A sensor failure prediction method and system

Abstract

The application relates to the technical field of industrial equipment health monitoring, and discloses a sensor fault prediction method and system.The method comprises the following steps: collecting sensor output values and working condition operation parameters, and performing filtering processing to obtain a stable signal sequence; comparing the stable signal sequence with historical health benchmarks to generate residual sequence time sequence dependent features; learning abnormal deviation probability density through a preset memory unit; mapping to obtain an initial fault probability score; separating and adjusting environmental disturbance factors to obtain inherent abnormal confidence; and outputting a sensor self-fault alarm signal when the conditions are met.The method can accurately distinguish between environmental disturbance and sensor inherent fault, improve fault prediction accuracy and reliability, and provide support for equipment predictive maintenance.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment health monitoring technology, and in particular to a sensor fault prediction method and system. Background Technology

[0002] Most current IoT chip technology methods for sensor fault detection rely on fixed thresholds or simple statistical models to determine signal anomalies. However, in real-world industrial scenarios, equipment operating conditions frequently change, and external environmental factors such as temperature, vibration, and electromagnetic interference are constantly fluctuating. These changes can cause significant fluctuations in normal sensor signals, leading to false alarms. Furthermore, the slow degradation of the sensor itself is often mixed with short-term environmental disturbances. Existing methods struggle to effectively distinguish between these two different sources of signal deviation. They cannot accurately identify early, weak inherent sensor faults and are prone to overreacting to normal fluctuations caused by environmental disturbances, ultimately resulting in a significant decrease in the reliability of fault alarms.

[0003] In summary, existing technologies cannot accurately distinguish between the actual deviation caused by sensor malfunctions and normal fluctuations caused by external environmental noise and changes in operating conditions in complex and ever-changing industrial environments, making it difficult to balance the sensitivity and accuracy of fault detection. Summary of the Invention

[0004] This invention provides a sensor fault prediction method and system to solve the problem that existing technologies cannot accurately distinguish between the true deviation caused by the sensor's own fault and the normal fluctuations caused by external environmental noise and changes in operating conditions in complex and ever-changing industrial environments, and it is difficult to balance the sensitivity and accuracy of fault detection.

[0005] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a sensor-based fault prediction method, comprising: The sensor output value and corresponding operating parameters are collected in real time, and adaptive Kalman filtering is performed on the sensor output value and corresponding operating parameters to obtain a stable signal sequence. The stationary signal sequence is subjected to point-by-point difference operation with the preset historical health benchmark signal to obtain the residual sequence, and the residual time-series dependency feature is obtained by performing time-series feature extraction operation based on the residual sequence. Based on the residual temporal dependency features, a memory unit temporal learning operation is performed to obtain the residual distribution features, and a probability density calculation operation is performed based on the residual distribution features to obtain the residual probability density. The residual probability density is compared with the preset fluctuation limit to obtain the potential fault labeling result. Based on the potential fault labeling result, the corresponding residual vector is extracted and saved to obtain the residual vector to be analyzed. Based on the residual vector to be analyzed, a feature mapping operation is performed to obtain the preliminary fault probability score. The preliminary fault probability score is compared with a preset warning threshold to obtain a multi-source environmental weight vector. The preliminary fault probability score is then weighted and adjusted according to the multi-source environmental weight vector to obtain an anomaly confidence value. If the abnormal confidence level value is higher than the preset confirmation threshold and the preset stability condition is met, the sensor itself will output a fault alarm signal to complete the fault prediction.

[0006] Secondly, the present invention provides a sensor fault prediction system, comprising: The signal filtering module is used to acquire sensor output values ​​and corresponding operating parameters in real time, and perform adaptive Kalman filtering on the sensor output values ​​and corresponding operating parameters to obtain a stable signal sequence. The feature extraction module is used to perform point-by-point difference calculation on the stationary signal sequence and the preset historical health benchmark signal to obtain the residual sequence, and to perform time-series feature extraction operation on the residual sequence to obtain the residual time-series dependency features. The probability calculation module is used to perform a memory unit time-series learning operation based on the residual time-series dependency features to obtain residual distribution features, and to perform a probability density calculation operation based on the residual distribution features to obtain residual probability density. The preliminary scoring module is used to compare the residual probability density with the preset fluctuation limit to obtain the potential fault labeling result, extract and save the corresponding residual vector according to the potential fault labeling result to obtain the residual vector to be analyzed, and perform feature mapping operation according to the residual vector to be analyzed to obtain the preliminary fault probability score. The confidence adjustment module is used to compare the preliminary fault probability score with a preset warning threshold to obtain a multi-source environmental weight vector, and to perform a weighted adjustment operation on the preliminary fault probability score based on the multi-source environmental weight vector to obtain an abnormal confidence value. The fault alarm module is used to output a fault alarm signal of the sensor itself if the abnormal confidence value is higher than the preset confirmation threshold and the preset stability condition is met, so as to complete the fault prediction.

[0007] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention filters out environmental noise interference by using a Kalman filter, extracts the time-dependent features of the residual sequence by using a preset memory unit, constructs a Gaussian mixture model to calculate the probability density of abnormal deviations, and realizes accurate identification of the early weak degradation trend of the sensor. This solves the problem that traditional fixed threshold or simple statistical models are difficult to distinguish between environmental disturbances and sensor fault deviations, and are not sensitive enough to early faults.

[0008] (2) This invention separates three types of environmental disturbance factors from the operating parameters, calculates the multi-source environmental weight vector and adjusts the preliminary fault probability score by conditional weighting, and removes the influence of environmental interference. This solves the problem that the existing technology has a high false alarm rate and a high false alarm rate due to the mixture of environmental factors, and cannot adapt to complex and variable industrial operating conditions.

[0009] (3) This invention calculates the confidence stability index and cumulative fault duration, and combines the multi-dimensional fault level matrix to classify the severity of faults, generating standardized abnormal status codes and alarm signal frames. This solves the problems of poor alarm reliability and ambiguous fault levels in traditional detection schemes, which make it difficult to support predictive maintenance decisions for equipment. Attached Figure Description

[0010] Figure 1 This is a schematic flowchart of a sensor fault prediction method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a sensor fault prediction system provided in the second embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Reference Figure 1 The first embodiment of the present invention provides a sensor fault prediction method, including the following steps: S101, Real-time acquisition of sensor output values ​​and corresponding operating parameters, adaptive Kalman filtering of sensor output values ​​and corresponding operating parameters to obtain a stable signal sequence; S102, perform point-by-point difference operation on the stationary signal sequence and the preset historical health benchmark signal to obtain the residual sequence, and perform time-series feature extraction operation on the residual sequence to obtain the residual time-series dependency feature; S103, perform memory unit temporal learning operation based on the residual temporal dependency features to obtain residual distribution features, and perform probability density calculation operation based on the residual distribution features to obtain residual probability density; S104, compare the residual probability density with the preset fluctuation limit to obtain the potential fault marking result, extract and save the corresponding residual vector according to the potential fault marking result to obtain the residual vector to be analyzed, and perform feature mapping operation according to the residual vector to be analyzed to obtain the preliminary fault probability score. S105, compare the preliminary fault probability score with the preset warning threshold to obtain a multi-source environmental weight vector, and perform a weighted adjustment operation on the preliminary fault probability score according to the multi-source environmental weight vector to obtain an abnormal confidence value. S106, if the abnormal confidence value is higher than the preset confirmation threshold and the preset stability condition is met, output the sensor's own fault alarm signal to complete the fault prediction.

[0013] In step S101, the sensor output value and corresponding operating parameters are acquired in real time, and adaptive Kalman filtering is performed on the sensor output value and the corresponding operating parameters to obtain a stationary signal sequence, including: The sensor output value and the corresponding operating parameters are used to construct a multi-dimensional state input vector. The multi-dimensional state input vector is preset with a kinematic model to generate a predicted state estimate. The single-point difference between the predicted state estimate and the sensor output value is calculated. The single-point differences at each sampling time are combined and spliced ​​in time stamp order to obtain the measurement residual sequence; A preset initial noise covariance matrix is ​​retrieved, and the initial noise covariance matrix is ​​adaptively iteratively updated using the measurement residual sequence to obtain the updated initial noise covariance matrix. The Kalman gain coefficient is then calculated based on the updated initial noise covariance matrix. The predicted state estimate is corrected using the Kalman gain coefficient, and the true component in the corrected state estimate is extracted to obtain the stationary signal sequence.

[0014] It should be noted that the sensor output values ​​include physical quantity measurements such as velocity, acceleration, and pressure, while the operating parameters include environmental data such as temperature, vibration frequency, and electromagnetic intensity. These two types of data are aligned according to the sampling timestamp and integrated into a multi-dimensional state input vector. The vector dimension is determined based on the sensor type and the number of operating parameters. In a vehicle-mounted scenario, the vector includes eight components: longitudinal acceleration, lateral acceleration, yaw rate, vehicle speed, etc. Each component is labeled with the acquisition time and data unit. The preset kinematic model is selected based on the application scenario; in a vehicle-mounted scenario, a vehicle dynamics model is used. The model parameters include fixed parameters such as vehicle mass, tire rolling radius, and transmission ratio, and these parameter values ​​are obtained through experimental calibration.

[0015] It is worth noting that the measurement residual sequence calculation performs subtraction operations on each corresponding component of the vector, preserving the original data accuracy without additional smoothing or filtering. The initial noise covariance matrix is ​​obtained by statistically analyzing measurement data from the historical health status of the sensor. The parameters are determined by statistically analyzing health status data from the sensor's continuous operation for more than 100 hours, sorting them in ascending order, and taking the 90th percentile value, with a sample size of no less than 5000 groups. The initial measurement noise variance of the vehicle speed sensor is set to 0.25 (km / h). 2 The process noise variance was set to 0.1 (km / h). 2 The initial measurement noise variance of the pressure sensor is set to 0.01 (kPa). 2 The process noise variance was set to 0.005 (kPa). 2 The initial measurement noise variance of the temperature sensor is set to 0.1 (°C). 2 The process noise variance was set to 0.05 (°C). 2 During dynamic updates, if multiple consecutive sampling points of the measurement residual sequence exceed the preset fluctuation range, the noise variance of the corresponding component is increased proportionally. When the fluctuation of the vehicle speed sensor residual exceeds 0.3 km / h for five consecutive sampling points, the measurement noise variance is adjusted to 0.16 (km / h). 2 The preset fluctuation range is determined based on the sensor accuracy and operating conditions. For example, it can be set to allow the vehicle speed sensor to fluctuate by more than 0.3 km / h. The increase ratio can be determined experimentally, such as by adjusting it to 2.5 times the original value, to reflect the increase in measurement uncertainty.

[0016] It should be noted that the Kalman gain coefficient is obtained through matrix operations on the initial noise covariance matrix, following the standard rules of Kalman filtering. The gain coefficient ranges from 0 to 1. A larger coefficient indicates that the system trusts the sensor measurements more, while a smaller coefficient indicates that the system trusts the model predictions more. In the vehicle scenario, the initial value of the Kalman gain coefficient corresponding to vehicle speed is set to 0.55, and it is adjusted to 0.72 when the residual fluctuation increases, achieving dynamic correction of the predicted state estimate. The weighted correction is performed component by component, and the corrected data removes abnormal fluctuation components, retaining the true trend of physical quantity changes.

[0017] It is worth noting that, to further eliminate any outliers that may remain after Kalman filtering, the corrected signal undergoes post-processing. A sliding window is used to calculate the local mean and standard deviation, with a window size of 10 sampling points. If a sampling point value exceeds the mean ± 3 times the standard deviation, it is identified as an outlier and replaced with the mean of the corrected values ​​at adjacent time points. Furthermore, a secondary verification is performed based on the sensor's physical reasonable range. For example, the reasonable range for a vehicle speed sensor is 0~120km / h; values ​​outside this range are directly replaced. After the above processing, the standard deviation of each component in the stationary signal sequence is effectively reduced. For instance, the standard deviation of the original vehicle speed sensor data is 0.8km / h, which is reduced to 0.25km / h after processing, effectively filtering out environmental noise. The output format of the stationary signal sequence is consistent with the multidimensional state input vector.

[0018] It should be noted that the parameter calibration of the preset kinematic model involved scenario testing with 10 identical devices, collecting sensor data under different operating conditions, and fitting the optimal values ​​of the model parameters. The adjustment ratio for the measurement noise variance during dynamic updates was determined experimentally. Fifty sets of sample data containing different operating conditions were selected to test the filtering effect under different ratios. The residual fluctuation amplitude was used as a trigger condition, for example, exceeding 30% of the normal fluctuation range of the physical quantity, to determine the adjustment ratio that effectively tracks changes while maintaining stability.

[0019] For example, in an application scenario of a vehicle-mounted wheel speed sensor, the collected raw vehicle speed data are 72.1 km / h, 71.8 km / h, and 72.5 km / h, corresponding to operating parameters including an ambient temperature of 25℃ and a vibration frequency of 50Hz. After constructing a multi-dimensional state input vector, the predicted state estimates are generated using a preset kinematic model: 71.9 km / h, 71.8 km / h, and 72.0 km / h. The calculated measurement residual sequence is 0.2 km / h, 0.0 km / h, and 0.5 km / h. When the residual sequence fluctuates by more than 0.3 km / h over three consecutive sampling points, a dynamic update is triggered, and the vehicle speed measurement noise variance changes from the initial 0.25 km / h. 2 Adjusted to 0.64 (km / h) 2 Based on this, the Kalman gain coefficient was recalculated to be 0.72. This gain coefficient was used to correct the predicted state estimates, yielding corrected values ​​of 71.92 km / h, 71.8 km / h, and 72.15 km / h. Subsequently, the corrected values ​​underwent post-processing, employing the 3σ principle to remove outliers, ultimately resulting in a stationary signal sequence. The standard deviation of this sequence decreased from 0.8 km / h in the original data to 0.22 km / h, effectively filtering out environmental noise.

[0020] In step S102, a point-by-point difference operation is performed on the stationary signal sequence and a preset historical health benchmark signal to obtain a residual sequence, and a time-series feature extraction operation is performed on the residual sequence to obtain residual time-series dependency features, including: Input the stable signal sequence and the preset historical health benchmark signal, perform point-by-point subtraction operation, and output the current residual sequence; The preset historical health benchmark signal is sliced ​​according to the time span of the current residual sequence, and the autocorrelation function of the slice is calculated. The autocorrelation function is quantified to obtain the inherent fluctuation pattern. The inherent fluctuation law is set as the convolution kernel, and a convolution operation is performed on the current residual sequence to identify and output local feature components; The local feature components are subjected to dimensionality reduction mapping to extract and output the temporal dependency features of the residual sequence.

[0021] It should be noted that the point-by-point subtraction operation is performed aligned with the sampling timestamps of the signals. The value of the stationary signal sequence corresponding to each sampling point is subtracted from the corresponding value of the preset historical healthy baseline signal, and the difference forms the current residual sequence. The values ​​of the residual sequence can be positive or negative, retaining the original calculation precision without truncation or smoothing, thus fully reflecting the deviation between the two types of signals.

[0022] It is worth noting that the preset historical health baseline signal is derived from typical data collected by the sensor during long-term stable operation under standard operating conditions. Standard operating conditions are defined as the operating state of the equipment under no-load conditions, ambient temperature of 20-25℃, vibration frequency ≤30Hz, and electromagnetic intensity ≤10μT. The data collection time is no less than 100 hours, and the sampling frequency is consistent with the current stable signal sequence. The preset historical health baseline signal is stored according to operating condition type, with different operating conditions corresponding to independent baseline signal segments to ensure consistency in the comparison scenario. The preset historical health baseline signal is updated every 30 days. During the update, abnormal data is removed, and abnormal data under standard operating conditions is retained, with a sample size of no less than 10,000 sets.

[0023] It should be noted that the time span of the slice is consistent with the length of the current residual sequence. If the current residual sequence contains 300 sampling points, corresponding to a time span of 6 seconds, then 300 consecutive sampling points are extracted from the historical health baseline signal as a slice. The autocorrelation function calculation is performed step by step according to the lag step, which starts from 0 and increases to 1 / 3 of the slice length. By calculating the correlation coefficient of the signal under different lag steps, the inherent fluctuation law of the historical health baseline signal is quantified. The correlation coefficient ranges from [-1, 1].

[0024] It is worth noting that the construction of the convolution kernel is based on the inherent fluctuation patterns of historical health baseline signals. The correlation coefficients calculated from the autocorrelation function are arranged in order of lag steps to form a one-dimensional convolution kernel. Considering that negative correlation may reflect reverse fluctuations, the absolute value of the correlation coefficient can be taken or only non-negative values ​​can be retained to highlight the positive correlation pattern. The specific choice can be made according to the actual signal characteristics. During the convolution operation, the convolution kernel slides on the current residual sequence with a stride of 1. This value was determined experimentally, and tests were conducted on residual samples from 20 different sensors. The results show that a stride of 1 achieves the best balance between feature recognition efficiency and information integrity, and this stride is universal for different types of sensors without additional adjustment. At each sliding position, the sum of the product of the kernel element and the corresponding segment of the sequence is calculated, and the result forms a local feature component. This component can highlight signal segments consistent with historical health fluctuation patterns while weakening random noise interference.

[0025] It should be noted that the dimensionality reduction mapping uses principal component analysis: a feature matrix is ​​constructed for the local feature components, their covariance matrix is ​​calculated, eigenvalues ​​and eigenvectors are solved, and the top k eigenvectors with the largest eigenvalues ​​are selected as principal component directions. Then, the local feature components are projected onto these directions to obtain k-dimensional temporal features. The value of k is determined through experimental statistics: 50 sets of health signal samples from different sensors are selected, and the eigenvalue distribution of the feature matrix of each sample is calculated. The results show that the cumulative contribution rate of the first two principal components is over 85%, which can completely reflect the temporal evolution of the residuals. Therefore, k=2 is chosen in this embodiment. If the cumulative contribution rate is less than 85% in practical applications, the number of principal components can be appropriately increased until the requirement is met. The principal components obtained after projection can be interpreted in conjunction with domain knowledge. For example, in a vehicle scenario, the first principal component may reflect the overall shift trend of the residuals, and the second principal component may reflect the oscillation phase characteristics.

[0026] For example, in the application scenario of an onboard vehicle speed sensor, the stable signal sequence is the vehicle speed data during constant speed cruising, containing 300 sampling points, with values ​​fluctuating between 71.9 km / h and 72.1 km / h. The preset historical health benchmark signal is the vehicle speed data collected by the sensor under standard operating conditions, with an average value of 72.0 km / h corresponding to 300 sampling points. After performing point-by-point subtraction, the current residual sequence (unit: km / h) exhibits small alternating positive and negative values ​​such as -0.1, +0.1, and -0.05. The historical health benchmark signal is sliced ​​according to the time span of 300 sampling points, and the autocorrelation function of the slice is calculated. The results show that the correlation coefficient remains above 0.75 when lagging by 1-3 sampling points, and decays to below 0.1 after lagging by 8 sampling points. Based on this, the inherent fluctuation pattern of the historical health benchmark signal is quantified.

[0027] Based on this pattern, correlation coefficients with lags of 0-4 steps were selected as 1.0, 0.8, 0.75, 0.5, and 0.3, respectively. After normalization, a 5-point convolution kernel [0.3, 0.7, 1.0, 0.75, 0.35] was constructed to highlight the main fluctuation patterns. A convolution operation was performed on the current residual sequence with a stride of 1, resulting in a local feature component sequence with a dimension of 296×1. The output value after convolving three consecutive positive residual sampling points when the vehicle passes through slight road undulations was 0.45, while the output value after convolving an isolated spike corresponding to random noise was only 0.12, indicating that the convolution kernel effectively enhanced the signal consistent with historical fluctuation patterns. Principal component analysis was performed on the local feature components for dimensionality reduction, and the first two principal components were extracted as temporal features. The first principal component scored 0.38, mainly reflecting the overall shift trend of the residuals; the second principal component scored 0.21, mainly reflecting the oscillation phase characteristics of the residuals.

[0028] In step S103, a memory unit temporal learning operation is performed based on the residual temporal dependency features to obtain residual distribution features, and a probability density calculation operation is performed based on the residual distribution features to obtain residual probability density, including: Input the residual temporal dependency features, update the cell state through the gating operation of the preset memory unit to obtain the updated cell state, filter the output based on the updated cell state, and obtain the long short-term memory hidden state vector containing effective temporal information. The long short-term memory hidden state vector is input into the terminal of a preset long short-term memory neural network for decoding to obtain Gaussian mixture modeling parameters describing the residual statistical distribution. A dynamic probability distribution model is constructed based on the Gaussian mixture modeling parameters. The measured residual values ​​are input into the dynamic probability distribution model to calculate the sum of weighted probability densities, and the residual probability density is obtained.

[0029] It should be noted that the preset memory unit adopts a Long Short-Term Memory (LSTM) network architecture, including a forget gate, input gate, and output gate. The gating operation uses the Sigmoid activation function, and the cell state update uses the Tanh activation function. The input dimension is consistent with the temporal feature dimension. The gating operation controls the flow of information through parameters determined during training. The forget gate determines the degree to which the cell state from the previous time step is retained, the input gate controls the proportion of the current candidate state added to the cell state, and the output gate selects the effective components from the cell states to generate hidden states. All gating parameters are optimized through training, enabling the network to automatically learn temporal dependencies related to fault prediction and filter out noise interference. Cell state updates are performed step-by-step, with each time step corresponding to a temporal feature of a sampling point, ensuring the continuity of information transmission.

[0030] It's worth noting that nonlinear activation maps cell state values ​​to a range of -1 to 1, adjusting the numerical distribution through element-wise operations to highlight the changing trends in temporal features. The dimension of the activated long short-term memory hidden state vector remains consistent with the temporal features; in the vehicle sensor scenario, the dimension is set to 2, with each dimension corresponding to the overall offset trend and oscillation phase characteristics of the residuals. The values ​​in the vector are arranged in chronological order, recording the dynamic evolution trajectory of the temporal features.

[0031] It should be noted that the fully connected layer contains two layers of neurons. The first layer has three times the number of neurons as the temporal feature dimension, while the second layer has the same number of neurons as the total number of Gaussian mixture modeling parameters. During decoding, the fully connected layer transforms the long short-term memory hidden state vector into parameters of multiple Gaussian distributions through weighted summation and nonlinear transformation, including the mean, variance, and weight coefficients of each distribution. These parameters are arranged sequentially to form a complete set of Gaussian mixture modeling parameters.

[0032] It is worth noting that the number of combinations in the Gaussian mixture model can be dynamically determined using the BIC Bayesian Information Criterion. In this embodiment, three Gaussian distributions are used, corresponding to the three states of normal fluctuation, slight deviation, and significant anomaly of the residuals, respectively, with the sum of the weight coefficients of each distribution being 1. The dynamic probability distribution model is composed of multiple Gaussian distributions. When calculating the sum of weighted probability densities, the measured residual values ​​are substituted into each Gaussian distribution to calculate the probability density, which is then multiplied by the corresponding weight coefficient. The sum of all products is the final anomaly deviation probability density. To facilitate threshold setting, this probability density value can be mapped to the 0-0.2 interval using max-min normalization or the Sigmoid function, and further divided as follows: 0-0.08 is normal, 0.08-0.15 is slightly anomaly, and 0.15-0.2 is significantly anomaly. A larger value indicates a higher degree of deviation of the residuals from the normal distribution.

[0033] It should be noted that the gating parameters of the preset memory unit were determined through training. Fifty sets of temporal feature samples from different types of sensors were selected, covering various operating conditions. The samples were divided into training, validation, and test sets in an 8:1:1 ratio. During training, the gating parameters were adjusted to ensure that the preset memory unit could accurately retain valid temporal information and filter out noise interference. The number of neurons in the fully connected layer was determined by testing different combinations. When the number of neurons in the first layer was 6 and the number in the second layer was 9, the decoding accuracy of the modeling parameters met the requirements.

[0034] For example, in the application scenario of an onboard vehicle speed sensor, the residual temporal features contain feature vectors of 300 time steps, each vector having a dimension of 2, representing the overall offset (value 0.38) and oscillation phase (value 0.21) of the vehicle speed residual (unit: km / h), respectively. After inputting these features into a pre-trained preset memory unit, the network automatically extracts fault-related temporal information through gating operations to generate an updated long short-term memory hidden state vector. After performing nonlinear activation on the cell state, the vector value is mapped to the interval between -1 and 1, highlighting the trend of continuous positive offset.

[0035] The updated hidden state vector is input into the fully connected layer, and the modeling parameters are decoded to obtain three Gaussian distributed parameters. The first distribution has a mean of 0.02 km / h and a variance of 0.01 (km / h). 2 The first distribution has a weight of 0.6; the second distribution has a mean of 0.15 km / h and a variance of 0.03 (km / h). 2 The third distribution has a weight of 0.3 and a mean of -0.1 km / h and a variance of 0.02 (km / h). 2 The weight is 0.1. After constructing the dynamic probability distribution model, the measured residual value of 0.25 km / h is input, and the probability densities under each Gaussian distribution are calculated to be 0.04, 0.015, and 0.008, respectively. After weighted summation, the abnormal deviation probability density is 0.06, which exceeds the preset fluctuation limit of 0.05, indicating the existence of potential anomalies.

[0036] In step S104, the residual probability density is compared with a preset fluctuation limit to obtain a potential fault labeling result. Based on the potential fault labeling result, the corresponding residual vector is extracted and saved to obtain the residual vector to be analyzed. Then, a feature mapping operation is performed on the residual vector to be analyzed to obtain a preliminary fault probability score, including: Input the residual probability density and a preset fluctuation limit. If the residual probability density exceeds the preset fluctuation limit after comparison, output a potential fault flag. Based on the potential fault markers, extract and store the corresponding current residual vector, and use the stored current residual vector as the residual vector to be analyzed; Input the residual vector to be analyzed and the historical health baseline residual pattern, calculate the cosine similarity value of each sampling point according to the sampling timestamp order, combine and sort the cosine similarity values ​​according to the timestamp order, and generate and output the similarity distribution sequence. A weighted matrix is ​​constructed based on the similarity distribution sequence. The weighted matrix is ​​used to perform a linear transformation on the residual vector to be analyzed to obtain the transformed residual vector, and the deviation directions of the first three principal components are extracted. The transformed residual vector is projected onto the deviation direction of the first three principal components to obtain the deviation feature vector. The deviation feature vector is then input into a preset probability mapping model for feature mapping processing, and the preliminary fault probability score is output.

[0037] It should be noted that the preset fluctuation limit is determined by statistically analyzing the probability density distribution range of abnormal deviations under historical normal operating conditions. This involves collecting health status data from sensors that have been running continuously for over 100 hours, with a sample size of no less than 5000 sets. The abnormal deviation probability density values ​​at each time point are then sorted in ascending order, and the 95th percentile is used as the fluctuation limit. For example, in a vehicle-mounted scenario, this limit can be set to 0.05. During real-time monitoring, the current abnormal deviation probability density is compared to this limit step by step. If it exceeds this limit, a potential fault marker signal is immediately generated. This signal includes the timestamp of the abnormality and the corresponding probability density value.

[0038] It is worth noting that the current residual vector extraction range is the residual values ​​of 5 sampling points before and after the corresponding time of the marked signal. These are combined in chronological order to obtain a vector containing 11 sampling points, with each element retaining its original physical quantity unit (e.g., km / h for vehicle speed sensors). During storage, the data is categorized and stored in a buffer according to timestamps, with a buffer capacity of 100 vectors. If the buffer exceeds the capacity, the oldest data is overwritten according to the first-in-first-out principle to ensure that recent key anomaly information is preserved.

[0039] It should be noted that the historical health baseline residual pattern is obtained by analyzing health data under standard operating conditions. From the historical health data that matches the current operating conditions, a large number of residual segments are extracted by sliding with the same window length as the current residual vector. All segments are averaged or clustered to obtain a typical vector as the baseline residual pattern under the operating conditions, and stored according to the operating conditions.

[0040] To obtain the local similarity of each sampling point, for each sampling point in the current residual vector, a local sub-vector is formed by taking the m sampling points before and after it. If the number of points at the boundary is insufficient, endpoint values ​​are used to fill the gaps, or only the valid parts are taken. The cosine similarity of this local sub-vector and the corresponding local sub-vector in the reference residual pattern is calculated to obtain the local similarity value for that point. The local similarities of all points are combined in timestamp order to obtain a similarity distribution sequence. The sequence length is the same as the current residual vector. Each value in this sequence reflects the degree of similarity between the local waveform near that position and the reference pattern.

[0041] It is worth noting that the weighted matrix is ​​constructed based on the similarity distribution sequence. Each similarity value in the sequence is used as a diagonal element of the matrix, while the off-diagonal elements are set to 0, forming a diagonal matrix. During matrix transformation, the current residual vector is multiplied element-wise with this diagonal matrix, i.e., dot product, to obtain the transformed residual vector. This strengthens the features of high-similarity regions and weakens the interference of low-similarity regions, laying the foundation for subsequent principal component extraction.

[0042] It should be noted that the principal component deviation directions are determined in advance by analyzing historical health data. Multiple transformed residual vector samples matching the current operating conditions are extracted from the historical health data, their covariance matrices are calculated, and eigenvalues ​​and eigenvectors are solved. The eigenvalues ​​are sorted from largest to smallest, and the top three eigenvectors with a cumulative contribution rate of not less than 85% are selected as the principal component deviation directions. During real-time processing, the current transformed residual vector is projected onto these three directions to obtain the three-dimensional deviation eigenvectors.

[0043] It is worth noting that the deviation feature vector is obtained by performing a dot product operation between the transformed residual vector and the deviation directions of the three principal components, and the three dot product results constitute a three-dimensional deviation feature vector. This vector is input into a preset probability mapping model, and the model outputs a value between 0 and 1 as a preliminary fault probability score. The higher the score, the greater the probability of a fault.

[0044] It should be noted that the probability mapping model is a two-layer fully connected network, with 16 neurons in the first layer and 1 neuron in the second layer. The activation function is Sigmoid, and the loss function is cross-entropy loss. The model input is a three-dimensional deviation feature vector, and the output is a fault probability score from 0 to 1. The model is trained using historical fault data and normal data, with the training, validation, and test sets divided in an 8:1:1 ratio. Training details include a learning rate of 0.001, 1000 iterations, and the Adam optimizer. The training data covers 10 common sensor types, with each type containing 500 fault samples and 500 normal samples. The fault samples are the deviation feature vectors when the sensor actually malfunctions, and the normal samples are the deviation feature vectors during healthy operation. For example, in the application scenario of an onboard wheel speed sensor, the calculated abnormal deviation probability density is 0.06, exceeding the preset fluctuation limit of 0.05, generating a potential fault marker with a timestamp of 1685234567 seconds. Based on the labeled signal, extract the residual values ​​(unit: km / h) of 5 sampling points before and after the corresponding time point to form the current residual vector [-0.04, 0.12, 0.31, -0.09, 0.07, -0.05, 0.15, 0.28, -0.11, 0.06, -0.08] and store it. For each sampling point in the current residual vector, take the two sampling points before and after it to form a local window, and fill the boundary with endpoint values. Calculate the cosine similarity with the local sub-vectors of the corresponding window in the historical health baseline residual pattern to obtain a similarity distribution sequence with values ​​of 0.90, 0.85, 0.80, 0.75, 0.70, 0.65, 0.60, 0.55, 0.50, 0.45, and 0.40 respectively. Based on this sequence, construct a diagonal weighted matrix and perform element-wise weighting on the current residual vector to obtain the transformed residual vector. The transformed residual vector is projected onto the bias directions of the first three principal components extracted from historical data to obtain a three-dimensional bias feature vector [0.72, 0.15, 0.08]. This vector is input into a trained probability mapping model, which outputs an initial fault probability score of 0.85, indicating a high risk of sensor failure.

[0045] In step S105, the preliminary fault probability score is compared with a preset warning threshold to obtain a multi-source environmental weight vector. The preliminary fault probability score is then weighted and adjusted based on the multi-source environmental weight vector to obtain an anomaly confidence value, including: Input a multi-dimensional operating condition parameter sequence containing temperature, vibration, and electromagnetic data, and decouple the multi-dimensional operating condition parameter sequence to obtain three types of disturbance components; An environmental disturbance energy spectrum matrix is ​​constructed based on the three types of disturbance components. The information entropy value of the environmental disturbance energy spectrum matrix is ​​calculated to obtain the disturbance intensity coefficient. The disturbance intensity coefficients are combined to obtain a set of disturbance intensity coefficients. The set of disturbance intensity coefficients is then normalized to output a multi-source environmental weight vector. The preliminary fault probability score is weighted and attenuated using the multi-source environmental weight vector to output the environmental correction probability value. Calculate the ratio of the difference between the preliminary fault probability score and the environmental correction probability value, perform an activation transformation on the ratio of the difference, and output the environmental disturbance confidence factor. The environmental interference confidence factor is used to perform a reverse weighted correction on the environmental correction probability value, and the anomaly confidence value is output.

[0046] It should be noted that the multidimensional operating condition parameter sequence is collected aligned with the sampling timestamp, including three types of data: temperature (unit: °C), vibration frequency (unit: Hz), and electromagnetic intensity (unit: μT), ensuring temporal consistency. Since these three types of physical quantities are inherently independent, they can be directly used as environmental disturbance components. If coupling information exists in the sensor operating condition parameters, it can be decoupled using wavelet transform. In this embodiment, the original measured values ​​are directly used as disturbance components, and the original acquisition accuracy of each component is preserved without additional smoothing processing.

[0047] It is worth noting that an environmental disturbance energy spectrum matrix is ​​constructed to quantify the intensity of environmental disturbances. First, the values ​​of the three types of disturbance components at each sampling time are squared to obtain the instantaneous energy value of each component at that time. For temperature, the squared deviation from the reference value can be used. The rows of the matrix correspond to the time sampling points (e.g., processing 100 sampling points each time), and the columns correspond to the three types of disturbance components. The matrix elements are the energy values ​​at the corresponding times. Then, the information entropy of the energy sequence for each type of disturbance component is calculated: the energy values ​​are divided into several equally spaced intervals, such as 10 intervals. The frequency of samples appearing in each interval is counted, the probability distribution is calculated, and then the information entropy value of that component is obtained according to the information entropy formula, serving as the disturbance intensity coefficient for that component. The larger the entropy value, the more dispersed the energy distribution of the disturbance component, and the higher the disturbance intensity.

[0048] It should be noted that dividing the three types of disturbance intensity coefficients (i.e., information entropy values) by the maximum value of the three yields three normalized coefficients ranging from 0 to 1, forming a multi-source environmental weight vector W=[w1,w2,w3]. This vector reflects the relative intensity of each environmental disturbance component. Next, a weighted attenuation operation is performed on the preliminary fault probability score Pinit to obtain the environmental correction probability value Penv. In this embodiment, the maximum value of the weight vector is taken as the comprehensive environmental disturbance intensity, i.e., m=max(w1,w2,w3), then Penv=Pinit×(1-m). This operation ensures that Penv≤Pinit, and that Penv significantly decreases when the environmental disturbance is strong.

[0049] It is worth noting that the difference ratio The calculation involves dividing the difference between the initial failure probability score and the environmentally corrected probability value by the initial failure probability score. The result reflects the probability attenuation ratio caused by environmental disturbances. The activation transform employs a nonlinear transformation to enhance the discriminative power of moderate attenuation ratios and weaken the influence of extreme attenuation ratios. The final output is an environmental disturbance confidence factor, ranging from 0 to 1. A Sigmoid function is obtained through fitting experimental data. , where parameters and The determination was made by fitting 30 sets of samples under different environmental disturbance scenarios using the least squares method.

[0050] It should be noted that the inverse weighted correction multiplies the environmental correction probability value by the difference between 1 and the environmental interference confidence factor to obtain the anomaly confidence value. This calculation process isolates the influence of environmental disturbances and highlights the anomaly probability caused by sensor malfunction; the larger the value, the higher the probability of sensor malfunction.

[0051] It should be noted that the preset warning and confirmation thresholds are set using health status data from sensors that have been running continuously for over 100 hours with a sample size of no less than 5000 sets. The 90th percentile of the initial fault probability score is used as the warning threshold (0.6 in a vehicle-mounted scenario), and the 95th percentile is used as the confirmation threshold (0.7 in a vehicle-mounted scenario). When the initial fault probability score is below the warning threshold, it is considered that there is no obvious anomaly, and no environmental disturbance analysis is required. In this case, the initial fault probability score can be directly used as the anomaly confidence value. When the initial fault probability score is above the warning threshold, environmental disturbance analysis is initiated. The multi-source environmental weight vector is calculated and weighted according to the aforementioned steps to obtain the anomaly confidence value. This comparison and judgment ensures that environmental correction is only performed in cases of potential anomalies, saving computational resources.

[0052] In step S106, if the abnormal confidence value is higher than a preset confirmation threshold and a preset stability condition is met, a sensor fault alarm signal is output to complete the fault detection, including: Input the historical sequence of the abnormal confidence scores, calculate the dispersion and trend slope of the historical sequence, and combine the dispersion and trend slope to obtain the confidence stability index. If the abnormal confidence value is higher than the preset confirmation threshold and the confidence stability index is in the convergence range, then the time of the state where the abnormal confidence value is higher than the preset confirmation threshold is accumulated, the duration of the fault is counted and output. The range of fault duration is set as the row dimension, and the amount by which the anomaly confidence value exceeds the preset confirmation threshold is set as the column dimension, thus constructing a multidimensional fault level matrix; The duration of the fault and the extent by which the abnormal confidence value exceeds the preset confirmation threshold are mapped to a multi-dimensional fault level matrix. Each intersection region in the multi-dimensional fault level matrix corresponds to a unique fault severity level identifier. The fault severity level identifier is extracted and output. An abnormal status code is generated based on the fault severity level identifier, and a fault alarm signal frame is constructed and output using the abnormal status code.

[0053] It should be noted that the historical sequence of outlier confidence scores is selected from the values ​​of the most recent 20 sampling points, arranged in chronological order. The dispersion is calculated by determining the standard deviation of the sequence; a smaller standard deviation indicates less numerical fluctuation. The trend slope is obtained by linearly fitting the sequence values; a slope close to 0 indicates no significant upward or downward trend. Both factors together constitute the confidence stability index. The calculation process retains two decimal places to ensure the accuracy of the index meets the judgment requirements.

[0054] It is worth noting that the preset confirmation threshold is determined by statistically analyzing the confidence levels of historical health sensors. Values ​​from 50 sensors operating normally over the past 6 months are selected, sorted in ascending order, and the 95th percentile is used as the threshold. In the vehicle scenario, this threshold is set to 0.7. The convergence interval is calibrated experimentally. The convergence range of the trend slope is [-0.01, 0.01], and the convergence range of the dispersion is [0, 0.05]. If both conditions are met, the stability index is deemed to meet the requirements. The preset stability condition is that the trend slope of the confidence sequence ∈ [-0.01, 0.01] and the dispersion ∈ [0, 0.05].

[0055] It should be noted that the fault duration is accumulated starting from the moment the confidence level first exceeds the preset confirmation threshold and the stability index enters the convergence interval, measured in seconds. The accumulated duration is updated with each sample until the value falls below the threshold or the stability index exceeds the convergence interval, at which point accumulation stops. The magnitude by which the confidence level exceeds the threshold is calculated by subtracting the preset confirmation threshold from the current confidence level, with the result rounded to two decimal places, reflecting the strength of the anomaly. The experimental statistical basis for dividing the fault duration intervals is as follows: 100 sets of sample data from different fault scenarios were selected, and the correlation between fault duration and fault severity was analyzed to determine the interval ranges of 0-30 seconds, 31-180 seconds, and 181-3600 seconds. 3600 seconds is considered a reasonable upper limit in industrial scenarios; durations exceeding this limit are counted as 3600 seconds.

[0056] It's worth noting that the rows of the multidimensional fault severity matrix correspond to the fault duration range, the columns correspond to the range of confidence levels exceeding the threshold, and the matrix elements are the corresponding fault severity level identifiers. The duration ranges are divided into 0-30 seconds, 31-180 seconds, and 181-3600 seconds; the amplitude ranges are divided into 0-0.05, 0.06-0.1, and above 0.1. The matrix is ​​divided into 9 overlapping regions, each corresponding to a unique severity level identifier. Specifically, a duration of 0-30 seconds plus an amplitude of 0-0.05 indicates a minor fault (L1), a duration of 0-30 seconds plus an amplitude of 0.06-0.1 indicates a minor fault (L2), and a duration of 0-30 seconds plus an amplitude above 0.1 indicates a moderate fault (M1). A duration of 31-180 seconds with an amplitude of 0-0.05 indicates a minor fault (L2); a duration of 31-180 seconds with an amplitude of 0.06-0.1 indicates a moderate fault (M1); a duration of 31-180 seconds with an amplitude greater than 0.1 indicates a moderate fault (M2). A duration of 181-3600 seconds with an amplitude of 0-0.05 indicates a moderate fault (M1); a duration of 181-3600 seconds with an amplitude of 0.06-0.1 indicates a moderate fault (M2); a duration of 181-3600 seconds with an amplitude greater than 0.1 indicates a severe fault (H1). The matrix adjustment method for different sensors / operating conditions is as follows: based on the importance coefficient of the sensor and the complexity of the operating condition, adjust the threshold for dividing the amplitude range. For scenarios with high importance coefficients or complex operating conditions, reduce the amplitude range threshold by 10%.

[0057] It should be noted that the abnormal status codes are assigned according to the severity level of the fault. Minor faults correspond to a two-digit alphanumeric code, moderate faults to a three-digit code, and severe faults to a four-digit code. The coding rules are preset by the system to ensure that the status codes for different fault levels are uniquely identifiable. For example, minor fault L1 corresponds to code A1, and L2 to code A2; moderate fault M1 corresponds to code B11, and M2 to code B12; and severe fault H1 corresponds to code C111. The fault alarm signal frame contains four core pieces of information: the abnormal status code, the fault start timestamp, the fault duration, and the confidence level. These are concatenated in a fixed order and stored as a binary stream.

[0058] It should be noted that the transmission of fault alarm signal frames uses a dedicated communication channel with a bandwidth of 1Mbps to ensure rapid transmission of the signal frames to the device control unit. The binary stream format of the signal frame includes a 4-byte frame header identifier, length fields for each core information segment, and data fields. A 2-byte checksum field is added to the frame tail for the receiving end to verify the integrity of the signal frame. The checksum field is obtained by calculating the sum of the frame header and data fields. The receiving end calculates and compares the checksum in the same way to ensure error-free signal transmission.

[0059] It should be noted that after the fault alarm signal frame is output, a device maintenance reminder is triggered, and data acquisition of the sensor is paused until sensor calibration / replacement is completed. After calibration, data is reacquired and the reference signal is updated. The sensor fault type differentiation rules are as follows: a confidence level between 0.7 and 0.8 and a fault duration ≤ 180 seconds is considered zero drift; a confidence level between 0.8 and 0.9 and a fault duration 181-3600 seconds is considered sensitivity degradation; a confidence level ≥ 0.9 or a fault duration ≥ 3600 seconds is considered hardware damage. The fault judgment range for different sensors is adjusted according to the sensor's accuracy level. For example, the overall confidence range threshold for high-precision sensors is lowered by 10%, i.e., 0.7 is adjusted to 0.63, 0.8 to 0.72, and 0.9 to 0.81, to differentiate fault types with a stricter standard.

[0060] For example, in the application scenario of an onboard accelerometer, the confidence level is 0.75, and the preset confirmation threshold is 0.7, which is higher than the threshold. The dispersion of its historical 20 sampling point sequences is calculated to be 0.03, and the trend slope is 0.005, both of which are within the convergence interval, satisfying the preset stability condition. The fault duration is accumulated from the first time the condition is met. After 5 minutes, the accumulated duration is 300 seconds. The preset stability condition is continuously met within 5 minutes. The system sampling frequency is 10Hz, with a total of 3000 samples, and the accumulated duration is 300 seconds. The confidence level exceeds the threshold by 0.05. The 300-second duration and the 0.05 exceedance are mapped to a multi-dimensional fault level matrix, and the fault severity level is identified as a moderate fault (M1). An abnormal status code B11 is generated based on this identifier. The status code, the start timestamp 1685234567 seconds, the duration of 300 seconds, and the confidence level 0.75 are concatenated in sequence to construct a fault alarm signal frame and output it. After the device maintenance reminder is triggered, the sensor data acquisition is paused, and after calibration, the data is reacquired and the reference signal is updated.

[0061] refer to Figure 2 The second embodiment of the present invention provides a sensor fault prediction system, comprising: The signal filtering module is used to acquire sensor output values ​​and corresponding operating parameters in real time, and perform adaptive Kalman filtering on the sensor output values ​​and corresponding operating parameters to obtain a stable signal sequence. The feature extraction module is used to perform point-by-point difference calculation on the stationary signal sequence and the preset historical health benchmark signal to obtain the residual sequence, and to perform time-series feature extraction operation on the residual sequence to obtain the residual time-series dependency features. The probability calculation module is used to perform a memory unit time-series learning operation based on the residual time-series dependency features to obtain residual distribution features, and to perform a probability density calculation operation based on the residual distribution features to obtain residual probability density. The preliminary scoring module is used to compare the residual probability density with the preset fluctuation limit to obtain the potential fault labeling result, extract and save the corresponding residual vector according to the potential fault labeling result to obtain the residual vector to be analyzed, and perform feature mapping operation according to the residual vector to be analyzed to obtain the preliminary fault probability score. The confidence adjustment module is used to compare the preliminary fault probability score with a preset warning threshold to obtain a multi-source environmental weight vector, and to perform a weighted adjustment operation on the preliminary fault probability score based on the multi-source environmental weight vector to obtain an abnormal confidence value. The fault alarm module is used to output a fault alarm signal of the sensor itself if the abnormal confidence value is higher than the preset confirmation threshold and the preset stability condition is met, so as to complete the fault prediction.

[0062] It should be noted that the sensor fault prediction system provided in this embodiment of the invention is used to execute all the process steps of the sensor fault prediction method in the above embodiment. The working principle and beneficial effect of the two correspond one-to-one, so they will not be described again.

[0063] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0064] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A sensor fault prediction method, characterized in that, include: The sensor output value and corresponding operating parameters are collected in real time, and adaptive Kalman filtering is performed on the sensor output value and corresponding operating parameters to obtain a stable signal sequence. The stationary signal sequence is subjected to point-by-point difference operation with the preset historical health benchmark signal to obtain the residual sequence, and the residual time-series dependency feature is obtained by performing time-series feature extraction operation based on the residual sequence. Based on the residual temporal dependency features, a memory unit temporal learning operation is performed to obtain the residual distribution features, and a probability density calculation operation is performed based on the residual distribution features to obtain the residual probability density. The residual probability density is compared with the preset fluctuation limit to obtain the potential fault labeling result. Based on the potential fault labeling result, the corresponding residual vector is extracted and saved to obtain the residual vector to be analyzed. Based on the residual vector to be analyzed, a feature mapping operation is performed to obtain the preliminary fault probability score. The preliminary fault probability score is compared with a preset warning threshold to obtain a multi-source environmental weight vector. The preliminary fault probability score is then weighted and adjusted according to the multi-source environmental weight vector to obtain an anomaly confidence value. If the abnormal confidence level value is higher than the preset confirmation threshold and the preset stability condition is met, the sensor itself will output a fault alarm signal to complete the fault prediction.

2. The sensor fault prediction method according to claim 1, characterized in that, The real-time acquisition of sensor output values ​​and corresponding operating parameters is followed by adaptive Kalman filtering to obtain a stationary signal sequence, including: The sensor output value and the corresponding operating parameters are used to construct a multi-dimensional state input vector. The multi-dimensional state input vector is preset with a kinematic model to generate a predicted state estimate. The single-point difference between the predicted state estimate and the sensor output value is calculated. The single-point differences at each sampling time are combined and spliced ​​in time stamp order to obtain the measurement residual sequence; A preset initial noise covariance matrix is ​​retrieved, and the initial noise covariance matrix is ​​adaptively iteratively updated using the measurement residual sequence to obtain the updated initial noise covariance matrix. The Kalman gain coefficient is then calculated based on the updated initial noise covariance matrix. The predicted state estimate is corrected using the Kalman gain coefficient, and the true component in the corrected state estimate is extracted to obtain the stationary signal sequence.

3. The sensor fault prediction method according to claim 1, characterized in that, The step of performing point-by-point difference calculations between the stationary signal sequence and a preset historical health benchmark signal to obtain a residual sequence, and then performing time-series feature extraction operations based on the residual sequence to obtain residual time-series dependency features, includes: Input the stable signal sequence and the preset historical health benchmark signal, perform point-by-point subtraction operation, and output the current residual sequence; The preset historical health benchmark signal is sliced ​​according to the time span of the current residual sequence, and the autocorrelation function of the slice is calculated. The autocorrelation function is quantified to obtain the inherent fluctuation pattern. The inherent fluctuation law is set as the convolution kernel, and a convolution operation is performed on the current residual sequence to identify and output local feature components; The local feature components are subjected to dimensionality reduction mapping to extract and output the temporal dependency features of the residual sequence.

4. The sensor fault prediction method according to claim 1, characterized in that, The step of performing a memory unit temporal learning operation based on the residual temporal dependency features to obtain residual distribution features, and then performing a probability density calculation operation based on the residual distribution features to obtain residual probability density, includes: Input the residual temporal dependency features, update the cell state through the gating operation of the preset memory unit to obtain the updated cell state, filter the output based on the updated cell state, and obtain the long short-term memory hidden state vector containing effective temporal information. The long short-term memory hidden state vector is input into the terminal of a preset long short-term memory neural network for decoding to obtain Gaussian mixture modeling parameters describing the residual statistical distribution. A dynamic probability distribution model is constructed based on the Gaussian mixture modeling parameters. The measured residual values ​​are input into the dynamic probability distribution model to calculate the sum of weighted probability densities, and the residual probability density is obtained.

5. The sensor fault prediction method according to claim 1, characterized in that, The process of comparing the residual probability density with a preset fluctuation limit to obtain a potential fault labeling result, extracting and saving the corresponding residual vector based on the potential fault labeling result to obtain the residual vector to be analyzed, and performing feature mapping operation based on the residual vector to be analyzed to obtain a preliminary fault probability score includes: Input the residual probability density and a preset fluctuation limit. If the residual probability density exceeds the preset fluctuation limit after comparison, output a potential fault flag. Based on the potential fault markers, extract and store the corresponding current residual vector, and use the stored current residual vector as the residual vector to be analyzed; Input the residual vector to be analyzed and the historical health baseline residual pattern, calculate the cosine similarity value of each sampling point according to the sampling timestamp order, combine and sort the cosine similarity values ​​according to the timestamp order, and generate and output the similarity distribution sequence. A weighted matrix is ​​constructed based on the similarity distribution sequence. The weighted matrix is ​​used to perform a linear transformation on the residual vector to be analyzed to obtain the transformed residual vector, and the deviation directions of the first three principal components are extracted. The transformed residual vector is projected onto the deviation direction of the first three principal components to obtain the deviation feature vector. The deviation feature vector is then input into a preset probability mapping model for feature mapping processing, and the preliminary fault probability score is output.

6. The sensor fault prediction method according to claim 1, characterized in that, The step of comparing the preliminary fault probability score with a preset warning threshold to obtain a multi-source environmental weight vector, and then performing a weighted adjustment operation on the preliminary fault probability score based on the multi-source environmental weight vector to obtain an anomaly confidence value, includes: Input a multi-dimensional operating condition parameter sequence containing temperature, vibration, and electromagnetic data, and decouple the multi-dimensional operating condition parameter sequence to obtain three types of disturbance components; An environmental disturbance energy spectrum matrix is ​​constructed based on the three types of disturbance components. The information entropy value of the environmental disturbance energy spectrum matrix is ​​calculated to obtain the disturbance intensity coefficient. The disturbance intensity coefficients are combined to obtain a set of disturbance intensity coefficients. The set of disturbance intensity coefficients is then normalized to output a multi-source environmental weight vector. The preliminary fault probability score is weighted and attenuated using the multi-source environmental weight vector to output the environmental correction probability value. Calculate the ratio of the difference between the preliminary fault probability score and the environmental correction probability value, perform an activation transformation on the ratio of the difference, and output the environmental disturbance confidence factor. The environmental interference confidence factor is used to perform a reverse weighted correction on the environmental correction probability value, and the anomaly confidence value is output.

7. The sensor fault prediction method according to claim 1, characterized in that, If the abnormal confidence value is higher than a preset confirmation threshold and a preset stability condition is met, a sensor fault alarm signal is output to complete fault prediction, including: Input the historical sequence of the abnormal confidence scores, calculate the dispersion and trend slope of the historical sequence, and combine the dispersion and trend slope to obtain the confidence stability index. If the abnormal confidence value is higher than the preset confirmation threshold and the confidence stability index is in the convergence range, then the time of the state where the abnormal confidence value is higher than the preset confirmation threshold is accumulated, the duration of the fault is counted and output. The range of fault duration is set as the row dimension, and the amount by which the anomaly confidence value exceeds the preset confirmation threshold is set as the column dimension, thus constructing a multidimensional fault level matrix; The duration of the fault and the extent by which the abnormal confidence value exceeds the preset confirmation threshold are mapped to a multi-dimensional fault level matrix. Each intersection region in the multi-dimensional fault level matrix corresponds to a unique fault severity level identifier. The fault severity level identifier is extracted and output. An abnormal status code is generated based on the fault severity level identifier, and a fault alarm signal frame is constructed and output using the abnormal status code.

8. A sensor fault prediction system, characterized in that, include: The signal filtering module is used to acquire sensor output values ​​and corresponding operating parameters in real time, and perform adaptive Kalman filtering on the sensor output values ​​and corresponding operating parameters to obtain a stable signal sequence. The feature extraction module is used to perform point-by-point difference calculation on the stationary signal sequence and the preset historical health benchmark signal to obtain the residual sequence, and to perform time-series feature extraction operation on the residual sequence to obtain the residual time-series dependency features. The probability calculation module is used to perform a memory unit time-series learning operation based on the residual time-series dependency features to obtain residual distribution features, and to perform a probability density calculation operation based on the residual distribution features to obtain residual probability density. The preliminary scoring module is used to compare the residual probability density with the preset fluctuation limit to obtain the potential fault labeling result, extract and save the corresponding residual vector according to the potential fault labeling result to obtain the residual vector to be analyzed, and perform feature mapping operation according to the residual vector to be analyzed to obtain the preliminary fault probability score. The confidence adjustment module is used to compare the preliminary fault probability score with a preset warning threshold to obtain a multi-source environmental weight vector, and to perform a weighted adjustment operation on the preliminary fault probability score based on the multi-source environmental weight vector to obtain an abnormal confidence value. The fault alarm module is used to output a fault alarm signal of the sensor itself if the abnormal confidence value is higher than the preset confirmation threshold and the preset stability condition is met, so as to complete the fault prediction.

Images