Airborne multi-sensor information fusion method based on interactive multi-model algorithm

Abstract

An airborne multi-sensor information fusion method based on an interactive multiple model algorithm, after data preprocessing and polynomial fitting and down-sampling, an interactive multiple model (IMM) structure containing multiple parallel sub-filters is established, each sub-filter corresponds to different working modes in which the aircraft sensor may be in and at the initialization, the initial probability of each mode is set and each sub-filter is run in parallel, at the beginning of each time cycle, interactive mixing and parallel Kalman filtering are carried out in turn; in each filtering cycle, according to the residual error and its covariance updated in step 2, the posterior probability of each model is updated in real time and the failure judgment and failure detection logic are executed, and then the posterior weighted fusion output is carried out. The model probability is used as the dynamic confidence in the application, the sensitive detection and smooth isolation of the sensor soft fault are realized, the signal jump problem caused by the existing hard threshold switching is solved, and the filtering smoothness and the maneuvering response speed are considered at the same time while adapting to the airborne platform with limited computing power.

Description

Technical Field

[0001] This invention relates to a technology in the field of aircraft control, specifically an airborne multi-sensor information fusion method based on the interactive multi-model (IMM) algorithm. Background Technology

[0002] In commercial aircraft automatic flight control systems, to ensure high safety, redundant sensor architectures (such as triple-redundant inertial navigation systems and air data systems) are typically used to measure aircraft status information. Existing IMM-based technologies have excessive computational loads and are sensitive to high-frequency oscillation noise. They cannot achieve smooth weighted transitions when faults occur, are difficult to avoid signal jumps, and lack clear fault discrimination logic and smoothing mechanisms, making them unable to effectively identify sensor soft faults. Summary of the Invention

[0003] This invention addresses the problems of existing hard threshold voting methods having low sensitivity to soft faults, making it difficult to detect slow degradation or slight drift in sensor performance, leading to erroneous signals entering the control system before the threshold is triggered; simple logic switching easily causing output signal jumps or high-frequency jitter when sensor noise is high or the sensor is in a critical fault state, affecting the stability of the autopilot; and existing technologies usually rely on a single low-pass filter, which is difficult to effectively filter out high-frequency noise while ensuring rapid dynamic response to aircraft maneuvers. To address these issues, this invention proposes an airborne multi-sensor information fusion method based on an interactive multi-model algorithm. This method is deployed in a cascaded signal processing architecture and controls each independent sensor channel of the automatic flight control system. It includes three core stages: data preprocessing, IMM filter estimation, and fault determination and fusion. It not only outputs a smoothed filtered signal but also outputs the health confidence of the channel in real time for use by the downstream redundancy management system.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to an airborne multi-sensor information fusion method based on an interactive multi-model algorithm, comprising:

[0006] Step 1, data preprocessing and polynomial fitting downsampling, specifically includes:

[0007] 1.1 Acquire high-frequency sampling data from raw sensors (such as FCM1 pitch angle, calibrated airspeed, etc.).

[0008] 1.2 To address the unique high-frequency noise characteristics of commercial aircraft sensors, an N-point n-order polynomial fitting algorithm is used to perform sliding window smoothing on the high-frequency sampling data collected in step 1.1. Specifically, the input data stream is divided into independent segments with continuous time, each segment consisting of N points; the least squares method is used to fit the data points within a single segment, and the center point of the fitted curve is used to replace the original noise points.

[0009] 1.3 After fitting, the data is significantly downsampled. The downsampled frequency satisfies the Nyquist sampling law and is higher than the calculation cycle frequency of the aircraft flight control law. Under the premise of ensuring no spectral aliasing and preserving the core dynamic characteristics of the signal, high-frequency measurement noise is effectively removed, significantly reducing the computational load of the subsequent IMM algorithm and enabling it to meet the real-time requirements of the airborne computer.

[0010] Step 2, Interactive Input and Parallel Filtering Estimation, specifically involves establishing an interactive multiple model (IMM) structure containing multiple parallel sub-filters. Each sub-filter corresponds to a different operating mode that the aircraft sensor may be in, and during initialization, the initial probabilities of each mode are set, allowing the sub-filters to run in parallel. At each time step... At the start of the loop, interactive mixing and parallel Kalman filtering are performed sequentially.

[0011] The aforementioned interactive mixing specifically refers to: based on the previous time step The posterior probabilities and Markov transition probability matrices of each sub-model are used to mix the state estimates and covariances of different models from the previous time step to calculate the mixed initial state of each sub-filter at the current time step. and mixed initial covariance Specifically: , , ,in: For state vectors, Let covariance matrix be the variance matrix. From the pattern Transfer to mode The probability, The normalization constant is indicated by the superscript. This indicates the matrix transpose.

[0012] The parallel Kalman filtering specifically involves: using initial values ​​obtained through interactive mixing to drive three parallel Kalman filters, each corresponding to one of the three operating modes of the aircraft sensor, to perform state estimation, including:

[0013] Operating mode one is a normal and stable mode, configured with a small process noise covariance. It is used to characterize the stable flight phase of an aircraft, focusing on filtering accuracy;

[0014] Operating mode two is a mobile mode, configured with a larger process noise covariance. It is used to characterize the phase of an aircraft's attitude change or violent maneuver, with an emphasis on tracking speed.

[0015] Operating mode three is a fault / high noise mode, configured with extremely high measurement noise covariance. It is used to characterize the situation where the sensor is abnormal, fails, or is subjected to strong interference.

[0016] The state estimation mentioned above refers to: each sub-filter using the current measurement value after preprocessing in step 1. Perform state and covariance prediction, calculate residuals and their covariance, and calculate Kalman gain. And update the state estimate Covariance Specifically, this involves utilizing a mixed initial state. and mixed initial covariance Update the time: , ,in: Here is the state transition matrix. The process noise covariance matrix for the corresponding mode; calculate the residuals. covariance ,in: For the current measurement vector, For the observation matrix, To observe the noise covariance matrix.

[0017] Step 3: Posterior update and soft fault detection for each model in the IMM filter bank. Specifically, in each filtering cycle, based on the residuals and covariance updated in Step 2, the posterior probability of each model is updated in real time, and fault judgment and failure detection logic is executed.

[0018] The aforementioned real-time update refers to: assuming the measurement residuals follow a Gaussian distribution, calculating the value of each filter... The likelihood function ,in: This represents the determinant of the matrix. Then, using Bayes' theorem combined with the likelihood function, the probabilities of each mode are updated to obtain the posterior probabilities of each mode at the current time step. .

[0019] The aforementioned fault diagnosis and failure detection refer to: real-time monitoring of the posterior probability of fault / high-noise modes. And calculate the NIS value at the same time, specifically: When the posterior probability The NIS value continues to exceed the preset safety threshold and is still within the chi-square distribution. When the value is within a reasonable range, the sensor is determined to be in a state of performance degradation or soft fault. If the posterior probability of mode three... The value is relatively high, and the NIS value consistently exceeds the chi-square distribution. If the upper bound of the confidence interval is reached, it is determined that the sensor has experienced a hard fault or has completely failed.

[0020] When a sensor is in a state of performance degradation or soft failure, the signal is not directly cut off, but the weight of the sensor in subsequent redundancy fusion is reduced.

[0021] The aforementioned safety threshold was determined statistically through Monte Carlo simulation experiments based on the system's requirements for the false alarm rate.

[0022] Step 4: Posterior weighted fusion output, specifically: based on the posterior probabilities of each mode calculated in Step 3. State estimates for each sub-filter By performing a weighted summation, we obtain the final estimate of the current channel. At the same time, output the health confidence of the channel.

[0023] When a fault is determined to be a soft fault during fault diagnosis and failure detection, the signal is not directly cut off. Instead, the signal is adjusted using the probability weighting mechanism inherent in the IMM. The increase automatically and smoothly reduces the weight of the sensor in the final output; if a hard fault is determined, a fault flag is output for physical removal by the downstream redundancy management system. Technical effect

[0024] This invention utilizes soft redundancy management based on IMM model probabilities, using the model probabilities output by the IMM algorithm as a dynamic confidence index for sensor health, enabling sensitive detection of soft faults. A fitting-downsampling cascade based on prior spectrum analysis is introduced at the IMM input. Compared to existing technologies, this invention allows the automatic flight control system to perform weighted smooth transitions among three redundant sensors based on this dynamic confidence. When a sensor detects a soft fault (increased probability in Mode 3), its weight in the overall control law calculation automatically and smoothly decreases, thus avoiding signal jumps and unnecessary control surface jitter. Simultaneously, polynomial fitting suppresses high-frequency oscillation noise, solving the technical problem of existing IMMs easily diverging in extremely high-frequency noise environments. Furthermore, the downsampling strategy effectively balances the high computational overhead of the IMM's multi-filter parallel architecture, making it adaptable to airborne embedded computing platforms with limited computing power. The downsampling frequency still satisfies the Nyquist sampling theorem and is higher than the calculation frequency of the aircraft control law. Attached Figure Description

[0025] Figure 1 This is a flowchart of the present invention;

[0026] Figure 2 This is a schematic diagram of the system in the embodiment;

[0027] Figure 3 This is a schematic diagram illustrating the effect of an example. Detailed Implementation

[0028] like Figure 2As shown in this embodiment, an airborne multi-sensor information fusion system based on an interactive multi-model algorithm is included, comprising: a data preprocessing module, an IMM filtering estimation module, and a fault determination and fusion module. Specifically: the data preprocessing module performs polynomial fitting smoothing and downsampling processing on the high-frequency raw data collected by the sensors to obtain low-frequency preprocessed data that retains core dynamic characteristics and satisfies the Nyquist sampling theorem. The IMM filtering estimation module performs interactive mixing and parallel Kalman filtering processing based on the posterior probabilities of each mode at the previous time step and the preprocessed data at the current time step to obtain the state estimates, covariance matrices, and measurement residuals of each sub-filter. The fault determination and fusion module performs posterior probability updates, NIS fault detection, and weighted summation based on the residuals, state estimates, and likelihood functions output by the IMM filtering estimation module to obtain the final state estimates, the health confidence of the sensor channels, and the discrimination results for both soft and hard faults.

[0029] The data preprocessing module includes a data acquisition unit, a polynomial fitting unit, and a downsampling unit. The data acquisition unit acquires high-frequency data from analog or digital signals output by airborne sensors to obtain an original high-frequency sampled data sequence. The polynomial fitting unit performs N-point n-order least squares fitting within a sliding window on the original high-frequency sampled data sequence to obtain smooth data points that suppress high-frequency noise. The downsampling unit performs frequency decimation processing based on the smoothed data points, satisfying the Nyquist sampling theorem, to obtain low-frequency preprocessed data that retains core dynamic characteristics.

[0030] The IMM filtering estimation module includes an input interaction unit and a parallel filtering unit. The input interaction unit performs model interaction mixing processing based on the Markov transition probability matrix according to the posterior probability of each mode at the previous time and the state estimate at the previous time to obtain the mixed initial state and mixed initial covariance of each sub-filter at the current time. The parallel filtering unit drives the three sub-filters corresponding to the stationary mode, the maneuvering mode and the fault mode to perform parallel Kalman filtering iterative processing according to the mixed initial state, the mixed initial covariance and the low-frequency preprocessed data to obtain the state estimate, covariance matrix, measurement residual and innovation covariance of each mode.

[0031] The fault determination and fusion module includes a probability update unit, a fault detection unit, and a weighted fusion unit. Specifically: the probability update unit performs likelihood function calculation and Bayesian probability update processing based on the measurement residuals and innovation covariance to obtain the posterior probability of each mode at the current time; the fault detection unit performs dual logical threshold discrimination processing for soft and hard faults based on the posterior probability of the fault mode and the normalized innovation square (NIS) value to obtain the health status flag of the sensor channel (including healthy, soft fault, or hard fault); and the weighted fusion unit performs probability weighted summation processing based on the posterior probability of each mode at the current time and the state estimate under each mode to obtain the final fused state estimate and channel health confidence.

[0032] like Figure 1 As shown in the figure, this embodiment relates to an airborne multi-sensor information fusion method based on an interactive multi-model algorithm, including:

[0033] Step 1, data preprocessing and polynomial fitting downsampling, specifically includes:

[0034] 1.1 Acquire high-frequency sampling data from raw sensors (such as FCM1 pitch angle, calibrated airspeed, etc.).

[0035] 1.2 To address the unique high-frequency noise characteristics of commercial aircraft sensors, an N-point n-order polynomial fitting algorithm is used to perform sliding window smoothing on the high-frequency sampling data collected in step 1.1. Specifically, the input data stream is divided into independent segments with continuous time, each segment consisting of N points; the least squares method is used to fit the data points within a single segment, and the center point of the fitted curve is used to replace the original noise points.

[0036] 1.3 After fitting, the data is significantly downsampled.

[0037] Step 2: Interactive Input and Parallel Filtering Estimation. An interactive multiple model (IMM) structure is established, comprising multiple parallel sub-filters. Each sub-filter corresponds to a different operating mode that the aircraft sensor may be in. During initialization, the initial probabilities of each mode are set, and the sub-filters run in parallel. At each time step... At the start of the loop, the interactive mixing and parallel Kalman filtering of the model are executed sequentially, specifically including:

[0038] 2.1 Based on the previous moment The posterior probabilities and Markov transition probability matrices of each sub-model are used to mix the state estimates and covariances of different models from the previous time step to calculate the mixed initial state of each sub-filter at the current time step. and mixed initial covariance .

[0039] 2.2 Based on the mixed initial state and mixed initial covariance, three parallel Kalman filters are driven using the current measurement values ​​after preprocessing in step 1. Calculate Kalman gain And update the state estimate Covariance .

[0040] Step 3: Posterior update and soft fault detection for each model in the IMM filter bank. Specifically, within each filtering cycle, based on the residuals and their covariance updated in Step 2, the posterior probability of each model is updated, and fault judgment and failure detection logic is executed. This includes:

[0041] 3.1 Assuming the measurement residuals follow a Gaussian distribution, calculate the values ​​for each filter. The likelihood function To obtain the posterior probabilities of each mode at the current time. ;

[0042] 3.2 Posterior probability of real-time monitoring of fault / high-noise mode And calculate the NIS value at the same time, specifically: When the posterior probability The NIS value continues to exceed the preset safety threshold and is still within the chi-square distribution. When the value is within a reasonable range, the sensor is determined to be in a state of performance degradation or soft fault. If the posterior probability of mode three... The value is relatively high, and the NIS value consistently exceeds the chi-square distribution. If the upper bound of the confidence interval is reached, it is determined that the sensor has experienced a hard fault or has completely failed.

[0043] The aforementioned safety threshold was determined statistically through Monte Carlo simulation experiments based on the system's requirements for the false alarm rate.

[0044] Step 4: Based on the posterior probabilities of each pattern calculated in Step 3 State estimates for each sub-filter By performing a weighted summation, we obtain the final estimate of the current channel. And the health confidence level of the channel.

[0045] If the fault is determined to be a soft fault, the system does not directly cut off the signal, but instead uses the probability weighting mechanism inherent in the IMM to adjust the signal accordingly. The increase automatically and smoothly reduces the weight of the sensor in the final output; if a hard fault is determined, a fault flag is output for physical removal by the downstream redundancy management system.

[0046] Through practical application experiments, in a hardware and software environment of an industrial control computer with an Intel Core i7 processor and a MATLAB / Simulink R2022b simulation platform, the airborne multi-sensor information fusion method based on the interactive multi-model algorithm of this invention was run by injecting different types of fault signals into a single-channel real stable flight data stream in a multi-channel configuration. Figure 3 As shown, to verify the dual detection and fault tolerance performance of hardware and software faults proposed in this invention, a simulation experiment with a full duration of 5 seconds was constructed.

[0047] 1. Soft Fault Suppression Verification: At time t=1.2s, a slow drift error of approximately 0.4° was injected into the sensor. The result, as shown in the signal amplitude (top figure), is that although the sensor observation (light blue line) shows a significant shift, the fused output (green line) of this invention still closely follows the true state (gray line), with the maximum tracking error controlled within 0.05°. As shown in the middle figure, the posterior probability of the fault mode (mode three) at this time... The value rapidly increased from 0.02 to 0.98, exceeding the preset alarm threshold (0.8); however, as shown in the figure below, the NIS value only rose to around 18, remaining within the soft fault / fault-tolerant range (the hard fault threshold is 25). Based on this, the system correctly determined it to be a soft fault and automatically reduced the sensor's weight using probability weighting to achieve smooth suppression.

[0048] 2. Hard Fault Clearance Verification: At time t=4.0s, a large-amplitude step signal is injected into the sensor to simulate a hard fault. At this time, the NIS value surges instantaneously and exceeds 100, entering the hard fault / clearance interval. Although the fault mode probability also remains at 1.0, because the NIS value exceeds the upper limit of the chi-square distribution signal, the system preferentially determines it as a hard fault and immediately outputs a hard fault clearance flag (red cross in the diagram) to prevent erroneous large-amplitude signals from entering the downstream control law.

[0049] It should be noted that, since the verification scenario in this embodiment is a stable flight phase, the posterior probability of Mode 2 (maneuver mode) is extremely low and close to zero. For the sake of clarity in the illustration, Figure 3 The text only shows a comparison of the probability evolution between Mode 1 (normal stable mode) and Mode 3 (failure mode).

[0050] Compared with existing technologies, this invention significantly improves system safety, detects minor sensor drift (soft faults) earlier, and prevents erroneous data from entering the flight control loop. Simultaneously, the method improves control quality, ensuring a smooth, jitter-free output signal and effectively reducing unnecessary control surface movements. Furthermore, by incorporating a downsampling fitting strategy, the complex IMM algorithm can be run efficiently and in real-time on an onboard computer with limited computing power.

[0051] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. An airborne multi-sensor information fusion method based on an interactive multi-model algorithm, characterized in that, include: Step 1: Data preprocessing and polynomial fitting downsampling; Step 2, Interactive Input and Parallel Filtering Estimation, specifically involves establishing an interactive multiple model (IMM) structure containing multiple parallel sub-filters. Each sub-filter corresponds to a different operating mode that the aircraft sensor may be in. During initialization, the initial probabilities of each mode are set, and the sub-filters are run in parallel. At each time step... At the beginning of the loop, interactive mixing and parallel Kalman filtering are performed sequentially; Step 3: Posterior update and soft fault detection for each model in the IMM filter bank. Specifically, in each filtering cycle, based on the residuals and their covariance updated in Step 2, the posterior probability of each model is updated in real time and the fault judgment and failure detection logic is executed. Step 4, posterior weighted fusion output, specifically: based on the posterior probabilities of each mode calculated in step 3, the state estimates of each sub-filter are weighted and summed to obtain the final estimate of the current channel, while outputting the health confidence of the channel.

2. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 1, characterized in that, Step 1 specifically includes: 1.1 Acquire the high-frequency sampling data from the raw sensor; 1.2 To address the unique high-frequency noise characteristics of commercial aircraft sensors, an N-point n-order polynomial fitting algorithm is used to perform sliding window smoothing on the high-frequency sampling data collected in step 1.

1. Specifically, the input data stream is divided into independent segments with continuous time, each segment consisting of N points; the least squares method is used to fit the data points within a single segment, and the center point of the fitted curve is used to replace the original noise points. 1.3 After fitting, the data is significantly downsampled.

3. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 1, characterized in that, The aforementioned interactive mixing specifically refers to: based on the previous time step The posterior probabilities and Markov transition probability matrices of each sub-model are used to mix the state estimates and covariances of different models from the previous time step to calculate the mixed initial state of each sub-filter at the current time step. and mixed initial covariance Specifically: , , ,in: For state vectors, Let covariance matrix be the variance matrix. From the pattern Transfer to mode The probability, The normalization constant is indicated by the superscript. This indicates the matrix transpose.

4. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 1, characterized in that, The parallel Kalman filtering specifically involves: using initial values ​​obtained through interactive mixing to drive three parallel Kalman filters, each corresponding to one of the three operating modes of the aircraft sensor, to perform state estimation, including: Operating mode one is a normal and stable mode, configured with a small process noise covariance. It is used to characterize the stable flight phase of an aircraft, focusing on filtering accuracy; Operating mode two is a mobile mode, configured with a larger process noise covariance. It is used to characterize the phase of aircraft attitude change or violent maneuver, focusing on tracking speed; Operating mode three is a fault / high noise mode, configured with extremely high measurement noise covariance. It is used to characterize the situation where the sensor is abnormal, fails, or is subjected to strong interference.

5. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 4, characterized in that, The state estimation mentioned above refers to: each sub-filter using the current measurement value after preprocessing in step 1. Perform state and covariance prediction, calculate residuals and their covariance, and calculate Kalman gain. And update the state estimate Covariance Specifically, this involves utilizing a mixed initial state. and mixed initial covariance Update the time: , ,in: Here is the state transition matrix. The process noise covariance matrix for the corresponding mode; calculate the residuals. covariance ,in: For the current measurement vector, For the observation matrix, To observe the noise covariance matrix.

6. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 1, characterized in that, The aforementioned real-time update refers to: assuming the measurement residuals follow a Gaussian distribution, calculating the value of each filter... The likelihood function ,in: The determinant of the matrix is ​​used, and then Bayes' theorem combined with the likelihood function is used to update the probabilities of each mode, thus obtaining the posterior probability of each mode at the current time. .

7. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 1, characterized in that, The aforementioned fault diagnosis and failure detection refer to: real-time monitoring of the posterior probability of fault / high-noise modes. And calculate the NIS value at the same time, specifically: When the posterior probability The NIS value continues to exceed the preset safety threshold and is still within the chi-square distribution. When the value is within a reasonable range, the sensor is determined to be in a state of performance degradation or soft fault. If the posterior probability of mode three is... The value is relatively high, and the NIS value consistently exceeds the chi-square distribution. If the upper bound of the confidence interval is reached, it is determined that the sensor has experienced a hard fault or has completely failed.

8. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 7, characterized in that, When a sensor is in a state of performance degradation or soft failure, the signal is not directly cut off, but the weight of the sensor in subsequent redundancy fusion is reduced.

9. The airborne multi-sensor information fusion method based on the interactive multi-model algorithm according to claim 7, characterized in that, When a fault is determined to be a soft fault during fault diagnosis and failure detection, the signal is not directly cut off. Instead, the signal is weighted by the probability factoring mechanism inherent in the IMM, based on the posterior probability. The increase automatically and smoothly reduces the weight of the sensor in the final output; if a hard fault is determined, a fault flag is output for physical removal by the downstream redundancy management system.

Images