An adaptive kalman filter edge control method based on subjective logic self-assessment

Abstract

The application provides a kind of adaptive Kalman filtering edge control method based on subjective logic self-evaluation, the application adopts four-layer architecture of offline lightweight modeling-online subjective logic self-evaluation-working condition adaptive parameter correction-robust control adaptation, realizes the quantitative evaluation of filtering state and multi-sensor data through subjective logic opinion triple, accurately identifies the root cause of filtering deviation;Combined with working condition identification, dynamically adjust the self-evaluation threshold and correction step, and correct the noise covariance matrix Q / R in a lightweight manner;After smoothing and fault-tolerant processing, output stable estimated value to support edge control decision.This technology can be widely applied to edge control scenarios in steel, power, chemical and other industrial fields (SCR denitration ammonia injection control, industrial furnace temperature regulation, fluid pressure control, motor speed regulation, etc.), can significantly improve the precision, stability and robustness of edge control, reduce equipment wear and tear, reduce the risk of environmental over-standard, and has advanced technology and industrial practicality.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically to an adaptive Kalman filter edge control method based on subjective logic self-evaluation. Background Technology

[0002] In edge control scenarios in industries such as steel, power, and chemicals, edge controllers rely on real-time data from multiple sensors to make rapid local decisions. Adaptive Kalman filtering is the core algorithm for sensor noise reduction and accurate state estimation, and its accuracy directly determines the stability and reliability of control. Industrial environments present problems such as dynamic fluctuations in operating conditions, environmental interference, and slow drift of equipment parameters, requiring filtering algorithms to have real-time self-evaluation and dynamic parameter correction capabilities.

[0003] The existing technology has the following bottlenecks: The lack of an online quantitative self-evaluation mechanism and reliance solely on residual thresholds make it impossible to distinguish between process noise, measurement noise, and sensor malfunctions, which can easily lead to incorrect parameter adjustments and increase estimation errors. The self-evaluation model is complex and computationally intensive, making it difficult to adapt to the limited computing power of edge controllers. It also has high execution latency and does not meet industrial real-time requirements. The self-evaluation index is fixed and cannot simultaneously adapt to slow drift and instantaneous change in operating conditions. It is prone to insufficient sensitivity or oversensitivity, which can cause control command chattering. In multi-sensor coupled scenarios, only one sensor is evaluated independently without consistency and conflict verification. Anomalies in a single sensor can easily lead to filtering distortion and control errors. The self-assessment and parameter correction are disconnected, there is no fault-tolerant smoothing mechanism, and instantaneous jumps trigger frequent adjustments, which aggravates wear and tear on the actuator and shortens its lifespan.

[0004] In summary, existing solutions struggle to balance high precision, low latency, lightweight design, and high robustness, failing to meet the actual needs of industrial edge control. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: S1: Offline lightweight modeling: Industrial historical data is preprocessed and operating condition clustered, adaptive Kalman filter parameters are initialized, and a lightweight subjective logic self-evaluation model is trained.

[0007] Collect industrial operation data covering at least three complete production cycles, including NO at the SCR denitrification inlet. x Concentration, flue gas flow rate, temperature, or motor speed, current, voltage, etc.; covering normal operating conditions, slow drift operating conditions, instantaneous change operating conditions, and sensor abnormal operating conditions.

[0008] The 3σ criterion is used to remove obvious abnormal data; correlation analysis is performed on multiple sensors to screen core sensing parameters, reduce the dimensions of online input, and reduce the edge computing load.

[0009] A lightweight K-means clustering method with a limited number of clusters of 3–5 is adopted. The clustering features are "core operating condition parameters - filter residuals - sensor data variance". The elbow rule is used to determine the optimal number of clusters, classify typical operating conditions, and provide a benchmark for online operating condition identification.

[0010] Based on the preprocessed historical data, the following was completed: Initialize the process matrix F and measurement matrix H; The initial process noise covariance matrix Q0 and the measurement noise covariance matrix R0 are calculated using the classical noise covariance estimation method. The confidence interval of NEES (normalized squared estimation error) is calculated based on the filter residuals and used as a benchmark for evaluating filter consistency.

[0011] Based on historical filtered states and sensor data, train a subjective logical opinion triplet model: Determine the base rate 'a' (prior reliability) for each sensor and filter state. Pre-calculate the DC benchmark threshold for conflict degree under different operating conditions; The complex matrix operations in subjective logic are simplified into scalar operations, and core parameters are pre-stored to ensure that the online calculation latency is ≤10ms.

[0012] S2: Online Subjective Logic Self-Assessment: Real-time acquisition and preprocessing of multi-sensor data; evaluation of filter state reliability and sensor data reliability based on subjective logical opinion triplets; calculation of conflict degree between sensors and between sensors and filter state; and fusion determination of the root cause of filter deviation.

[0013] The edge controller acquires multi-sensor data with a control cycle of 10–100ms; for data exceeding the process range, the average of the previous three cycles is used as a replacement; simultaneously, one-step prediction using Kalman filtering is performed, and the residual γ is calculated. k NEES with normalized estimation error squared k .

[0014] With NEES k Using residual variance as input, construct the subjective logical opinion of the filter state: ω X =(b X ,u X ,a X ) belief b X: Characterizes the consistency of filtering. NEES is 0.8–0.95 within the confidence interval, and decreases linearly with increasing deviation when it exceeds the limit; Uncertainty u X : Characterizes the degree of uncertainty in filtering. Under stable operating conditions, it is taken as 0.05–0.2. When fluctuating, it increases linearly with the increase of residual variance. base rate a X Offline pre-stored prior value, fixed at 0.5.

[0015] Opinions are mapped to filtered reliability probability P using projection probability. X And classify and determine: P X ≥0.8: Filtering is normal; 0.5≤P X <0.8: Filtering bias; P X <0.5: Filtering error.

[0016] Develop opinions for each sensor: ω S （S） =(b S （S） ,u S （S） ,a S （S） ) Where: b S （S） For the reliability of sensor data, u S （S） For uncertainty, a S （S） This is the offline pre-stored base rate.

[0017] Calculate the two types of conflict degree DC: Inter-sensor collision degree DC(S¹,S²); Sensor and filter state conflict degree DC(S) （S） ,X).

[0018] Classified by conflict level: DC≤0.3: No conflict; 0.3 < DC ≤ 0.6: Mild conflict; DC > 0.6: Severe conflict.

[0019] Accurate identification of the root cause of filtering bias: The root cause of the deviation is determined solely by fusing the filter evaluation results, sensor evaluation results, and conflict degree according to the following rules: (1) Filter deviation + severe conflict between a few sensors and the filter + no conflict between sensors → abnormal measurement noise; (2) Filter deviation + severe conflict between most sensors and the filter + no conflict between sensors → abnormal process noise; (3) Filter deviation + severe conflict between a certain sensor and other sensors and filters → sensor failure; (4) Filter deviation + slight conflict of all sensors + slight conflict with filter → overall fluctuation of operating conditions.

[0020] S3: Adaptive parameter correction for operating conditions: Real-time identification of operating conditions, dynamic adaptation of self-evaluation thresholds, and lightweight adjustment of process noise covariance matrix Q and measurement noise covariance matrix R based on subjective logical trust, with dynamic setting of correction step size according to operating conditions.

[0021] Based on the variance change rate and filter residual change trend of multi-sensor data, combined with offline clustering benchmarks, the conditions are divided into three categories in real time: Stable operating conditions; slow drift operating conditions; sudden change operating conditions.

[0022] The subjective logic threshold is adjusted in real time according to the type of working condition to achieve adaptive sensitivity. Stable operating conditions: Standard thresholds are used to ensure evaluation accuracy; Slow drift condition: Reduce the collision threshold to 0.2 and increase the filter reliability threshold to 0.85 to improve sensitivity; Transient change conditions: Increase the conflict threshold to 0.5 and decrease the filter reliability threshold to 0.7 to avoid misjudgment.

[0023] With Subjective Logic Trust Correction (TR) as its core, it adopts pure scalar multiplication to avoid complex matrix operations and is adapted to edge computing power.

[0024] Measurement noise R correction: For anomalous sensor s*, calculate the confidence correction factor RF. R Press R new =RF R ×R old renew; Process noise Q correction: When an abnormal process noise is identified, calculate the RFQ and adjust it according to Q. new =RF Q ×Q old renew.

[0025] Adjust the step size according to operating conditions to balance stability and response speed: Stable operating conditions: 0.1–0.3, parameters remain stable; Slow drift: 0.4–0.6, gradual correction; Transient mutation: 0.7–0.9, rapid adaptation.

[0026] S4: Robust control adaptation: The filtered state estimate is smoothed, the self-evaluation index is continuously and periodically fault-tolerantly determined, and the quantitative evaluation result is output to the edge control module to coordinately adjust the control strategy.

[0027] The six-point window averaging method is adopted: the most recent six filtered estimates are taken, the maximum and minimum values ​​are removed, and then the average is calculated to suppress the instantaneous fluctuations caused by parameter adjustment.

[0028] The following rules are set for determining whether an indicator exceeds the threshold for three consecutive sampling periods: only when the indicator exceeds the threshold for three consecutive periods will it be determined as a filter abnormality or sensor failure, and a three-level alarm will be output: Slight deviation; severe deviation; sensor malfunction, to avoid malfunctions caused by momentary interference.

[0029] Self-assessment quantitative indicators (P) X P S （S） Directly input to the edge control module to achieve dynamic policy adjustment: P X ≥0.8: Control uses standard parameters; 0.5≤P X <0.8: Increase the control correction step size to improve anti-interference capability; P X <0.5: Switch to multi-sensor fusion fault-tolerant mode to avoid control failure caused by abnormal filtering.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention introduces a subjective logic opinion triplet, which transforms the filter state and sensor state into calculable indicators of belief, uncertainty and base rate, and establishes clear deviation root cause judgment rules in combination with conflict degree analysis, so as to achieve accurate identification of abnormal types and fundamentally improve the rationality of filtering decision; (2) The present invention uses offline lightweight modeling to complete working condition clustering, parameter initialization and subjective logic model pre-training, and puts complex calculations in the front; in the online stage, matrix operation is simplified to scalar operation, reducing floating point calculation and iteration times, avoiding the computational power consumption caused by complex probability reasoning and statistical models, so that the algorithm can be executed efficiently on edge hardware; and through real-time working condition identification, the conflict degree threshold and reliability judgment threshold of subjective logic self-evaluation are dynamically adjusted, and a dynamic correction step size is set: when facing slow drift, the evaluation sensitivity is improved to ensure timely correction of deviation; when facing instantaneous change, the evaluation sensitivity is reduced to avoid false triggering and frequent adjustment, so as to achieve stable adaptation under different working conditions. Attached Figure Description

[0031] Figure 1 This is a schematic diagram illustrating the specific process of the present invention. Detailed Implementation

[0032] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0033] Please refer to the reference. Figure 1 This invention provides an adaptive Kalman filter edge control method based on subjective logic self-evaluation.

[0034] S1: Key technologies for offline lightweight modeling layer implementation: Industrial field data preprocessing and operating condition clustering: Collect historical operating data of industrial edge control scenarios (such as inlet NOx concentration, flue gas flow and temperature of SCR denitrification, speed, current and voltage of motor control, etc.), covering typical scenarios such as normal operating conditions, slow drift operating conditions, instantaneous change operating conditions, and sensor abnormal operating conditions, with a time span covering at least 3 complete production cycles.

[0035] Lightweight K-means clustering (limiting the number of clusters to 3-5 to reduce computational load) is adopted. The clustering feature is "core operating condition parameters - filter residuals - sensor data variance". The elbow rule is used to determine the optimal number of clusters and classify typical operating conditions. Abnormal data is removed based on the 3σ criterion. At the same time, correlation analysis is performed on multi-sensor data to screen core sensing parameters, reduce the online input dimension, and reduce the edge computing load.

[0036] Adaptive Kalman filter basic parameter initialization: Based on the filtered historical data, the process matrix F and measurement matrix H of the Kalman filter are initialized. The process noise covariance matrix Q0 and measurement noise covariance matrix R0 are calculated using the classical noise covariance estimation method. At the same time, the initial consistency threshold (NEES confidence interval) is calculated based on the filter residuals to provide a benchmark for online self-evaluation.

[0037] Lightweight training of subjective logic self-evaluation model: Based on the filtering state and sensor data of historical data, train the basic model of subjective logic self-evaluation opinion triple (belief b, uncertainty u, base rate a), determine the base rate distribution a of each sensor and filtering state (based on prior knowledge of sensor reliability in industrial field); pre-calculate the conflict degree (DC) baseline threshold under different working conditions, and simplify the calculation process of subjective logic conflict degree, trust correction, etc. (such as simplifying matrix operation to scalar operation and pre-storing core calculation parameters) to ensure online calculation latency ≤10ms and adapt to the computing power of edge controller.

[0038] S2: Key Technology Implementation of Online Subjective Logic Self-Evaluation Layer Edge multi-sensor data real-time acquisition and preprocessing: Through the local acquisition module of the edge controller, multi-sensor data is acquired in real time (the acquisition frequency is consistent with the edge control period, 10 - 100 ms). The data is standardized and preliminary outliers are removed according to the parameters of the offline modeling layer (if a certain parameter exceeds the process allowable range, the mean value of the previous 3 sampling periods is used for substitution). At the same time, one-step prediction and residual calculation of the Kalman filter are completed to obtain the filtered prediction value and residual γ k , normalized estimated error squared NEES k .

[0039] Subjective logic self-evaluation of filtering state: Quantify consistency and reliability; Taking the NEES k value and residual variance of the Kalman filter as inputs, construct the subjective logic opinion of the filtering state ω_X=(b_X,u_X,a_X): Belief b_X: Characterize the consistency of the filtering state. When NEES k is within the confidence interval, b_X takes 0.8 - 0.95; when NEES k exceeds the confidence interval, it linearly decreases with the increase of the deviation (0 - 0.8); Uncertainty u_X: Characterize the unknown degree of the filtering state. When the working condition is stable, u_X takes 0.05 - 0.2; when the working condition fluctuates, it linearly increases with the increase of the residual variance (0.2 - 0.5); Base rate a_X: Use the prior value pre-stored in the offline modeling layer, fixed at 0.5. Map the opinion ω_X to the reliability probability P_X of the filtering state through projection probability calculation. If P_X≥0.8, it is judged as "filtering normal"; if 0.5≤P_X<0.8, it is judged as "filtering deviation"; if P_X<0.5, it is judged as "filtering abnormal".

[0040] Subjective logic self-evaluation of multi-sensor data: Quantify reliability and conflict: For the real-time data of each sensor, based on the data variance and the deviation from the filtered prediction value, construct the subjective logic opinion of the sensor ω_S⁽ˢ⁾=(b_S⁽ˢ⁾,u_S⁽ˢ⁾,a_S⁽ˢ⁾) (s is the sensor number, s = 1,2,...,V), where b_S⁽ˢ⁾ characterizes the reliability of the sensor data, u_S⁽ˢ⁾ characterizes the uncertainty of the sensor data, and a_S⁽ˢ⁾ is the base rate of the sensor reliability pre-stored offline. Calculate the conflict degree DC(S¹,S²) between sensors and the conflict degree DC(S⁽ˢ⁾,X) between the sensor and the filtering state, and quantify the data consistency between sensors and the matching degree between the sensor data and the filtering state based on the conflict degree: If DC≤0.3, it is judged as "no conflict"; if 0.3<DC≤0.6, it is judged as "mild conflict"; if DC>0.6, it is judged as "severe conflict".

[0041] Accurate Identification of Filter Deviation Root Causes: By fusing filter state self-assessment results, sensor self-assessment results, and conflict indexes, logical judgment rules are established to accurately identify the root causes of filter deviations. Filtering deviation + severe conflict between a few sensors and the filtering status + no conflict between sensors → abnormal measurement noise of the corresponding sensor. Filtering deviation + severe conflict between most sensors and the filtering status + no conflict between sensors → abnormal process noise; Filter deviation + severe conflict between a sensor and other sensors + severe conflict with filter status → sensor malfunction; Filtering deviation + slight conflict between all sensors + slight conflict with filtering status → overall fluctuation in operating conditions.

[0042] S3: Key Technology Implementation of Adaptive Parameter Correction Layer for Operating Conditions: Real-time identification of industrial operating conditions: Based on the variance change rate and the change trend of filter residuals of multi-sensor data, combined with the operating condition clustering results of the offline modeling layer, real-time identification of industrial operating conditions is realized, which are divided into three categories: stable operating conditions, slow drifting operating conditions, and instantaneous change operating conditions, providing a basis for the dynamic adaptation of self-evaluation indicators and correction strategies.

[0043] Dynamic adaptation of self-assessment indicators: matching the needs of different working conditions; Based on the operating condition identification results, dynamically adjust the conflict degree threshold and the filtering reliability probability threshold of the subjective logic self-evaluation: Stable operating conditions: Standard thresholds are used to ensure evaluation accuracy; Slow drift condition: Reduce the collision threshold (0.2), increase the filter reliability probability threshold (0.85), improve self-evaluation sensitivity, and capture small deviations in a timely manner; Transient change conditions: Increase the conflict threshold (0.5), decrease the filter reliability probability threshold (0.7), reduce the self-evaluation sensitivity, and avoid misjudgment caused by oversensitivity.

[0044] Lightweight filter parameter adaptive correction: based on self-evaluation quantization index; Based on the subjective logic of Trust Correction (TR), and combined with self-evaluation quantification indicators (filter reliability probability P_X, sensor conflict degree DC, belief b), a correction factor RF is calculated to achieve lightweight correction of the process noise covariance matrix Q and the measurement noise covariance matrix R. The correction process uses scalar multiplication to avoid complex matrix operations and is adaptable to edge computing power. Measurement noise correction: For the identified abnormal sensor s*, calculate its confidence correction factor RF_R⁽ˢ*⁾=(1-RW), where RW is the correction weight, which is jointly determined by the conflict degree between the sensor and the filter state DC(S⁽ˢ*⁾,X) and the filter belief b_X. After correction, R⁽ˢ*⁾_new=RF_R⁽ˢ*⁾×R⁽ˢ*⁾_old (RF_R>1 indicates increased measurement noise, RF_R<1 indicates decreased measurement noise). Process noise correction: When a process noise anomaly is identified, the process noise correction factor RF_Q=(1-RW_X) is calculated, where RW_X is determined by the filter reliability probability P_X and the residual variance. After correction, Q_new=RF_Q×Q_old. Dynamic adjustment of correction step size: Under stable operating conditions, the correction step size is 0.1-0.3 to ensure parameter stability; under slow drift conditions, it is 0.4-0.6 to achieve gradual correction; under instantaneous change conditions, it is 0.7-0.9 to achieve rapid adaptation.

[0045] S4: Key technologies for robust control adaptation layer implementation: Smoothing of filtered state estimates: The state estimates after Kalman filtering are smoothed using a 6-point window averaging method (taking the most recent 6 estimates, removing the maximum and minimum values, and then averaging them) to avoid instantaneous fluctuations in the estimates caused by parameter correction and to provide stable input data for edge control.

[0046] Self-assessment results with fault tolerance and alarm output: A continuous sampling judgment rule is set for the self-evaluation quantitative indicators (filter reliability probability P_X, sensor conflict degree DC): only when the indicator exceeds the threshold for 3 consecutive sampling cycles will it be judged as "filter abnormality / sensor failure" and a graded alarm signal (slight deviation / severe deviation / sensor failure) will be output to the edge controller to avoid false alarms and control adjustments caused by instantaneous abnormalities.

[0047] Coordinated adaptation of filtering and edge control: The quantitative indicators of subjective logic self-evaluation (filter reliability probability P_X, sensor reliability probability P_S⁽ˢ⁾) are output to the edge control module to provide a basis for adjusting the control strategy: When P_X≥0.8, the control strategy adopts conventional parameters; when 0.5≤P_X<0.8, the control strategy increases the correction step size to improve anti-interference capability; when P_X<0.5, the control strategy switches to fault-tolerant mode, using multi-sensor fusion data for decision-making to avoid control errors caused by filtering anomalies.

[0048] Compared to existing adaptive Kalman filtering algorithms in industrial edge control scenarios, this technology, based on subjective logic self-evaluation, has the following core advantages, fully meeting the "high precision, low latency, lightweight, and high robustness" requirements of industrial edge control: By introducing subjective logic to achieve quantitative self-evaluation of the filter state and multi-sensor data, the limitations of the traditional "threshold judgment" are overcome. It can accurately identify the root cause of filter deviation (process noise / measurement noise / sensor failure), providing a clear basis for parameter correction. The state estimation error is reduced by more than 60%, and the edge control accuracy is improved by more than 5%.

[0049] Modeling and lightweight simplification are completed offline, while online computation uses scalar operations to replace complex matrix operations. The overall execution latency of self-evaluation and parameter correction is ≤10ms, which meets the cycle requirements of 10-100ms for industrial edge control and can be directly deployed on low-cost edge controllers.

[0050] Balancing drift and sudden changes: Based on the dynamic adjustment of self-evaluation indicators and correction step size according to the working condition identification, it can simultaneously cope with the slow drift and instantaneous changes in the industrial field, avoiding the problems of insufficient or excessive sensitivity of evaluation, and improving the accuracy and timeliness of parameter correction by more than 70%.

[0051] By considering the conflict and consistency analysis among multiple sensors, the self-evaluation results are not distorted due to the anomaly of a single sensor. The robustness of the filtering results is improved by more than 80%, effectively reducing the probability of control decision errors.

[0052] The design incorporates a smoothing and self-evaluation fault-tolerance mechanism to prevent actuator "chattering." Simultaneously, the self-evaluation index is output to the control module, achieving synergistic optimization of filtering and control. This extends the service life of industrial equipment by more than 30% and significantly improves control stability.

[0053] The present invention has been described by the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. An adaptive Kalman filter edge control method based on subjective logic self-evaluation, characterized in that: The specific steps of the control method are as follows: S1: Offline lightweight modeling: For industrial historical data preprocessing and operating condition clustering, the adaptive Kalman filter parameters are initialized, and a lightweight subjective logic self-evaluation model is trained. S2: Online Subjective Logic Self-Assessment: Real-time acquisition and preprocessing of multi-sensor data; evaluation of filter state reliability and sensor data reliability based on subjective logical opinion triplets; calculation of conflict degree between sensors and between sensors and filter state; and fusion determination of the root cause of filter deviation. S3: Adaptive parameter correction for operating conditions: Real-time identification of operating condition types, dynamic adaptation of self-evaluation thresholds, and lightweight adjustment of process noise covariance matrix Q and measurement noise covariance matrix R based on subjective logical trust, and dynamic setting of correction step size according to operating conditions. S4: Robust control adaptation: The filtered state estimate is smoothed, the self-evaluation index is continuously and periodically fault-tolerantly determined, and the quantitative evaluation result is output to the edge control module to coordinately adjust the control strategy.

2. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The working condition clustering described in S1 adopts a lightweight K-means algorithm with a limited number of 3–5 categories. The core working condition parameters, filter residuals, and sensor data variance are used as clustering features, and the optimal number of clusters is determined by the elbow rule.

3. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The subjective logical opinion triplet described in S2 includes belief b, uncertainty u, basic law a, and filtered state opinion ω. X =(b X ,u X ,a X The filter reliability probability P is obtained by projecting the probability. X Press P X The filtering is classified as normal, biased, or abnormal.

4. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The collision degree DC mentioned in S2 is divided into inter-sensor collision degree DC (S 1 ,S 2 ) and sensor-filter conflict degree DC (S （S） (X), according to DC≤0.3, 0.3<DC≤0.6、DC> The 0.6 rating is divided into no conflict, minor conflict, and major conflict.

5. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The deviation root cause identification rules described in S2 include: a) Filter deviation + severe conflict between a few sensors and the filter + no conflict between sensors → abnormal measurement noise; b) Filter deviation + severe conflict between most sensors and the filter + no conflict between sensors → abnormal process noise; c) Filtering deviation + severe internal and external conflict of a single sensor → sensor failure; d) Filtering deviation + global mild conflict → overall fluctuation in operating conditions.

6. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The operating conditions described in S3 include stable operating conditions, slow drift operating conditions, and instantaneous change operating conditions, which correspond to dynamically adjusting the conflict degree threshold and the filter reliability probability threshold.

7. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The lightweight correction described in S3 uses scalar multiplication, with measurement noise R and process noise Q adjusted by the confidence correction factor RF. R RF Q Perform a proportional update.

8. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The correction step size described in S3 is: 0.1–0.3 for stability, 0.4–0.6 for slow drift, and 0.7–0.9 for transient changes.

9. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The smoothing process described in S4 uses a 6-point window averaging method with head and tail removed; the fault tolerance determination is to determine an anomaly and output an alarm only when the threshold is exceeded for three consecutive sampling periods.

10. The adaptive Kalman filter edge control method based on subjective logic self-evaluation as described in claim 1, characterized in that: The collaborative adjustment strategy described in S4: P X ≥0.8 uses standard parameters; 0.5≤P X <0.8 Increase the correction step size; P X <0.5 Switch to multi-sensor fusion fault-tolerant control mode.

Images