixed mathematical models. A process-noise covariance Q is selected, a measurement-noise covariance R is assigned to each sensor, and the estimator uses these values throughout operation.

In an idealized system, this can be written as

\[ \mathbf{x}_{k+1}=f(\mathbf{x}_k,\mathbf{u}_k)+\mathbf{w}_k, \qquad \mathbf{w}_k\sim\mathcal N(\mathbf 0,Q_k) \]

and

\[ \mathbf{z}_k=h(\mathbf{x}_k)+\mathbf{v}_k, \qquad \mathbf{v}_k\sim\mathcal N(\mathbf 0,R_k). \]

A conventional implementation may initially assume

\[ Q_k=Q, \qquad R_k=R. \]

Real spacecraft sensors, however, do not always behave according to fixed statistical properties. Measurement quality can vary with temperature, illumination, geometry, signal strength, target range, vibration, radiation effects, processing conditions, field of view, and sensor operating mode. Similarly, uncertainty in the spacecraft dynamics can change during maneuvers, thruster firings, flexible-body disturbances, momentum-management events, or other operating conditions.

Consequently,

\[ \boxed{ Q_k\ \text{and}\ R_k\ \text{need not remain constant} } \]

and a sensor that is highly reliable under one operating condition may become much less informative under another.

Adaptive sensor fusion allows the navigation estimator to modify how it represents and combines information according to the observed or predicted quality of the sensors and system dynamics.

The central idea is

\[ \boxed{ \text{observe estimation quality} \rightarrow \text{adapt uncertainty model} \rightarrow \text{change fusion behaviour} } \]

Adaptive fusion therefore does not simply mean switching sensors ON and OFF. It includes continuous modification of measurement weighting, online estimation of noise statistics, innovation-based adaptation, dynamic covariance tuning, quality-dependent sensor weighting, and mode-dependent estimator behaviour.

4.2.27.1 Why Adaptive Sensor Fusion Is Needed

Consider two sensors measuring information about the same spacecraft state:

\[ z_1=h_1(x)+v_1, \]

\[ z_2=h_2(x)+v_2. \]

Suppose initially

\[ R_1

The estimator correctly gives sensor 1 greater influence because its measurement uncertainty is smaller.

Later, environmental conditions degrade sensor 1 so that its true uncertainty becomes

\[ R_{1,\text{actual}}\gg R_{1,\text{nominal}}. \]

If the estimator continues using the original value

\[ R_{1,\text{filter}}=R_{1,\text{nominal}}, \]

it still treats the degraded sensor as highly accurate.

The Kalman gain therefore assigns the measurement more influence than it deserves.

This illustrates the fundamental motivation for adaptive fusion:

\[ \boxed{ \text{Changing sensor quality should produce changing estimator confidence} } \]

Without adaptation, the mathematical confidence assigned by the estimator can become disconnected from actual sensor performance.

4.2.27.2 Fixed Versus Adaptive Sensor Fusion

In fixed sensor fusion,

\[ Q_k=Q_0, \qquad R_k=R_0. \]

The estimator assumes that the underlying process and measurement statistics remain sufficiently close to these nominal values.

In adaptive fusion,

\[ Q_k=f_Q(\text{operating condition, residuals, dynamics, health}), \]

and/or

\[ R_k=f_R(\text{measurement quality, innovation, geometry, environment, health}). \]

Therefore, the estimator can respond when its assumptions cease to describe the actual system adequately.

The purpose is not to make Q and R change continuously without justification. Adaptation should occur because measurable evidence indicates that the uncertainty model should change.

4.2.27.3 What Can Be Adapted?

Adaptive sensor fusion is broader than adaptive measurement covariance.

Possible adaptive quantities include:

Thus,

\[ \boxed{ \text{Adaptive fusion} \neq \text{only adaptive }R } \]

although adapting R is one of the most common approaches.

4.2.27.4 Measurement-Quality-Dependent Weighting

Suppose two independent scalar sensors measure the same quantity:

\[ z_1=x+v_1, \qquad z_2=x+v_2. \]

If

\[ v_1\sim\mathcal N(0,R_1), \qquad v_2\sim\mathcal N(0,R_2), \]

the information-weighted fused estimate is

\[ \hat x= \frac{R_1^{-1}z_1+R_2^{-1}z_2} {R_1^{-1}+R_2^{-1}}. \]

The measurement having smaller variance receives greater influence.

If sensor quality varies,

\[ R_i=R_i(t), \]

then the fusion weights automatically change.

For example,

\[ R_1(t)\uparrow \quad\Rightarrow\quad R_1^{-1}(t)\downarrow. \]

Therefore,

\[ \boxed{ \text{larger uncertainty} \Rightarrow \text{smaller measurement influence} } \]

This is one of the simplest interpretations of adaptive sensor fusion.

4.2.27.5 Adaptive Measurement Covariance

For a Kalman-type estimator,

\[ K_k=P_k^-H_k^T \left( H_kP_k^-H_k^T+R_k \right)^{-1}. \]

Therefore R_k directly controls how strongly a measurement affects the state correction.

If

\[ R_k\downarrow, \]

the estimator generally trusts the measurement more.

If

\[ R_k\uparrow, \]

the estimator generally trusts the prediction more.

This gives a smooth mechanism for changing sensor influence without immediately rejecting the sensor.

A degraded sensor can therefore transition through

\[ \text{nominal} \rightarrow \text{reduced weight} \rightarrow \text{strongly down-weighted} \rightarrow \text{rejected}. \]

This is usually more flexible than treating every sensor as simply healthy or failed.

4.2.27.6 Measurement Quality Indicators

Many sensors provide information beyond the measurement itself.

A generic sensor interface may provide

\[ \left\{ z_i,\, t_i,\, H_i,\, Q_i^{\text{sensor}},\, C_i \right\}, \]

where H_i may represent health, Q_i^{\text{sensor}} a quality indicator, and C_i some confidence or tracking metric.

Examples of useful quality information include:

The estimator may use such information to construct

\[ R_i=f(q_i). \]

The essential requirement is that the relationship between the quality indicator and actual measurement uncertainty be characterized.

4.2.27.7 Confidence Is Not Automatically Covariance

Suppose a sensor reports

\[ c_i=0.92. \]

This does not automatically mean

\[ \sigma_i^2=1-0.92. \]

A confidence score may represent classification confidence, tracking probability, feature quality, or an internally defined vendor metric.

Covariance has a specific statistical interpretation:

\[ R_i=E[v_iv_i^T]. \]

Therefore,

\[ \boxed{ \text{sensor confidence score} \neq \text{measurement covariance} } \]

unless a validated relationship between them has been established.

A mapping such as

\[ R_i=f(c_i) \]

should therefore be derived from sensor characterization, simulation, calibration, or experimental data rather than chosen arbitrarily.

4.2.27.8 Innovation-Based Adaptation

The innovation provides one of the most useful indicators of estimator consistency.

For measurement z_k,

\[ \nu_k=z_k-h(\hat x_k^-). \]

The predicted innovation covariance is

\[ S_k=H_kP_k^-H_k^T+R_k. \]

If the filter is statistically well tuned, the observed innovations should be broadly consistent with S_k.

If the innovations repeatedly become larger than predicted, one possibility is that the assumed uncertainty is too small.

Conceptually,

\[ \nu_k\nu_k^T>S_k \]

over a statistically meaningful interval may suggest that the filter is overconfident.

This can motivate adaptation of R, Q, or both.

4.2.27.9 Normalized Innovation Squared

A commonly useful consistency statistic is the Normalized Innovation Squared:

\[ \boxed{ \epsilon_k=\nu_k^TS_k^{-1}\nu_k } \]

For a correctly modelled Gaussian system, \epsilon_k has a known statistical interpretation related to the measurement dimension.

Persistent large values can indicate:

Therefore, NIS is valuable for adaptation, but it does not by itself identify the cause of inconsistency.

4.2.27.10 Innovation Covariance Matching

Instead of examining a single innovation, a window of innovations can be used.

For N samples,

\[ \hat C_\nu = \frac{1}{N} \sum_{j=k-N+1}^{k} \nu_j\nu_j^T. \]

The theoretical innovation covariance is approximately

\[ S_k=H_kP_k^-H_k^T+R_k. \]

An adaptive algorithm can compare

\[ \hat C_\nu \]

with

\[ S_k. \]

If the observed innovation variance is persistently greater than predicted, the assumed uncertainty may be increased.

If it is persistently smaller, the uncertainty model may be conservative.

This family of techniques is commonly called covariance matching.

4.2.27.11 Why a Window Is Useful

A single measurement can be unusually large simply because of random noise.

Immediately changing R after every residual can therefore make the estimator unstable or excessively reactive.

Instead, statistics may be accumulated over a window:

\[ \mathcal W_k= \{ \nu_{k-N+1},\ldots,\nu_k \}. \]

The window provides evidence about persistent behaviour.

However, window size creates a trade-off.

A short window provides faster adaptation but noisier statistical estimates.

A long window provides smoother statistics but slower response to real changes.

Therefore,

\[ \boxed{ \text{adaptation speed} \leftrightarrow \text{statistical reliability} } \]

is an important design trade-off.

4.2.27.12 Innovation Mean Monitoring

Noise variance is not the only quantity that matters.

For an unbiased, correctly modelled estimator,

\[ E[\nu_k]\approx0. \]

A moving innovation mean can be estimated as

\[ \bar\nu_k= \frac{1}{N} \sum_{j=k-N+1}^{k}\nu_j. \]

If

\[ \bar\nu_k\neq0 \]

persistently, the problem may not simply be increased random noise.

It may indicate:

Simply increasing R can hide such a problem rather than solve it.

4.2.27.13 Adaptive Process Noise

Adaptive fusion can also modify process covariance.

The prediction covariance is

\[ P_{k+1}^- = F_kP_k^+F_k^T+Q_k. \]

The matrix Q_k represents uncertainty introduced by dynamics and unmodelled processes.

During quiet operation,

\[ Q_k=Q_{\text{quiet}} \]

may be appropriate.

During stronger or less predictable dynamics,

\[ Q_k=Q_{\text{dynamic}}, \]

where

\[ Q_{\text{dynamic}}>Q_{\text{quiet}} \]

in relevant state directions.

This causes the estimator to acknowledge greater uncertainty in its prediction.

4.2.27.14 Why Process Noise Can Change

A spacecraft dynamics model is never perfect.

Prediction uncertainty can change because of:

Thus a single fixed Q may not adequately represent every operating mode.

4.2.27.15 Mode-Dependent Process Noise

One practical strategy is to define different covariance models for different operating modes.

For example,

\[ Q_k= \begin{cases} Q_{\text{coast}}, & \text{coasting},\\ Q_{\text{maneuver}}, & \text{maneuvering},\\ Q_{\text{settling}}, & \text{post-maneuver settling}. \end{cases} \]

This is adaptive behaviour even though the matrices themselves may have been determined offline.

The adaptation occurs because the estimator selects the uncertainty model appropriate to the current physical regime.

4.2.27.16 Continuous Versus Discrete Adaptation

Adaptive fusion can operate continuously:

\[ R_k=f(q_k) \]

or through discrete levels:

\[ R_k= \begin{cases} R_{\text{nominal}}, & q_k>q_H,\\ R_{\text{degraded}}, & q_L

Continuous adaptation produces smooth changes in sensor influence.

Discrete adaptation is often easier to understand, test, and verify.

A practical system may combine both.

4.2.27.17 Covariance Inflation

A simple adaptive response to degraded sensor quality is

\[ R_i'=\alpha_iR_i, \qquad \alpha_i\ge1. \]

For nominal conditions,

\[ \alpha_i=1. \]

As confidence decreases,

\[ \alpha_i>1. \]

For severe degradation,

\[ \alpha_i\gg1. \]

As R_i' becomes large, the measurement naturally contributes less to the update.

Covariance inflation therefore provides a continuous bridge between full trust and practical exclusion.

4.2.27.18 Direction-Dependent Adaptation

Sensor quality does not necessarily degrade equally in every measurement direction.

Suppose

\[ R= \begin{bmatrix} \sigma_x^2&0&0\\ 0&\sigma_y^2&0\\ 0&0&\sigma_z^2 \end{bmatrix}. \]

If only one direction becomes poorly observed,

\[ \sigma_z^2\uparrow \]

while

\[ \sigma_x^2,\sigma_y^2 \]

remain approximately unchanged.

Increasing the entire covariance matrix uniformly would unnecessarily discard good information.

Therefore adaptive fusion should, where justified, preserve the directional structure of uncertainty.

4.2.27.19 Geometry-Dependent Adaptation

Measurement quality often depends strongly on geometry.

The information contribution of a linearized measurement is related to

\[ H^TR^{-1}H. \]

Even if intrinsic sensor noise remains unchanged, poor geometry can make certain state directions weakly observable.

Consequently, adaptive fusion may consider both

\[ R_i \]

and

\[ H_i. \]

This is particularly important when the relative orientation between the sensor, observed object, reference direction, or measurement geometry changes substantially during operation.

4.2.27.20 Range-Dependent Measurement Noise

Some sensors have measurement accuracy that changes with range.

Instead of assuming

\[ R=R_0, \]

a model may use

\[ R=R(r). \]

For example, an optical measurement may have different angular or position sensitivity as apparent target size changes, while an active ranging system may have quality dependent on received signal characteristics.

Therefore, the estimator can modify measurement covariance according to operating range when supported by sensor characterization.

4.2.27.21 Optical Measurement Adaptation

Optical navigation quality can depend on:

An image-processing algorithm may provide a quality quantity

\[ q_{\text{img}}. \]

The estimator may then use

\[ R_{\text{camera}} = f(q_{\text{img}},r,\text{geometry},\ldots). \]

The important point is that the estimator should not treat every successful image-processing output as having identical accuracy.

4.2.27.22 GNSS Measurement Adaptation

GNSS measurement quality can vary with factors such as:

Therefore,

\[ R_{\text{GNSS}} = f(\text{geometry},\text{signal quality},\text{receiver status},\ldots). \]

A valid GNSS solution does not necessarily imply constant measurement accuracy.

4.2.27.23 Star-Tracker Adaptation

A star tracker may provide highly accurate absolute attitude information, but its quality and availability can vary.

Factors include:

An adaptive estimator may therefore distinguish among

\[ \text{high-quality track}, \quad \text{degraded track}, \quad \text{acquisition}, \quad \text{invalid}. \]

Each state can correspond to a different measurement treatment.

4.2.27.24 IMU Adaptation

Inertial sensor behaviour can change with:

The estimator may adapt uncertainty associated with gyro or accelerometer measurements or associated bias models.

For example, a gyro bias model may be

\[ b_{g,k+1}=b_{g,k}+w_{b,k}, \]

with

\[ w_{b,k}\sim\mathcal N(0,Q_b). \]

If bias variability is known to increase under particular conditions,

\[ Q_b \]

may be adjusted accordingly.

4.2.27.25 LiDAR and Active-Ranging Adaptation

Active-ranging measurement quality may depend on:

Thus,

\[ R_{\text{range}} = f(r,\text{return quality},\text{geometry},\ldots). \]

The same principle applies to other active ranging sensors.

Adaptive fusion should use the sensor's actual information quality rather than assuming that every valid range measurement is statistically identical.

4.2.27.26 Adaptive Multi-Sensor Weighting

Suppose m sensors contribute information.

A conceptual adaptive weighted estimate can be written

\[ \hat x= \frac{\sum_{i=1}^{m}w_i(t)z_i} {\sum_{i=1}^{m}w_i(t)}. \]

The weights may depend on

\[ w_i= f(R_i,\epsilon_i,q_i,H_i,G_i,\ldots), \]

where:

This illustrates an important shift:

\[ \boxed{ \text{sensor dominance can change dynamically} } \]

rather than being permanently assigned during design.

4.2.27.27 Adaptive Fusion Is Not Arbitrary Weighting

Assigning empirical weights such as

\[ w_1=0.7,\qquad w_2=0.3 \]

without a physical or statistical basis is not robust adaptive estimation.

Weights should represent something meaningful, such as:

Otherwise, tuning may work for one simulation but fail when operating conditions change.

4.2.27.28 Sensor Health and Adaptive Fusion

Sensor health and sensor measurement quality should be related but distinct.

A sensor can be

\[ \text{healthy but noisy}, \]

or

\[ \text{healthy but poorly observable}, \]

or

\[ \text{healthy but temporarily low quality}. \]

Therefore,

\[ \boxed{ \text{hardware health} \neq \text{measurement quality} } \]

A useful adaptive architecture considers both.

A hardware fault may require exclusion, whereas reduced measurement quality may only require covariance inflation.

4.2.27.29 Adaptive Gating

Measurement gating commonly uses

\[ \epsilon_i= \nu_i^TS_i^{-1}\nu_i. \]

A fixed threshold may be

\[ \epsilon_i<\gamma. \]

In some systems, gating behaviour itself may depend on operating mode, expected transient behaviour, measurement dimension, or sensor state.

However, adaptive thresholds must be designed carefully.

Making the gate wider simply because residuals are large can cause actual faults to be accepted.

The adaptation must preserve the intended statistical meaning of the validation test.

4.2.27.30 Outlier Handling

Adaptive fusion should distinguish gradual degradation from isolated outliers.

An isolated abnormal measurement may be handled through:

Persistent deterioration may justify modification of the sensor covariance model.

Therefore,

\[ \boxed{ \text{outlier} \neq \text{persistent noise change} \neq \text{bias fault} } \]

These conditions require different responses.

4.2.27.31 Robust Adaptive Fusion

Robust estimation and adaptive estimation can be combined.

A robust estimator reduces the influence of extreme residuals.

An adaptive estimator changes its uncertainty assumptions according to longer-term evidence.

Conceptually,

\[ w_i= w_{\text{robust}}(r_i) \, w_{\text{adaptive}}(q_i). \]

The first component protects against individual abnormal measurements.

The second responds to changing measurement quality.

This combination can provide greater resilience than either technique alone.

4.2.27.32 Adaptation Rate

Suppose an adaptive covariance parameter follows

\[ R_{k+1} = (1-\lambda)R_k + \lambda\hat R_k. \]

The parameter

\[ 0<\lambda\le1 \]

controls adaptation speed.

Large \lambda produces rapid adaptation.

Small \lambda produces slow, smooth adaptation.

Too-fast adaptation can cause the estimator to chase random noise.

Too-slow adaptation may fail to respond quickly enough to genuine sensor degradation.

4.2.27.33 Forgetting Factors

Another common concept is exponential weighting.

For a statistic C_k,

\[ C_k= \lambda C_{k-1} + (1-\lambda)\nu_k\nu_k^T, \]

where

\[ 0<\lambda<1. \]

Older measurements gradually receive less influence.

A value of \lambda close to one provides long memory.

A smaller value responds more rapidly to recent changes.

Forgetting factors therefore provide a convenient mechanism for tracking slowly changing noise characteristics.

4.2.27.34 Adaptation Bounds

Adaptive parameters should usually be bounded.

For example,

\[ R_{\min} \preceq R_k \preceq R_{\max}. \]

Similarly,

\[ Q_{\min} \preceq Q_k \preceq Q_{\max}. \]

Without bounds, transient residuals or numerical problems could drive the estimator toward unrealistic covariance values.

Bounds provide an important safety and verification mechanism.

4.2.27.35 Positive-Semidefinite Covariance

Any adapted covariance must remain physically meaningful.

For measurement covariance,

\[ R_k\succeq0. \]

For process covariance,

\[ Q_k\succeq0. \]

An adaptation law that produces a negative variance or indefinite covariance is invalid.

Therefore, adaptive algorithms often require:

This becomes especially important for full matrix adaptation.

4.2.27.36 Diagonal Versus Full-Matrix Adaptation

A simple implementation may adapt only diagonal terms:

\[ R_k= \operatorname{diag} ( \sigma_1^2(k), \dots, \sigma_m^2(k) ). \]

This is comparatively easy to tune and verify.

A more sophisticated method may adapt the full covariance:

\[ R_k= \begin{bmatrix} \sigma_1^2&r_{12}&\cdots\\ r_{21}&\sigma_2^2&\cdots\\ \vdots&\vdots&\ddots \end{bmatrix}. \]

Full-matrix adaptation can represent changing correlations but requires substantially more data and careful numerical treatment.

4.2.27.37 Correlation Must Still Be Considered

Adaptive weighting does not remove the correlation problem.

Suppose two sensors share a common error:

\[ z_1=x+b_c+v_1, \]

\[ z_2=x+b_c+v_2. \]

Even if their individual variances are adapted correctly, treating the measurements as independent can still overestimate their combined information.

Therefore,

\[ \boxed{ \text{adaptive covariance} \neq \text{automatic independence} } \]

Cross-sensor and temporal correlations must still be considered.

4.2.27.38 Adaptive Fusion and Redundancy

Section 4.2.26 introduced redundant sensing.

Adaptive fusion extends that concept.

Instead of treating redundant sensors as simply

\[ \text{available} \quad\text{or}\quad \text{unavailable}, \]

their contributions can vary continuously:

\[ \text{high confidence} \rightarrow \text{nominal} \rightarrow \text{degraded} \rightarrow \text{low weight} \rightarrow \text{excluded}. \]

Thus redundancy provides alternative information sources, while adaptation determines how strongly those sources should contribute under current conditions.

4.2.27.39 Adaptive Sensor Selection

Sometimes the best response is not continuous weighting but selection.

Suppose

\[ Q_i(t) \]

represents the current quality of sensor i.

Then

\[ i^*=\arg\max_i Q_i(t). \]

The selected sensor may change as conditions change.

However, rapid switching can create undesirable estimator transients.

Therefore adaptive selection usually requires:

4.2.27.40 Hysteresis

Suppose sensor B becomes preferred when

\[ Q_B-Q_A>\Delta Q_{\text{switch}}. \]

Returning immediately when the difference becomes slightly smaller can cause repeated switching.

Different transition thresholds can be used:

\[ \gamma_{\text{in}} < \gamma_{\text{out}}. \]

This creates hysteresis and prevents chattering.

Adaptive fusion therefore benefits from the same transition-management concepts used in redundancy management and sensor handover.

4.2.27.41 Dwell Time

A new quality condition may need to persist for

\[ T\ge T_{\text{dwell}} \]

before changing the fusion configuration.

This prevents temporary measurement fluctuations from causing unnecessary adaptation.

A useful sequence is

\[ \text{quality change detected} \rightarrow \text{confirmed} \rightarrow \text{adapt} \rightarrow \text{monitor}. \]

4.2.27.42 Smooth Sensor Reweighting

Rather than abruptly changing

\[ R_1\rightarrow R_2, \]

the covariance can be blended:

\[ R(\alpha) = (1-\alpha)R_1+\alpha R_2, \]

where

\[ 0\le\alpha\le1. \]

Similarly, measurement weights can transition smoothly.

This reduces sudden changes in estimator gain and state corrections.

However, covariance interpolation must preserve appropriate statistical interpretation.

4.2.27.43 Adaptive Fusion During Sensor Transitions

When one sensor becomes less useful and another becomes more useful, an overlap interval can be exploited.

Initially,

\[ w_A\gg w_B. \]

During transition,

\[ w_A\downarrow, \qquad w_B\uparrow. \]

Eventually,

\[ w_B\gg w_A. \]

This is preferable to a purely geometric switch when both sensors provide useful measurements over an overlap region.

The transition should be driven by validated measurement quality rather than only by a nominal switching point.

4.2.27.44 Mode-Dependent Adaptive Fusion

Different mission phases can require different sensor-fusion behaviour.

A general architecture may define

\[ \mathcal M = \{ M_1,M_2,\ldots,M_n \}. \]

Each mode can specify:

Adaptive fusion then operates inside these mode-specific constraints.

This combines predictable architecture with online flexibility.

4.2.27.45 General RPO Example

In rendezvous and proximity operations, relative-navigation sensor quality may change significantly with separation, viewing geometry, target visibility, illumination, and sensor operating range.

A general sensor suite might include combinations of:

At longer separation, one sensor family may provide stronger relative-state information.

As separation decreases, optical or active-ranging measurements may become more informative.

Rather than using an instantaneous switch,

\[ S_A\rightarrow S_B, \]

adaptive fusion can exploit an overlap interval:

\[ S_A+S_B\rightarrow\text{cross-check and fusion}. \]

The estimator can gradually change measurement weighting as actual quality changes.

This is a general RPO navigation principle and does not depend on any particular spacecraft architecture.

4.2.27.46 Range Alone Should Not Determine Sensor Weight

Suppose an optical sensor is normally expected to improve at close range.

It would be tempting to define

\[ R_{\text{optical}}=f(r) \]

using range alone.

But at the same range, measurement quality may differ because of:

Therefore, a more realistic model may be

\[ R_{\text{optical}} = f(r,q_{\text{image}},G,\ldots). \]

The same principle applies to other sensor types.

4.2.27.47 Adaptive Fusion and Observability

A sensor may produce low-noise measurements but contribute little information about a particular state direction.

Therefore adaptation should not focus exclusively on scalar quality scores.

For sensor i,

\[ \mathcal I_i = H_i^TR_i^{-1}H_i \]

provides a useful representation of measurement information.

Changing geometry changes H_i.

Changing sensor quality changes R_i.

Therefore both determine actual information contribution.

4.2.27.48 Adaptive Fusion and Covariance Directionality

Suppose relative position covariance is

\[ P_r= \begin{bmatrix} \sigma_x^2 & p_{xy} & p_{xz}\\ p_{yx} & \sigma_y^2 & p_{yz}\\ p_{zx} & p_{zy} & \sigma_z^2 \end{bmatrix}. \]

A new sensor may significantly reduce uncertainty in two directions but contribute little in the third.

Therefore navigation quality should not always be summarized using one scalar such as

\[ \operatorname{tr}(P_r). \]

Adaptive sensor management may need to consider eigenvalues, principal uncertainty directions, or mission-specific state requirements.

4.2.27.49 Adaptive Fusion and Navigation Integrity

Adaptive fusion must preserve integrity.

An estimator that continually changes its uncertainty model can appear to perform well while actually becoming statistically inconsistent.

Therefore the system should monitor not only estimated state accuracy but also whether its covariance remains credible.

Useful quantities include:

\[ \text{NIS}, \qquad \text{NEES when truth is available}, \qquad \text{innovation mean}, \qquad \text{innovation covariance}. \]

Adaptation should improve consistency rather than merely make residual plots look smaller.

4.2.27.50 Adaptive Fusion and FDIR

Adaptive fusion and fault detection interact strongly.

Suppose residuals increase.

There are at least two possible interpretations:

\[ \text{sensor quality degraded} \]

or

\[ \text{sensor fault}. \]

If every large residual causes R to increase, a genuine sensor fault might simply be down-weighted indefinitely rather than isolated.

Conversely, if every residual increase causes immediate rejection, normal changes in measurement quality may produce unnecessary sensor loss.

Therefore,

\[ \boxed{ \text{adaptation and fault isolation must be coordinated} } \]

4.2.27.51 Adaptation Must Not Hide Faults

Consider a sensor developing a growing bias:

\[ z=x+b(t)+v. \]

If the adaptive algorithm responds only by increasing

\[ R, \]

the estimator may tolerate the growing residual without identifying the underlying bias.

A robust architecture should separately examine:

Adaptation should not become a mechanism for concealing model or hardware failures.

4.2.27.52 Fault-Aware Adaptation

Sensor health can explicitly influence covariance.

For example,

\[ R_i^{\text{eff}} = \alpha_iR_i, \]

where

\[ \alpha_i = f(H_i,q_i,\epsilon_i). \]

Possible behaviour might be

\[ \alpha_i=1 \]

for nominal operation,

\[ \alpha_i>1 \]

for degraded operation,

and measurement exclusion for a confirmed fault.

This integrates statistical quality information with FDIR state.

4.2.27.53 Adaptive Bias Estimation

Instead of only changing measurement covariance, some systematic sensor errors can be represented explicitly as states.

For example,

\[ z=x+b+v \]

with

\[ b_{k+1}=b_k+w_b. \]

The estimator can estimate b.

If bias variability changes,

\[ Q_b \]

can itself be adapted.

This is often more physically meaningful than treating every systematic residual as additional white measurement noise.

4.2.27.54 Model Adaptation Versus Noise Adaptation

Suppose the measurement model is incorrect:

\[ z=h_{\text{true}}(x)+v, \]

while the estimator uses

\[ z=h_{\text{model}}(x)+v. \]

Increasing R may reduce the effect of the mismatch but does not correct the model.

Therefore,

\[ \boxed{ \text{model error} \neq \text{measurement noise} } \]

Adaptive covariance tuning should not substitute for correcting known physical modelling errors.

4.2.27.55 Multiple-Model Adaptive Estimation

Sometimes the system can operate according to several possible models:

\[ M_1,M_2,\ldots,M_n. \]

Each model may maintain its own estimate

\[ \hat x_i,P_i \]

and likelihood

\[ \Lambda_i=p(z|M_i). \]

Model probabilities can be updated according to measurement consistency.

The combined estimate can then depend on the relative model probabilities.

This provides another form of adaptive estimation: rather than continuously changing one covariance matrix, the estimator changes confidence among several predefined models.

4.2.27.56 Interacting Multiple Model Concept

An Interacting Multiple Model approach can maintain several dynamic hypotheses simultaneously.

For example,

\[ M_1=\text{nominal dynamics}, \]

\[ M_2=\text{maneuvering dynamics}, \]

\[ M_3=\text{disturbed dynamics}. \]

Each model produces a state estimate.

Measurement likelihoods modify the probability associated with each model.

The resulting estimate represents a probability-weighted combination of the model-conditioned estimates.

This is particularly useful when system behaviour changes between distinct dynamic regimes.

4.2.27.57 Adaptive Estimation and Machine Learning

Machine-learning models can also estimate sensor quality or noise characteristics.

For example,

\[ R_k = f_\theta( \text{temperature}, \text{signal quality}, \text{geometry}, \text{image metrics}, \ldots ). \]

A learned model can capture complex relationships that are difficult to express analytically.

However, using ML inside a navigation estimator introduces additional concerns:

Therefore ML-based adaptation should not automatically be assumed superior to physics-based or statistically derived adaptation.

4.2.27.58 Physics-Guided Adaptive Fusion

A useful compromise is to preserve the physical estimator structure while allowing data-driven models to estimate particular quantities.

For example,

\[ \text{sensor/environment data} \xrightarrow{\text{ML}} \hat R_k \]

followed by

\[ \hat R_k \xrightarrow{\text{bounded validation}} R_k \xrightarrow{} \text{Kalman update}. \]

The state-estimation equations remain physically interpretable, while the difficult sensor-quality relationship is learned from data.

Bounds and validation can prevent arbitrary ML outputs from directly controlling estimator confidence.

4.2.27.59 Adaptation Using External Context

Sensor quality may sometimes be predicted before residuals become poor.

For example, the system may know:

Then

\[ R_k = f(\text{predicted operating condition}) \]

can be adjusted proactively.

This is different from purely residual-driven adaptation, which reacts after the measurements begin behaving differently.

4.2.27.60 Reactive Versus Predictive Adaptation

Two broad approaches can therefore be distinguished.

Reactive adaptation

\[ \text{residual changes} \rightarrow \text{detect degradation} \rightarrow \text{adapt}. \]

Predictive adaptation

\[ \text{known operating condition} \rightarrow \text{predict quality change} \rightarrow \text{adapt}. \]

A practical spacecraft navigation architecture can use both.

Predictive adaptation prepares the estimator for expected changes.

Reactive adaptation detects behaviour that was not predicted correctly.

4.2.27.61 Adaptive Fusion With Asynchronous Sensors

Spacecraft sensors frequently operate at different rates.

Suppose

\[ f_1=100\text{ Hz}, \qquad f_2=10\text{ Hz}, \qquad f_3=1\text{ Hz}. \]

Adaptive fusion should respect the timestamp and quality of each individual measurement.

The estimator may therefore perform

\[ \text{propagate} \rightarrow z_1 \rightarrow \text{propagate} \rightarrow z_1 \rightarrow z_2 \rightarrow\cdots \]

while updating the appropriate covariance model for whichever measurement arrives.

Adaptation should not require artificially synchronizing all sensors.

4.2.27.62 Adaptive Fusion and Measurement Latency

Measurement quality is not the only changing property.

Latency can also vary.

If a measurement corresponds to

\[ t_m \]

but is processed at

\[ t_p=t_m+\Delta t, \]

then using it as a current-time measurement can produce an apparent residual.

An adaptive algorithm must not incorrectly interpret latency-induced disagreement as increased sensor noise.

Time alignment therefore remains fundamental.

4.2.27.63 Adaptive Fusion and Frame Errors

Similarly, a frame transformation error can produce persistent innovation.

Suppose

\[ z_B \]

must be transformed through

\[ C_{AB}. \]

If C_{AB} is incorrect, increasing R may make residual statistics appear less abnormal, but the physical problem remains.

Therefore large innovations should trigger diagnostic reasoning before covariance is modified blindly.

4.2.27.64 Adaptive Fusion Architecture

A useful conceptual architecture is

\[ \boxed{ \text{Sensors} \rightarrow \text{Validation} \rightarrow \text{Quality Assessment} \rightarrow \text{Adaptive Noise Model} \rightarrow \text{Estimator} } \]

with feedback:

\[ \boxed{ \text{Estimator Innovations} \rightarrow \text{Consistency Monitor} \rightarrow \text{Adaptive Noise Model} } \]

and a parallel path:

\[ \text{Sensor Health} \rightarrow \text{FDIR} \rightarrow \text{Sensor Manager}. \]

Thus adaptive fusion is not simply one equation inside the Kalman filter. It is a system-level navigation function.

4.2.27.65 Practical Processing Sequence

For every measurement z_i, a practical adaptive-fusion sequence may be:

\[ z_i \rightarrow \text{timestamp validation} \rightarrow \text{frame/unit validation} \rightarrow \text{hardware health} \rightarrow \text{quality assessment} \rightarrow \text{innovation test} \rightarrow R_i^{\text{eff}} \rightarrow \text{measurement update}. \]

Afterward,

\[ \nu_i,\ S_i \]

can be returned to the quality and FDIR logic.

This produces a closed information loop.

4.2.27.66 Adaptation State Machine

Adaptive behaviour can be organized using sensor states such as

\[ \mathcal H_i= \{ \text{nominal}, \text{degraded}, \text{suspicious}, \text{excluded}, \text{recovering} \}. \]

For example,

\[ \text{nominal} \rightarrow \text{degraded} \]

may cause covariance inflation.

\[ \text{degraded} \rightarrow \text{suspicious} \]

may tighten monitoring.

\[ \text{suspicious} \rightarrow \text{excluded} \]

may remove the measurement.

Recovery may require sustained good behaviour before returning to nominal operation.

4.2.27.67 Preventing Adaptation Chatter

Suppose sensor quality fluctuates around a threshold.

Without protection,

\[ R_{\text{nominal}} \leftrightarrow R_{\text{degraded}} \]

could switch repeatedly.

This can cause rapidly changing Kalman gains and undesirable state transients.

Useful protections include:

4.2.27.68 Rate-Limited Adaptation

The covariance may be constrained so that

\[ \left\| R_{k+1}-R_k \right\| \le \Delta R_{\max}. \]

This prevents one unusual measurement from producing a dramatic change in estimator weighting.

Rate limiting can make adaptive behaviour smoother and easier to verify.

However, emergency fault logic may still need the ability to reject clearly invalid measurements immediately.

4.2.27.69 Adaptation and Numerical Stability

Rapidly changing Q and R affect covariance propagation and Kalman gain computation.

Poorly designed adaptation can create:

Therefore adaptive fusion requires the same numerical discipline as the underlying estimator, often with additional safeguards.

4.2.27.70 Minimum and Maximum Sensor Influence

A system may impose limits on measurement influence.

For example,

\[ R_i\ge R_{i,\min} \]

prevents the estimator from becoming unrealistically confident in a sensor.

Likewise,

\[ R_i\le R_{i,\max} \]

may define the largest covariance for which the sensor remains meaningfully fused.

Beyond that point, explicit exclusion may be clearer than retaining an effectively zero-weight measurement.

4.2.27.71 Adaptive Fusion and Navigation Availability

Adaptive fusion can improve navigation availability because a degraded sensor does not necessarily need to be discarded.

Instead,

\[ \text{nominal sensor} \rightarrow \text{degraded sensor with larger }R \]

may still provide useful information.

This can preserve partial navigation capability when a fixed-threshold architecture might otherwise remove the sensor entirely.

However, availability should never be increased by pretending poor measurements are more accurate than they actually are.

4.2.27.72 Graceful Degradation

Adaptive fusion naturally supports graceful degradation.

Suppose a high-quality sensor deteriorates.

Instead of

\[ \text{nominal navigation} \rightarrow \text{navigation lost}, \]

the system may transition through

\[ \text{nominal} \rightarrow \text{reduced accuracy} \rightarrow \text{degraded mode}. \]

The covariance should increase accordingly.

Downstream guidance and control can then determine which operations remain permissible.

4.2.27.73 Navigation Quality and Control Authority

The adaptive estimator should expose not only

\[ \hat x \]

but also

\[ P, \qquad H_{\text{nav}}, \qquad M_{\text{nav}}. \]

If adaptive fusion determines that sensor quality has deteriorated and covariance has increased, control authority may need to change.

For example,

\[ P>P_{\max} \]

may inhibit a precision operation even though the estimator still produces a numerical state.

Thus,

\[ \boxed{ \text{adaptive navigation confidence should propagate into GNC decisions} } \]

4.2.27.74 Adaptation and Safety-Critical Operations

Adaptive behaviour should become especially controlled during operations requiring high navigation confidence.

The system may impose stricter requirements on:

Adaptive fusion should therefore operate inside mission-level safety constraints rather than independently of them.

4.2.27.75 Adaptive Fusion Verification

Adaptive estimators are harder to verify than fixed estimators because their behaviour depends on both measurements and internal adaptation logic.

Verification should therefore include cases with:

The objective is to verify both estimation performance and adaptation behaviour.

4.2.27.76 Monte Carlo Verification

Monte Carlo analysis is particularly useful.

For run j,

\[ e_k^{(j)} = x_k^{(j)}-\hat x_k^{(j)}. \]

Across many realizations, the designer can examine:

Adaptive logic that appears effective in one deterministic simulation may behave very differently across randomized noise and fault realizations.

4.2.27.77 Adaptation Response Time

Suppose sensor degradation begins at

\[ t=t_f. \]

The adaptive algorithm responds at

\[ t=t_a. \]

Define

\[ T_{\text{adapt}} = t_a-t_f. \]

A useful adaptive system should respond quickly enough to prevent unacceptable navigation degradation while avoiding excessive reaction to random fluctuations.

Thus adaptation response time is itself an important performance metric.

4.2.27.78 False Adaptation

An adaptive algorithm can change its parameters even when the nominal sensor model remains valid.

Define conceptually

\[ P_{FA,\text{adapt}} = P( \text{adaptation triggered} \mid \text{nominal conditions} ). \]

Excessive false adaptation can cause unnecessary covariance inflation, reduced navigation accuracy, sensor switching, or operational restrictions.

Therefore adaptation sensitivity must be evaluated statistically.

4.2.27.79 Recovery Verification

After sensor quality returns to normal, the estimator should recover appropriately.

Important questions include:

Recovery is as important as degradation detection.

4.2.27.80 Testing Abrupt and Gradual Changes

Adaptive algorithms should be tested against both abrupt changes,

\[ R: R_0\rightarrow R_1, \]

and gradual changes,

\[ R(t)=R_0+\alpha t. \]

Abrupt changes test response speed.

Gradual changes test sensitivity to slow degradation.

An algorithm designed only for one type may perform poorly for the other.

4.2.27.81 Testing Misleading Quality Indicators

A sensor quality indicator can itself be wrong.

For example,

\[ q_i=\text{high} \]

while actual measurement quality is poor.

The adaptive estimator should therefore not depend blindly on one quality metric.

Cross-checking quality indicators against innovations and independent measurements improves robustness.

This leads to the principle

\[ \boxed{ \text{reported quality should itself be validated} } \]

4.2.27.82 Testing Correlated Sensors

Adaptive fusion should also be tested when sensor errors are correlated.

For example,

\[ E[v_1v_2^T]=R_{12}\neq0. \]

If the adaptation algorithm treats both sensors as independently improving, it may reduce covariance excessively.

Therefore Monte Carlo campaigns should include realistic common-mode and cross-sensor error sources.

4.2.27.83 Adaptive Fusion Requirements

A practical adaptive-fusion requirement should be measurable.

Instead of saying:

“The filter shall adapt to changing sensor quality.”

a stronger requirement might specify:

Adaptive behaviour becomes verifiable only when it is defined quantitatively.

4.2.27.84 Logging Requirements

Adaptive estimators require extensive logging.

Useful quantities include:

\[ \hat x_k, \qquad P_k, \qquad \nu_k, \qquad S_k, \qquad Q_k, \qquad R_k, \]

together with:

Without these signals, diagnosing why an adaptive estimator changed behaviour can be extremely difficult.

4.2.27.85 Explainability and Traceability

For a safety-relevant navigation system, it should be possible to answer:

Why did the estimator stop trusting this sensor?

A traceable explanation might be:

\[ \text{image quality decreased} \rightarrow R_{\text{camera}}\uparrow \rightarrow K_{\text{camera}}\downarrow \rightarrow \text{camera contribution decreased}. \]

This is much easier to verify than an opaque mechanism whose weighting changes cannot be related to measurable physical quantities.

4.2.27.86 Adaptive Fusion Design Workflow

A systematic development process is:

  1. Characterize each sensor under expected operating conditions.
  2. Identify variables that affect measurement quality.
  3. Develop nominal R and Q models.
  4. Determine measurable quality indicators.
  5. Establish mappings from quality to covariance or weight.
  6. Define innovation-based consistency monitoring.
  7. Define adaptation bounds and rates.
  8. Coordinate adaptation with FDIR.
  9. Introduce hysteresis and dwell logic where required.
  10. Test individual degradation mechanisms.
  11. Test multiple-sensor interactions.
  12. Perform Monte Carlo verification.
  13. Validate using hardware or representative sensor data.
  14. Verify downstream GNC behaviour under degraded navigation.

4.2.27.87 Common Mistake: Adapting Everything

An adaptive estimator with too many free parameters can become difficult to understand and verify.

If every element of

\[ Q_k \]

and

\[ R_k \]

changes independently, the system may have many combinations of behaviour.

Furthermore, Q and R can sometimes produce similar effects on innovation statistics, making them difficult to identify simultaneously.

Therefore,

\[ \boxed{ \text{adapt only quantities that are observable and physically justified} } \]

is a strong design principle.

4.2.27.88 Distinguishing `Q` and `R` Errors

Suppose innovations are larger than expected.

This could mean

\[ R_{\text{actual}}>R_{\text{model}}, \]

but it could also mean

\[ Q_{\text{actual}}>Q_{\text{model}}. \]

Both can increase innovation magnitude.

Therefore blindly increasing R whenever residuals increase can misdiagnose process-model uncertainty as sensor degradation.

Additional evidence is needed, such as:

4.2.27.89 Identifiability of Adaptive Parameters

Suppose both Q and R are unknown.

The measurements may not always contain enough information to uniquely determine both.

This is an identifiability problem.

An adaptive estimator can appear mathematically sophisticated while estimating quantities that cannot actually be distinguished from available data.

Therefore adaptive parameters should be selected with observability and identifiability in mind.

4.2.27.90 Offline Versus Online Adaptation

Not every changing covariance must be estimated online.

Sensor characterization may establish an offline model:

\[ R=f(T,r,q,\ldots). \]

During operation, measured values of temperature T, range r, or quality q are inserted into this predefined model.

This can be easier to verify than estimating R directly from innovations.

A practical architecture may combine:

\[ \text{offline characterization} + \text{online quality scheduling} + \text{innovation monitoring}. \]

4.2.27.91 Scheduled Covariance

A covariance can be scheduled according to operating conditions.

For example,

\[ R(r)= \begin{cases} R_1,&r>r_1,\\ R_2,&r_2

This is simple and deterministic.

A continuous model,

\[ R(r)=f(r), \]

can provide smoother behaviour.

The choice depends on sensor characterization, implementation complexity, and verification requirements.

4.2.27.92 Hybrid Adaptive Fusion

A strong practical architecture can combine several mechanisms:

\[ R_i^{\text{base}} \]

from offline characterization,

\[ \alpha_{\text{quality}} \]

from real-time sensor quality,

and

\[ \alpha_{\text{innovation}} \]

from estimator consistency.

Then, conceptually,

\[ R_i^{\text{eff}} = \alpha_{\text{quality}} \alpha_{\text{innovation}} R_i^{\text{base}}, \]

subject to bounds:

\[ R_{\min} \preceq R_i^{\text{eff}} \preceq R_{\max}. \]

This combines prior physical knowledge with online evidence.

4.2.27.93 Sensor-Specific Adaptation

Different sensors should not necessarily use the same adaptation law.

A star tracker, IMU, GNSS receiver, camera, and ranging sensor have different:

Therefore,

\[ R_i=f_i(q_i,\text{environment},\nu_i,\ldots) \]

is usually more meaningful than imposing one generic adaptation rule on every sensor.

4.2.27.94 Adaptive Fusion and Computational Load

Adaptive fusion adds computations beyond the nominal estimator.

Additional tasks can include:

The architecture must therefore consider processor capability and update-rate requirements.

Adaptation that is mathematically attractive but cannot execute within the navigation cycle is not useful onboard.

4.2.27.95 Deterministic Execution

For real-time implementation, adaptive algorithms should have predictable execution behaviour where required.

For example, an algorithm whose number of optimization iterations varies greatly may create timing difficulties.

Bounded computations, fixed-size windows, and predefined adaptation limits can make implementation more deterministic.

This connects directly to practical filter implementation.

4.2.27.96 Adaptive Fusion and Initialization

At estimator startup, innovation statistics may not yet represent steady-state behaviour.

Immediately adapting Q or R during initialization can therefore be misleading.

A system may use

\[ T_{\text{adapt-enable}} \]

or require convergence criteria before full adaptation begins.

Alternatively, initialization may use separate adaptation parameters.

4.2.27.97 Adaptive Fusion After Sensor Reacquisition

A reacquired sensor should not necessarily return immediately to maximum weight.

A sequence may be

\[ \text{reacquired} \rightarrow \text{validated} \rightarrow \text{low weight} \rightarrow \text{nominal weight}. \]

As evidence accumulates, covariance can decrease gradually.

This protects the estimator from unstable measurements immediately after reacquisition.

4.2.27.98 Adaptation Memory

An important design question is whether the system remembers previous degradation.

Suppose a sensor repeatedly alternates between nominal and degraded behaviour.

A purely instantaneous adaptation algorithm may repeatedly restore full trust.

A health-history variable can retain information about repeated instability.

However, long memory can unfairly penalize a sensor that has genuinely recovered.

Thus adaptation memory must balance responsiveness and historical evidence.

4.2.27.99 Adaptive Fusion Under Multiple Sensor Degradation

A challenging case occurs when several sensors degrade simultaneously.

If every sensor residual increases, the cause may be:

Therefore adaptive fusion should not automatically inflate every R_i.

Cross-sensor patterns and system-level diagnostics become important.

4.2.27.100 Common-Mode Degradation

Suppose

\[ z_i=x+b_c+v_i \]

for several sensors.

If the common bias b_c changes, all sensors may remain mutually consistent while collectively becoming wrong.

Adaptive weighting among those sensors cannot detect the problem.

An independent sensing principle or analytical reference may be required.

Therefore sensor diversity remains essential even in sophisticated adaptive fusion.

4.2.27.101 Adaptive Fusion Does Not Replace Sensor Modelling

Adaptive methods are sometimes viewed as a way to avoid accurate sensor models.

That is a mistake.

A strong adaptive estimator begins with good models for:

Adaptation then addresses legitimate variation around those models.

It should not compensate indefinitely for poor modelling.

4.2.27.102 Adaptive Fusion Does Not Guarantee Robustness

An adaptive algorithm introduces another feedback loop.

The estimator affects innovations.

Innovations affect adaptation.

Adaptation changes Q or R.

Those changes affect the estimator again.

Conceptually,

\[ \hat x \rightarrow \nu \rightarrow R \rightarrow K \rightarrow \hat x. \]

Poorly designed adaptation can therefore destabilize or distort estimation.

Adaptive does not automatically mean robust.

4.2.27.103 Stability of the Adaptation Loop

The adaptation rate should normally be slower than the estimator's response to individual measurement noise.

If adaptation reacts strongly to every innovation,

\[ \nu_k\uparrow \rightarrow R_k\uparrow \rightarrow K_k\downarrow \]

and then

\[ \nu_{k+1}\downarrow \rightarrow R_{k+1}\downarrow, \]

the covariance may oscillate.

Filtering, hysteresis, bounded rates, and statistical windows help separate persistent changes from random variation.

4.2.27.104 Adaptive Fusion Performance Metrics

Useful evaluation metrics include:

No single metric is sufficient.

An adaptive estimator can reduce RMSE while becoming statistically overconfident, or improve consistency while unnecessarily degrading accuracy.

4.2.27.105 Comparison With Fixed Fusion

A meaningful evaluation compares adaptive fusion with a fixed baseline.

For identical scenarios, compare

\[ \text{Fixed }Q,R \]

against

\[ \text{Adaptive }Q_k,R_k. \]

Under nominal conditions, adaptive fusion should not introduce unnecessary degradation.

Under changing conditions, it should demonstrate measurable benefits such as:

Without this baseline, it can be difficult to demonstrate whether adaptation actually adds value.

4.2.27.106 Verification Pyramid

Adaptive fusion can be verified progressively.

Level 1 — Mathematical tests

Check covariance equations, limits, positive semidefiniteness, and adaptation laws.

Level 2 — Synthetic sensor tests

Inject controlled noise, bias, outliers, and quality changes.

Level 3 — Closed-loop simulation

Evaluate navigation–guidance–control interaction.

Level 4 — Monte Carlo

Randomize sensor and system uncertainties.

Level 5 — Recorded sensor data

Replay representative measurements.

Level 6 — SIL/HIL testing

Evaluate timing, interfaces, numerical behaviour, and implementation.

Level 7 — Integrated system testing

Verify the complete navigation and control chain.

4.2.27.107 Adaptive Fusion in Simulink

A conceptual Simulink implementation can separate the architecture into:

\[ \text{Sensor Models} \rightarrow \text{Measurement Bus} \rightarrow \text{Validation} \rightarrow \text{Quality Estimator} \rightarrow \text{Adaptive }R/Q \rightarrow \text{Estimator}. \]

Useful logged signals include:

\[ z_i, \quad \nu_i, \quad S_i, \quad R_i, \quad Q, \quad K, \quad \hat x, \quad P. \]

Separate subsystems for quality estimation and covariance adaptation make the model easier to inspect and verify than embedding all logic inside one estimator block.

4.2.27.108 Recommended Separation of Responsibilities

A clean architecture separates:

Sensor processing

Produces calibrated, timestamped measurements and quality information.

Sensor validation

Checks status, units, range, timing, and basic plausibility.

FDIR

Determines health and fault state.

Adaptive quality logic

Maps operating conditions and statistical evidence into effective uncertainty.

Estimator

Performs state prediction and measurement updates.

Navigation monitor

Checks covariance, consistency, availability, and mode readiness.

This separation improves traceability and verification.

4.2.27.109 Example Adaptive Measurement Path

A generic measurement path can be represented as

\[ z_i^{raw} \rightarrow z_i^{cal} \rightarrow z_i^{valid} \rightarrow q_i \rightarrow R_i^{eff} \rightarrow \nu_i \rightarrow \text{gate} \rightarrow \text{update}. \]

After the update,

\[ \nu_i,S_i \]

feed back into quality monitoring.

Thus adaptation occurs both before and after the estimator evaluates the measurement.

4.2.27.110 Key Design Questions

When designing adaptive sensor fusion, the engineer should ask:

These questions are more important than simply choosing a sophisticated adaptive algorithm.

4.2.27.111 Practical Design Principle

A useful hierarchy is

\[ \boxed{ \text{Physical sensor model} \rightarrow \text{Offline characterization} \rightarrow \text{Real-time quality assessment} \rightarrow \text{Bounded adaptation} \rightarrow \text{Statistical monitoring} \rightarrow \text{FDIR} } \]

The estimator should begin with physically meaningful uncertainty models.

Online adaptation should then modify those models only when justified by measurable conditions or statistically significant evidence.

4.2.27.112 Complete Conceptual Adaptive Sensor-Fusion Architecture

A complete spacecraft adaptive-fusion architecture can be represented conceptually as

\[ \boxed{ \begin{array}{c} \text{Sensor Measurements}\\ \downarrow\\ \text{Calibration / Time / Frame Processing}\\ \downarrow\\ \text{Measurement Validation}\\ \downarrow\\ \text{Sensor Quality Assessment}\\ \downarrow\\ \text{Adaptive }R_i\text{ / Weight Generation}\\ \downarrow\\ \text{Innovation Gating}\\ \downarrow\\ \text{Multi-Sensor Estimator}\\ \downarrow\\ \hat x,\ P \end{array} } \]

with feedback

\[ \boxed{ \nu_i,S_i \rightarrow \text{Consistency Monitoring} \rightarrow \text{Adaptive Logic} } \]

and a parallel FDIR path

\[ \boxed{ \text{Sensor Status + Residuals + Cross-Checks} \rightarrow \text{FDIR} \rightarrow \text{Sensor Health / Reconfiguration}. } \]

At the system level,

\[ \hat x,\ P,\ H_{\text{nav}} \]

are then provided to guidance and control.

4.2.27.113 Key Takeaways

Adaptive sensor fusion responds to changing information quality rather than assuming fixed sensor statistics.

The most important ideas are:

The central principle is

\[ \boxed{ \text{Trust each sensor according to the information it can reliably provide now, not merely according to its nominal specification.} } \]

The next topic, 4.2.28 Practical Filter Implementation, naturally follows from this. Once the sensor models, fusion algorithms, redundancy management, and adaptive logic have been established mathematically, they must be converted into numerically stable, computationally efficient, real-time software suitable for practical spacecraft GNC implementation.

Next: 4.2.28 Practical Filter Implementation

After developing adaptive sensor fusion through measurement-quality-dependent weighting, adaptive process and measurement covariance, innovation-based adaptation, sensor-health interaction, bounded adaptation, graceful degradation, and verification, the next section converts estimator theory into numerically stable, computationally efficient, real-time spacecraft navigation software.

4.2.28 Practical Filter Implementation

Continuous-to-discrete conversion, state propagation, covariance propagation, numerical stability, covariance symmetry, positive-semidefinite covariance, matrix conditioning, Joseph-form covariance updates, computational load, and embedded implementation.

Continue to Practical Filter Implementation →