Automobile multi-sensor fault signal recognition method based on sparse self-encoding network

Abstract

This invention discloses a method for identifying fault signals from multiple automotive sensors based on sparse autoencoder networks. To address the problem of inaccurate consistency monitoring and fault attribution under uncertainties such as missing data, noise, and hysteresis in redundant channels, this invention constructs a redundant graph structure, employs a shared dictionary and channel-specific dictionaries for sparse encoding, performs bidirectional cross-channel reconstruction in a shared subspace, calculates self-reconstruction and reconstruction residuals using time masks, and integrates evidence such as cross-correlation, hysteresis, and missing data ratios to estimate edge-level evidence weights and generate a time-varying redundancy graph. A joint objective is established for training, and consistency deviation and adaptive thresholds are calculated during the online phase. Combined with evidence gating and security degradation, this method achieves precise detection and localization of abnormal channels, robust identification under time-varying conditions and data loss, and the technical effects of uncertainty quantification and security degradation.

Description

Technical Field

[0001] This invention relates to the field of automotive sensors, and more particularly to a method for identifying fault signals from multiple automotive sensors based on sparse autoencoder networks. Background Technology

[0002] With the development of intelligent connectivity and electrification, vehicles generally adopt multi-sensor redundancy configurations for the same physical quantity (such as pressure, temperature, speed, acceleration, steering angle, etc.) to support functional safety and high availability monitoring.

[0003] Existing consistency monitoring and fault diagnosis methods mainly include: voting and median filtering based on engineering experience, and consistency verification threshold method; analytical redundancy and residual generation based on models (such as Kalman filtering, parity space, observer); and data-driven principal component analysis, sparse representation, autoencoder, deep time series model, etc. Meanwhile, some works have attempted to use the correlation between multiple sensors for cross-channel comparison or use evidence theory for uncertainty fusion.

[0004] However, existing technologies still have the following shortcomings, which limit the reliable monitoring and fault attribution of redundant channel consistency under time-varying operating conditions and data quality fluctuation scenarios:

[0005] 1. Coupling of consistency measurement and channel-specific bias: Many methods directly compare across channels on the signal or overall characterization without clearly separating the "common operating condition components" from the "channel-specific bias". This makes them susceptible to calibration bias, slow drift and noise interference, resulting in false alarms and unclear fault attribution.

[0006] 2. Redundant relationships lack time-varying modeling and evidence-based edge-level weighting: Existing methods mostly use fixed connectivity or static weights, which are difficult to adapt to the fact that cross-channel correlations change dynamically with operating conditions, lag and missing proportions. The fusion granularity of evidence such as missing samples, cross-correlation changes, and lag uncertainty is coarse, and there is a lack of edge-level gating and weighting, resulting in insufficient robustness and interpretability.

[0007] 3. Rigid judgment and threshold strategies, weak security degradation mechanisms: Common fixed or global thresholds are difficult to adapt to the historical statistical differences and time-varying distribution drift of different channels. When cross-channel evidence is insufficient or unreliable, there is a lack of clear uncertainty measurement and controllable degradation strategies, which can easily lead to misjudgment or failure to detect in online scenarios.

[0008] Therefore, there is an urgent need for a multi-sensor fault identification method that can characterize the common operating conditions of redundant channels in a shared subspace, decouple the channel-specific deviations, dynamically adjust the relationships and judgment thresholds between channels based on evidence, and perform safety degradation under high uncertainty. Summary of the Invention

[0009] One objective of this invention is to propose a method for identifying automotive multi-sensor fault signals based on sparse autoencoder networks. Addressing the problems of existing technologies such as coupling consistency measurement with channel deviation, static redundancy relationships lacking evidence weighting, rigid thresholds lacking uncertainty and security degradation, this invention proposes a method that constructs a redundant graph structure, uses a shared dictionary and a channel-specific dictionary for sparse encoding, performs bidirectional cross-channel reconstruction only in the shared subspace, and integrates self-reconstruction residuals, reconstruction residuals, cross-correlation, lag, and missing proportion estimates to form a time-varying redundancy graph. Based on evidence-weighted reconstruction constraints, graph Laplace consistency constraints, and sparse regularization, the network is trained. During the online phase, consistency deviation is calculated, an adaptive threshold is set, and security degradation is triggered when evidence is insufficient. This invention achieves accurate detection and localization of abnormal channels, robust adaptation to time-varying operating conditions and data missingness, and the technical effects of uncertainty quantification and reduced false alarms and missed alarms.

[0010] A method for identifying fault signals from multiple automotive sensors based on a sparse autoencoder network according to an embodiment of the present invention is characterized by comprising the following steps:

[0011] S1. Collect the multi-channel time series of the same physical quantity redundancy group in the vehicle system, perform time synchronization and resampling on the time series, and window it according to a fixed length to obtain windowed input data, and generate a time mask for marking missing and invalid sampling points, and output the windowed input data and time mask.

[0012] S2. Using windowed input data and time mask as input, establish an initial redundancy map of the redundancy group. Send the input into a sparse autoencoder network containing a shared dictionary and a channel private dictionary. For each channel, obtain the shared sparse code, the channel private sparse code, and the self-reconstruction signal. Calculate the self-reconstruction residual based on the time mask and output the initial redundancy map, the shared sparse code, the channel private sparse code, the self-reconstruction signal, and the self-reconstruction residual.

[0013] S3. Taking the initial redundancy map, shared sparse coding, channel-specific sparse coding, self-reconstruction signal, and self-reconstruction residual as input, perform bidirectional cross-channel mutual reconstruction on the initial redundancy map based only on the shared sparse coding and shared dictionary, and calculate the mutual reconstruction residual based on the time mask. Input the self-reconstruction residual, mutual reconstruction residual, and evidence reflecting cross-channel correlation and data integrity into the evidence estimation module to obtain the edge-level evidence weight. Use the edge-level evidence weight to assign weights to the initial redundancy map to generate a time-varying redundancy map, and output the time-varying redundancy map and the edge-level evidence weight.

[0014] S4. Using the time-varying redundancy graph and its edge-level evidence weights, shared sparse coding, channel-private sparse coding, self-reconstruction signal, self-reconstruction residual and mutual reconstruction residual as input, construct a joint objective function based on the time-varying redundancy graph and the edge-level evidence weights, and iteratively update the sparse autoencoder network, outputting the trained network parameters and the updated shared dictionary and channel-private dictionary.

[0015] S5. Using the trained network parameters, the updated shared dictionary and channel-private dictionary, and the initial redundancy graph as input, the new windowed input data and time mask are fed into the sparse autoencoder network to obtain the shared sparse code, channel-private sparse code, and self-reconstructed signal. The self-reconstruction residual is calculated based on the time mask. Cross-channel mutual reconstruction is performed on the initial redundancy graph based on the shared sparse code and the shared dictionary. The mutual reconstruction residual is calculated based on the time mask. Based on this, the online edge-level evidence weights are estimated and the initial redundancy graph is weighted to generate an online time-varying redundancy graph. The output is the online time-varying redundancy graph, online edge-level evidence weights, shared sparse code, channel-private sparse code, self-reconstruction residual, and mutual reconstruction residual.

[0016] S6. Taking the online time-varying redundancy graph, online edge-level evidence weight, shared sparse coding, channel private sparse coding, self-reconstruction residual and mutual reconstruction residual as input, the mutual reconstruction residual of each channel is weighted and summarized on the online time-varying redundancy graph according to the online edge-level evidence weight, and the consistency deviation is calculated by combining the channel private sparse coding and self-reconstruction residual statistics of each channel. The adaptive threshold of each channel is determined based on the statistics of the historical window, and the consistency deviation of each channel and the corresponding adaptive threshold are output.

[0017] S7. Using the consistency deviation of each channel and the corresponding adaptive threshold as input, compare the consistency deviation of each channel with its threshold: when any channel exceeds the threshold, output a fault alarm and give the abnormal channel identifier of that channel; when none exceed the threshold, output the redundancy group consistency judgment and use the online edge-level evidence weight as the uncertainty measure. When the online edge-level evidence weight of each edge connected to a channel and other channels is lower than the threshold, trigger the security degradation to the judgment based only on the self-reconstruction residual of that channel, and output a high uncertainty state identifier.

[0018] Optionally, step S1 specifically includes:

[0019] Collect multiple raw time-series data from redundant sensor groups targeting the same physical quantity from the vehicle system;

[0020] The multiple data streams are aligned to a common time base based on the time stamps of each channel to obtain time-synchronized data.

[0021] The synchronized channel signals are resampled on a unified sampling grid using a known resampling method to obtain a multi-channel sequence with consistent sampling times.

[0022] The resampled signals of each channel are divided into time windows according to a preset fixed window length, allowing partial overlap between adjacent windows to obtain windowed input data;

[0023] For each sampling point within each window, at least based on the quality indicators provided by the acquisition system or based on the determination of missing samples, duplicate samples, numerical out-of-bounds values, and obvious unreasonable abrupt changes, a time mask is generated that is aligned point by point with the windowed input data. The time mask is used to mark missing sampling points and invalid sampling points.

[0024] Output the windowed input data and the time mask.

[0025] Optionally, step S2 specifically includes:

[0026] An initial redundancy graph is constructed based on the redundancy relationships between channels within the redundancy group. Each channel is treated as a node in the graph, and edge connections are established between channels with redundant relationships.

[0027] The edges of the initial redundant graph can be directed or undirected. The initial weights of the edges are determined based on engineering priors or statistical similarity. When not specified in advance, they are taken as equal weights.

[0028] The windowed input data and the time mask input are combined into a sparse autoencoder network containing a shared dictionary and a channel-private dictionary. The shared dictionary is used by all channels in the redundancy group, and the channel-private dictionary corresponds to each channel.

[0029] The sparse autoencoder network is used to obtain a shared sparse code and a channel-specific sparse code for each channel in the redundancy group. The windowed input data of the channel is then reconstructed based on the shared dictionary and the channel-specific dictionary to obtain a reconstructed signal. The shared dictionary is used to characterize the common operating conditions of the redundancy group, and the channel-specific dictionary is used to characterize the unique deviation of the channel.

[0030] The difference between the self-reconstructed signal and the windowed input data of the channel is calculated at the effective sampling points based on the time mask, and then normalized according to the number of effective sampling points in the window to obtain the self-reconstructed residual.

[0031] Output the initial redundancy map, the shared sparse coding, the channel-private sparse coding, the self-reconstructed signal, and the self-reconstructed residual.

[0032] Optionally, step S3 specifically includes:

[0033] The initial redundancy map, the shared sparse coding of each channel, the channel-specific sparse coding of each channel, the self-reconstructed signal of each channel, and the self-reconstructed residual of each channel are used as inputs.

[0034] Based solely on shared sparse coding and a shared dictionary, cross-channel mutual reconstruction is performed in two directions between the two channels connected by each edge of the initial redundant graph to obtain a mutual reconstruction signal;

[0035] Based on the time mask corresponding to the reconstructed channel, the difference between the cross-reconstructed signal and the windowed input data of the channel is calculated at the sampling positions marked as valid, and the cross-reconstruction residual is obtained.

[0036] The self-reconstruction residual, mutual reconstruction residual, cross-channel cross-correlation index, cross-channel lag estimation, and missing proportion obtained from time mask statistics are input into the evidence estimation module to obtain the edge-level evidence weights of each edge in the initial redundancy graph.

[0037] The corresponding edges in the initial redundancy graph are weighted using the edge-level evidence weights to generate a time-varying redundancy graph;

[0038] Output the time-varying redundancy graph and the edge-level evidence weights.

[0039] Optionally, step S4 specifically includes:

[0040] Using time-varying redundancy graphs and their edge-level evidence weights, as well as shared sparse coding, channel-private sparse coding, self-reconstructed signals, self-reconstructed residuals and mutual reconstruction residuals as inputs, a joint objective function is constructed and the sparse autoencoder network is iteratively updated.

[0041] The joint objective function includes at least the weighted sum of the following constraints: a self-reconstruction constraint that measures the difference between the self-reconstructed signal and the windowed input data of the corresponding channel at the position marked as valid by the time mask; a mutual reconstruction constraint that uses only shared sparse coding and shared dictionary for cross-channel mutual reconstruction and calculates the mutual reconstruction residual at the valid sampling position of the reconstructed channel and weights it according to the edge-level evidence weight of the corresponding edge; a shared sparse coding consistency constraint that constructs a graph Laplacian operator based on the time-varying redundancy graph and weights it according to the edge-level evidence weight to ensure that the shared sparse coding of connected channels maintains consistency under the graph structure; and a sparse regularization constraint that imposes sparsity requirements on the shared sparse coding and the channel-specific sparse coding, so that common operating conditions are represented by shared sparse coding and channel-specific deviations are represented by channel-specific sparse coding.

[0042] By minimizing the joint objective function, the parameters of the sparse autoencoder network, as well as the shared dictionary and the channel private dictionary, are iteratively updated, and the trained network parameters and the updated shared dictionary and channel private dictionary are output.

[0043] Optionally, step S5 specifically includes:

[0044] Given the trained network parameters and the updated shared dictionary and channel private dictionary, and using the initial redundancy graph as the graph structure reference, the new windowed input data and its corresponding time mask are input into the sparse autoencoder network. For each channel in the redundancy group, the shared sparse code, the channel private sparse code, and the self-reconstruction signal are obtained, and the self-reconstruction residual is calculated at the position marked as valid by the time mask.

[0045] Based solely on shared sparse coding and a shared dictionary, cross-channel mutual reconstruction is performed in both directions of each edge of the initial redundant graph to obtain the mutual reconstruction signal, and the mutual reconstruction residual is calculated at the effective position marked by the time mask of the reconstructed channel.

[0046] The self-reconstruction residual, mutual reconstruction residual, cross-channel cross-correlation index, cross-channel lag estimation, and missing proportion obtained based on time mask statistics are input into the evidence estimation module to obtain the online edge-level evidence weights of each edge of the initial redundancy graph.

[0047] The corresponding edges of the initial redundant graph are weighted using the online edge-level evidence weights to generate an online time-varying redundant graph;

[0048] Outputs online time-varying redundancy graph, online edge-level evidence weights, shared sparse coding, channel-private sparse coding, self-reconstruction residuals, and mutual reconstruction residuals.

[0049] Optionally, step S6 specifically includes:

[0050] Using online time-varying redundancy graph, online edge-level evidence weights, shared sparse coding, channel-private sparse coding, self-reconstruction residuals, and mutual reconstruction residuals as inputs, the following processing is performed on each channel within the redundancy group: Select each edge connected to the channel in the online time-varying redundancy graph, and sum the mutual reconstruction residuals obtained with the channel as the reconstructed end according to the online edge-level evidence weights of the corresponding edges to obtain the cross-channel inconsistency measure of the channel;

[0051] The weighted summation result is superimposed with the channel-private sparse coding energy of the channel and the statistics of the channel's self-reconstruction residual within the window to obtain the consistency deviation of the channel.

[0052] The adaptive threshold is calculated based on the sliding quantile of the consistency deviation of the channel across multiple historical windows.

[0053] Output the consistency deviation of each channel and its corresponding adaptive threshold.

[0054] Optionally, step S7 specifically includes:

[0055] Using the consistency deviation of each channel and the corresponding adaptive threshold as input, the consistency deviation of each channel in the redundancy group is compared with the adaptive threshold of that channel.

[0056] When the consistency deviation of any channel exceeds its adaptive threshold, a fault alarm is output and the abnormal channel identifier of that channel is given.

[0057] When the consistency deviation of all channels in the redundancy group does not exceed their respective adaptive thresholds, the output determines that the redundancy group is in a consistent state.

[0058] Furthermore, using the online edge-level evidence weight as an uncertainty measure, for any channel, when the online edge-level evidence weights corresponding to all edges connected to that channel in the online time-varying redundancy graph are all lower than a preset threshold, a security downgrade is triggered, switching to a mode that judges based solely on the self-reconstruction residual of that channel, while simultaneously outputting a high uncertainty state flag.

[0059] The beneficial effects of this invention are:

[0060] (1) Improve the accuracy of consistency monitoring and fault attribution: By decoupling common operating conditions and channel-specific deviations through shared dictionary and channel private dictionary, cross-channel mutual reconstruction relies only on shared subspace, focusing on the "should be consistent" components, significantly suppressing calibration deviation, slow drift and noise interference, and achieving accurate positioning and low false alarms of abnormal channels;

[0061] (2) Improve the robustness and interpretability of time-varying working conditions and data uncertainty: integrate evidence such as self-reconstruction residuals, mutual reconstruction residuals, cross-correlation, lag and missing ratio to estimate the weight of edge evidence, generate online time-varying redundancy graphs and implement evidence gating and security degradation, provide uncertainty quantification, reduce false detection and false negative detection, and maintain controllable degradation when there is high missing or weak correlation.

[0062] (3) Enhance online usability and maintainability: Through joint optimization including self-reconstruction constraints, evidence-weighted mutual reconstruction constraints, graph Laplace consistency constraints and sparse regularization, combined with sliding quantile adaptive threshold, a stable consistency deviation index is formed, reducing the cost of manual threshold tuning and scenario migration, and improving the reliability of online deployment. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a method for identifying fault signals from multiple automotive sensors based on sparse autoencoder networks, as proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 A method for identifying fault signals from multiple automotive sensors based on sparse autoencoder networks, characterized by the following steps:

[0067] S1. Collect the multi-channel time series of the same physical quantity redundancy group in the vehicle system, perform time synchronization and resampling on the time series, and window it according to a fixed length to obtain windowed input data, and generate a time mask for marking missing and invalid sampling points, and output the windowed input data and time mask.

[0068] S2. Using windowed input data and time mask as input, establish an initial redundancy map of the redundancy group. Send the input into a sparse autoencoder network containing a shared dictionary and a channel private dictionary. For each channel, obtain the shared sparse code, the channel private sparse code, and the self-reconstruction signal. Calculate the self-reconstruction residual based on the time mask and output the initial redundancy map, the shared sparse code, the channel private sparse code, the self-reconstruction signal, and the self-reconstruction residual.

[0069] S3. Taking the initial redundancy map, shared sparse coding, channel-specific sparse coding, self-reconstruction signal, and self-reconstruction residual as input, perform bidirectional cross-channel mutual reconstruction on the initial redundancy map based only on the shared sparse coding and shared dictionary, and calculate the mutual reconstruction residual based on the time mask. Input the self-reconstruction residual, mutual reconstruction residual, and evidence reflecting cross-channel correlation and data integrity into the evidence estimation module to obtain the edge-level evidence weight. Use the edge-level evidence weight to assign weights to the initial redundancy map to generate a time-varying redundancy map, and output the time-varying redundancy map and the edge-level evidence weight.

[0070] S4. Using the time-varying redundancy graph and its edge-level evidence weights, shared sparse coding, channel-private sparse coding, self-reconstruction signal, self-reconstruction residual and mutual reconstruction residual as input, construct a joint objective function based on the time-varying redundancy graph and the edge-level evidence weights, and iteratively update the sparse autoencoder network, outputting the trained network parameters and the updated shared dictionary and channel-private dictionary.

[0071] S5. Using the trained network parameters, the updated shared dictionary and channel-private dictionary, and the initial redundancy graph as input, the new windowed input data and time mask are fed into the sparse autoencoder network to obtain the shared sparse code, channel-private sparse code, and self-reconstructed signal. The self-reconstruction residual is calculated based on the time mask. Cross-channel mutual reconstruction is performed on the initial redundancy graph based on the shared sparse code and the shared dictionary. The mutual reconstruction residual is calculated based on the time mask. Based on this, the online edge-level evidence weights are estimated and the initial redundancy graph is weighted to generate an online time-varying redundancy graph. The output is the online time-varying redundancy graph, online edge-level evidence weights, shared sparse code, channel-private sparse code, self-reconstruction residual, and mutual reconstruction residual.

[0072] S6. Taking the online time-varying redundancy graph, online edge-level evidence weight, shared sparse coding, channel private sparse coding, self-reconstruction residual and mutual reconstruction residual as input, the mutual reconstruction residual of each channel is weighted and summarized on the online time-varying redundancy graph according to the online edge-level evidence weight, and the consistency deviation is calculated by combining the channel private sparse coding and self-reconstruction residual statistics of each channel. The adaptive threshold of each channel is determined based on the statistics of the historical window, and the consistency deviation of each channel and the corresponding adaptive threshold are output.

[0073] S7. Using the consistency deviation of each channel and the corresponding adaptive threshold as input, compare the consistency deviation of each channel with its threshold: when any channel exceeds the threshold, output a fault alarm and give the abnormal channel identifier of that channel; when none exceed the threshold, output the redundancy group consistency judgment and use the online edge-level evidence weight as the uncertainty measure. When the online edge-level evidence weight of each edge connected to a channel and other channels is lower than the threshold, trigger the security degradation to the judgment based only on the self-reconstruction residual of that channel, and output a high uncertainty state identifier.

[0074] In this specific embodiment, S1 specifically refers to:

[0075] For redundant sensor arrays of the same physical quantity, firstly, multi-channel raw sequences containing time stamps and observations are acquired and aligned to a common time reference based on the time stamps to establish a unified, equally spaced sampling grid.

[0076] ;

[0077] in Represents a common time set, Indicates the first Alignment time, This indicates the total number of samples after alignment and the sampling step size. ;

[0078] Then, resampling is performed on each channel within that time grid, denoted as the resampling operator. This yields an isochronous sequence:

[0079] ;

[0080] in Indicates channel index, Indicates the first Channel at time Alignment value, Indicates step size For the channel At any moment The resampling operator for the output value preferably performs time sorting and merges repeated samples before resampling to eliminate artifacts, and can use piecewise linear interpolation or zero-order preservation to complete the resampling.

[0081] The aligned channel sequences are then divided into overlapping time windows with a fixed window length and a sliding step size, with the window length set to 1. The sliding step size is Window index is And the total number of windows is Within the same window, the input data matrix is ​​formed by stacking the data row by row according to the channel. To mark missing and invalid sampling points, a time mask is generated by combining the acquisition system's quality flag, missing and duplicate sampling detection, and judgment of numerical out-of-bounds and unreasonable abrupt changes. The quality flag is defined as follows: (A value of 1 indicates valid quality, and a value of 0 indicates invalid quality). The allowable amplitude range is... (respectively the first) The lower and upper limits of the channel, and the mutation detection threshold are... and define adjacent increments (when season ), and accordingly in the The structure within a window and Point-aligned mask elements ;

[0082] in Indicates the first The first window Channel 1 The mask value of each sampling point, For indicator functions, Indexing sampling points within the window To be compatible with the window number The global time index corresponding to each sampling point, and then derived from all Composition of time mask matrix Final output and .

[0083] In this specific embodiment, S2 specifically refers to:

[0084] Based on windowed input and time masking, an initial redundancy graph is first constructed according to redundancy relationships. Then, each channel signal is fed into a sparse autoencoder network containing a shared dictionary and a channel-specific dictionary to achieve decoupled representation. Specifically, the signal vector of the c-th channel in the w-th time window is denoted as... And the time mask that is aligned point by point with it is ;

[0085] in Indicates window index, Indicates channel index, Represents a column vector with a length equal to the number of sampling points in the window. They are binary column vectors of the same length;

[0086] Constructing an initial redundancy graph serves as a priori for the redundancy relationships between channels:

[0087] ;

[0088] in A set of nodes that correspond one-to-one with each channel. For a set of edges with redundant associations, It is an edge weight matrix, and in the absence of prior knowledge, it can be taken as equal weight and set as symmetric as needed;

[0089] Let the shared dictionary be To characterize the common operating conditions of the redundant group, let the private dictionary of the c-th channel be... To characterize the channel-specific bias, the shared sparse code and the private sparse code are obtained by solving the mask-weighted sparse reconstruction criterion for the sparse code of the c-th channel in the w-th window, written as:

[0090] ;

[0091] in and These represent the shared and private sparse coding of the c-th channel in the w-th window, respectively. and To optimize the use of variables to carry the shared and private encoding to be determined, element-wise multiplication Represents the Euclidean norm, express norm, and The sparse regularization coefficients of shared and private encodings are respectively. This represents the parameter pair that minimizes the objective function;

[0092] The self-reconstructed signal of this channel, reconstructed from the obtained sparse coding, is as follows:

[0093] ;

[0094] in This represents the reconstruction result of the c-th channel in the w-th window;

[0095] Further, the self-reconstruction residuals are calculated based on the mask at the valid locations and normalized according to the number of valid points:

[0096] ;

[0097] in For the self-reconstruction residual of channel c in window w, for The number of elements with a value of 1 represents the number of valid sampling points. This means that when the number of valid points is 0, 1 is substituted to avoid division by zero;

[0098] This allows us to output the initial redundancy map. and its border rights Shared sparse coding Channel private sparse coding Self-reconfigured signals With self-reconstruction residual Proceed to the next steps.

[0099] In this specific embodiment, S3 specifically refers to:

[0100] Using the initial redundancy map and shared sparse coding for each channel as input, bidirectional cross-channel mutual reconstruction is performed only on connected channels in the shared subspace. The mutual reconstruction residual is calculated at the effective sampling position of the reconstructed end. Simultaneously, the weights of edge-level evidence are estimated by combining evidence such as cross-channel correlation, hysteresis, and missing value ratios, and a time-varying redundancy map is generated. Specifically: for the first... A time window, using only the shared dictionary. Shared sparse coding with source channel Reconstruct the target channel The shared components yield the reconstructed signal:

[0101] ;

[0102] in The shared dictionary representing redundant groups, whose dimensions are determined by the window length. With the number of shared atoms Determined and used to characterize common operating conditions. Indicates the first One channel in the window Shared sparse coding, dimension is Indicates direction The reconstructed signal with a length of and This is an index for the source and target channels;

[0103] The cross-reconstruction residuals are calculated and normalized at the effective sampling locations of the reconstructed channels:

[0104] ;

[0105] in Indicates the first The passage is in the window The observation vector, with a length of To and A point-by-point aligned binary time mask, with a value of 1 indicating that the sampling point is valid. This indicates that Hadamard multiplies element by element. Represents the L2 norm, For mask The number of elements with a value of 1, used to count valid sampling points. Used to avoid division by zero when there are no valid points;

[0106] To obtain a symmetric side-level metric, a two-way average residual is defined:

[0107] ;

[0108] in For direction mutual reconstruction residuals, For the edge The two-way average cross-reconstruction residual;

[0109] Using the above residuals together with cross-channel statistical correlation, lag and missing proportions as evidence, we obtain the edge-level evidence weights:

[0110] ;

[0111] in and They represent the first With the The passage is in the window The self-reconstruction residual (from the calculation result of step S2, used to reflect single-channel self-consistency). This represents the normalized cross-correlation peak value obtained after mask alignment and within the allowable hysteresis range, used to measure cross-channel linear correlation, and its value range is... This represents the discrete hysteresis corresponding to the relevant peak value, expressed in integer terms based on the number of sampling points. This represents the edge-level missing percentage obtained from the statistics of the two-channel mask, used to quantify data integrity. To map multi-source evidence to monotonic fusion function Its parameter vector and its dimensions are set by the specific implementation. Indicates in window Top opposite edge The credibility of the evidence;

[0112] Finally, the initial edge weights are weighted using evidence weights to obtain time-varying edge weights:

[0113] ;

[0114] in Representing an edge The initial edge weights, derived from the prior settings of the initial redundant graph, Display window Time-varying edge weights;

[0115] Based on this convergence, a time-varying redundancy map is obtained. (Its node set and edge set are inherited from the initial redundant graph, and the edge weights are determined by...) (Composition) and output and Used for subsequent joint training and online evaluation.

[0116] In this specific embodiment, S4 specifically refers to:

[0117] Using the time-varying redundancy graph and its edge-level evidence weights as constrained extrinsic parameters, and co-constructing a joint objective function with each channel's shared sparse coding and private sparse coding, self-reconstructed signal, self-reconstructed residual, and mutual reconstruction residual, the sparse autoencoder network and dictionary parameters are jointly trained by alternating minimization and gradient update. Specifically: for the ... Let there be a time window, and denote the set of channel nodes as... And the edge set is Time-varying edge weights Representing an edge In the window The weighted strength of evidence, Indicates the first Window signals for each channel, This represents its point-by-point time mask. This indicates the number of valid mask points within the window. For shared dictionaries, For the first A private dictionary for each channel. and The first One channel in the window To highlight "self-reconstruction consistency," "cross-channel sharing consistency," and "graph structure consistency" while suppressing redundant representations, a joint objective function is constructed using shared and private sparse coding. as follows:

[0118] in These are the weights for the self-reconstruction term, used to measure the self-reconstruction error at valid mask locations. For evidence-weighted cross-channel interoperability reconstruction item weights, only shared subspaces are used to highlight common operating conditions. For shared encoding Figure 1 Consistency term weights are used to ensure that adjacent channels maintain consistent coding under time-varying graphs. and These are the sparse regularization coefficients for shared and private encodings, respectively. This indicates that Hadamard multiplies element by element. Denotes the Euclidean norm. express Norm, Used to avoid division by zero when there are no valid points;

[0119] During training, the sum of objectives across multiple windows is used as the overall optimization target, employing an alternating minimization strategy. This begins with a fixed dictionary. and Under the condition of updating, the solution is obtained by sparse coding with mask and evidence weight. Then, under fixed encoding conditions, update using mask-weighted least squares combined with atomic column normalization. and Simultaneously, backpropagation updates the encoded network parameters to reduce the joint objective until the early stopping criterion on the validation set is satisfied. Finally, the trained network parameters and the updated shared dictionary are output. and channel private dictionary .

[0120] In this specific embodiment, S5 specifically includes:

[0121] Given the trained network parameters and updated dictionary, and using the initial redundant graph as the graph structure baseline, online inference and evidence-weighted updates are performed on new windowed inputs and time masks. Specifically, for the first online input... The window records the first... The observation vectors and masks for each channel are respectively and ;

[0122] in Indicates window index, Indicates channel index, A column vector with a length equal to the window length. Each is a binary column vector of the same length, and each valid sampling point is marked with 1.

[0123] Using the shared dictionary obtained during training With channel private dictionary The forward reasoning module obtains the first... The passage is in the window Shared sparse coding With private sparse coding Based on this, a self-reconfiguration signal is formed:

[0124] ;

[0125] in Indicates the first The passage is in the window The self-reconstruction results This indicates that the shared dictionary is used to represent common operating conditions. Indicates the first The channel's private dictionary is used to characterize channel-specific biases. and These are shared and private sparse coding, respectively.

[0126] Calculate and normalize the self-reconstruction residuals at the valid locations of the mask markers:

[0127] ;

[0128] in Indicates the first The passage is in the window Self-reconstruction residuals This indicates that Hadamard multiplies element by element. Represents the Euclidean norm, express The number of elements with a value of 1 is used to count the valid sampling points. This indicates that the larger of the two values ​​should be taken to avoid division by zero;

[0129] Bidirectional cross-channel reconstruction is performed on the edge set of the initial redundant graph using only the shared subspace, for any channel pair connected by an edge. In direction Above calculation:

[0130] ;

[0131] in Indicates source channel Shared components for target channels The results of mutual reconstruction Represents the mutual reconstruction residual in this direction, Indicates the first The passage is in the window Shared sparse coding, and The first The channel mask and its effective number of sampling points;

[0132] The mutual reconstruction residuals, self-reconstruction residuals, and cross-channel statistical evidence are input into the evidence estimation module to obtain online edge-level evidence weights:

[0133] ;

[0134] in Display window Above Online evidence weighting This indicates mapping multi-source evidence to monotonic fusion function For its parameter vector, The peak value of the cross-channel normalized cross-correlation calculated after mask alignment and within the allowable hysteresis range. The discrete hysteresis corresponding to this peak value, expressed in units of sampling points. The edge-level missing ratio obtained from the statistics of two-channel masks, and This refers to the self-reconstruction residual of the corresponding channel;

[0135] Online time-varying edge weights are obtained by assigning weights to the initial edge weights using online evidence weights:

[0136] ;

[0137] in Represents the edges in the initial redundant graph Prior weights, Display window Online time-varying edge weights are used to generate an online time-varying redundancy graph and output online edge-level evidence weights. Online time-varying edge weights Shared and private sparse coding , self-reconstruction residual With mutual reconstruction residual This is used for the next step of consistency determination and threshold adaptation.

[0138] In this specific embodiment, S6 specifically refers to:

[0139] Using online time-varying redundancy graphs, online edge-level evidence weights, shared and private sparse coding, self-reconstruction residuals, and mutual reconstruction residuals as inputs, the consistency deviation is calculated for each channel within the redundancy group, and an adaptive threshold is given. Specifically: Let the i-th channel be... The online time-varying redundancy diagram for each time window is as follows:

[0140] ;

[0141] in For the set of channel nodes, For the set of connected edges, It is an edge weight matrix and its elements Indicates in window Above Online time-varying weights;

[0142] For any channel Define its adjacency set Indicates and channel In the window The set of connected source channels is used, and the edge weights are normalized to obtain directional weights for weighted aggregation:

[0143] ;

[0144] in Indicates direction Normalized weights To avoid small constants with a denominator of zero, For channel indexing;

[0145] Only for channels in shared subspace The cross-channel inconsistency measure is obtained by weighting and summing the evidence-based residuals of the reconstructed ends:

[0146] ;

[0147] in Indicates channel In the window Cross-channel inconsistency Indicates direction The mutual reconstruction residuals are calculated from the effective sampling positions at the reconstructed end in the preceding steps;

[0148] To introduce auxiliary evidence of channel-specific bias and self-consistency, the first... Channel-private sparse coding energy:

[0149] ;

[0150] in Display window Private energy, Indicates the first The passage is in the window Private sparse coding, Indicates the number of private atoms, Represents the Euclidean norm;

[0151] Simultaneously adopting the first Channel self-reconstruction residual As a single-channel self-consistency statistic, it originates from the self-reconstruction error after normalization at the effective sampling positions in the preceding steps; based on this, the consistency deviation is constructed:

[0152] ;

[0153] in Indicates the first The passage is in the window Consistency deviation The weighting coefficients for the three pieces of evidence can be optimized according to the calibration or validation set;

[0154] To achieve online adaptive thresholding, the first threshold is defined. Historical index set of channels:

[0155] ;

[0156] in Represents a set of integers Let the length of the sliding history be a positive integer, and let the empirical quantile function be denoted as . ,in For quantile levels, for example, take To control the false alarm rate, then the first The passage is in the window The adaptive threshold is:

[0157] ;

[0158] in The threshold is determined by experience from the historical deviation sequence. Quantiles are given when historical data is insufficient. When calculating within a window, the consistency deviation of each channel can be calculated based on existing history, ultimately outputting the consistency deviation of each channel. and corresponding adaptive threshold For use in the next step of the determination.

[0159] In this specific embodiment, S7 specifically refers to:

[0160] Using the consistency deviation of each channel and the adaptive threshold for each time window as input, and combining online edge-level evidence weights and online time-varying redundancy graphs, alarms, abnormal channel location, redundancy group consistency determination, and security degradation are completed. Specifically, for the first... The time window order The consistency deviation of each channel is And its adaptive threshold is ,in Indicates window index, Indicates channel index, The deviation calculated in the preceding step S6, For the corresponding adaptive inter-value, first perform channel-by-channel determination and generate a binary result:

[0161] ;

[0162] in Indicates the first The passage is in the window The criteria for determining whether an alarm has been triggered. This indicates that the function takes the value 1 if the condition inside the parentheses is true, and 0 otherwise.

[0163] Based on this, a set of window-level exception channels is defined:

[0164] ;

[0165] in Display window abnormal channel set, For the set of channel nodes in the redundant group, if If not empty, output a fault alarm and provide the following information: The abnormal indicator of the middle channel, if If empty, output the consistency determination of the redundant group;

[0166] To provide uncertainty measurement and degradation triggering, the cross-channel evidence strength of each channel is measured based on online edge-level evidence weights, allowing the channel... Connected and with The set of incoming neighbors for the reconstructed end is:

[0167] ;

[0168] in For online time-varying redundancy graphs in the window Let the set of connected edges be Representing an edge The online evidence weights are then defined as follows: Channel-level uncertainty and degradation gating are defined as follows:

[0169] ;

[0170] in Indicates the first The passage is in the window Uncertainty measurement Indicates taking the maximum value. A gating variable indicating whether security degradation is triggered. The preset evidence threshold is defined if and only if it is related to the channel. All connected incoming edge evidence weights are lower than Downgrade triggered at time ;

[0171] In the security degradation mode, the judgment is made solely based on the self-reconstruction residual of the channel, Indicates the first The passage is in the window Self-reconstruction residuals, making This represents the downgraded inter-level value (e.g., selected quantile level) obtained based on the moving quantile of the channel's historical self-reconstructed residuals. And in the near (calculated from a historical window), then the downgrade decision is written as:

[0172] ;

[0173] in The value is based on historical statistics to ensure robustness when cross-channel evidence is insufficient, while when Output a high uncertainty state flag and As an uncertain reporting quantity that updates with the window, this step ultimately outputs the channel-level alarm results. Abnormal channel set Uncertainty measurement and high uncertainty downgrade indicator .

[0174] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying fault signals of multiple sensors of a vehicle based on a sparse auto-encoding network, characterized in that, Includes the following steps: S1. Collect the multi-channel time series of the same physical quantity redundancy group in the vehicle system, perform time synchronization and resampling on the time series, and window it according to a fixed length to obtain windowed input data, and generate a time mask for marking missing and invalid sampling points, and output the windowed input data and time mask. S2. Using windowed input data and time mask as input, establish an initial redundancy map of the redundancy group. Send the input into a sparse autoencoder network containing a shared dictionary and a channel private dictionary. For each channel, obtain the shared sparse code, the channel private sparse code, and the self-reconstruction signal. Calculate the self-reconstruction residual based on the time mask and output the initial redundancy map, the shared sparse code, the channel private sparse code, the self-reconstruction signal, and the self-reconstruction residual. S3. Taking the initial redundancy graph, shared sparse coding, channel-specific sparse coding, self-reconstruction signal, and self-reconstruction residual as input, for any edge connecting two channels in the initial redundancy graph, generate the mutual reconstruction signal of the target channel in both directions based only on the shared sparse coding and shared dictionary of the source channel, and calculate the mutual reconstruction residual at the effective sampling points based on the time mask of the target channel; obtain the evidence quantity for the edge, which consists of the self-reconstruction residual at both ends, the bidirectional mutual reconstruction residual, the cross-channel normalized cross-correlation peak value, the discrete hysteresis corresponding to the cross-correlation peak value, and the missing proportion obtained by statistical analysis of the time masks at both ends; use the evidence quantity together as evidence, and obtain the edge-level evidence weight by mapping to a monotonic fusion function of [0,1]; The initial edge weights of the initial redundant graph are assigned the edge-level evidence weights to generate a time-varying redundant graph, and the time-varying redundant graph and edge-level evidence weights are output. S4. Using the time-varying redundancy graph and its edge-level evidence weights, shared sparse coding, channel-private sparse coding, self-reconstruction signal, self-reconstruction residual and mutual reconstruction residual as input, construct a joint objective function based on the time-varying redundancy graph and the edge-level evidence weights, and iteratively update the sparse autoencoder network, outputting the trained network parameters and the updated shared dictionary and channel-private dictionary. S5. Using the trained network parameters, the updated shared dictionary and channel private dictionary, and the initial redundancy graph as input, the new windowed input data and time mask are fed into the sparse autoencoder network to obtain the shared sparse code, channel private sparse code, and self-reconstruction signal. The self-reconstruction residual is calculated based on the time mask. Cross-channel mutual reconstruction is performed on the initial redundancy graph based on the shared sparse code and the shared dictionary. The mutual reconstruction residual is calculated based on the time mask. The online edge-level evidence weights are estimated according to the same evidence normalization and fusion rules as in step S3, and the initial redundancy graph is weighted to generate an online time-varying redundancy graph. The output is the online time-varying redundancy graph, online edge-level evidence weights, shared sparse code, channel private sparse code, self-reconstruction residual, and mutual reconstruction residual. S6. Taking the online time-varying redundancy graph, online edge-level evidence weight, shared sparse coding, channel private sparse coding, self-reconstruction residual and mutual reconstruction residual as input, the mutual reconstruction residual of each channel is weighted and summarized on the online time-varying redundancy graph according to the online edge-level evidence weight, and the consistency deviation is calculated by combining the channel private sparse coding and self-reconstruction residual statistics of each channel. The adaptive threshold of each channel is determined based on the statistics of the historical window, and the consistency deviation of each channel and the corresponding adaptive threshold are output. S7. Using the consistency deviation of each channel and the corresponding adaptive threshold as input, compare the consistency deviation of each channel with its threshold: when any channel exceeds the threshold, output a fault alarm and give the abnormal channel identifier of that channel; when none exceed the threshold, output the redundancy group consistency judgment and use the online edge-level evidence weight as the uncertainty measure. When the online edge-level evidence weight of each edge connected to a channel and other channels is lower than the threshold, trigger the security degradation to the judgment based only on the self-reconstruction residual of that channel, and output a high uncertainty state identifier.

2. The method for identifying automotive multi-sensor fault signals based on sparse autoencoder networks according to claim 1, characterized in that, Step S1 is as follows: Collect multiple raw time-series data from redundant sensor groups targeting the same physical quantity from the vehicle system; The multiple data streams are aligned to a common time base based on the time stamps of each channel to obtain time-synchronized data. The synchronized channel signals are resampled on a unified sampling grid to obtain a multi-channel sequence with consistent sampling times. The resampled signals of each channel are divided into time windows according to a preset fixed window length, allowing partial overlap between adjacent windows to obtain windowed input data; For each sampling point within each window, at least based on the quality indicators provided by the acquisition system or based on the determination of missing samples, duplicate samples, numerical out-of-bounds values, and obvious unreasonable abrupt changes, a time mask is generated that is aligned point by point with the windowed input data. The time mask is used to mark missing sampling points and invalid sampling points. Output the windowed input data and the time mask.

3. The method for identifying fault signals of automobiles based on sparse autoencoder networks according to claim 1, characterized in that, Step S2 is as follows: An initial redundancy graph is constructed based on the redundancy relationships between channels within the redundancy group. Each channel is treated as a node in the graph, and edge connections are established between channels with redundant relationships. The edges of the initial redundant graph can be directed or undirected. The initial weights of the edges are determined based on engineering priors or statistical similarity. When not specified in advance, they are taken as equal weights. The windowed input data and the time mask input are combined into a sparse autoencoder network containing a shared dictionary and a channel-private dictionary. The shared dictionary is used by all channels in the redundancy group, and the channel-private dictionary corresponds to each channel. The sparse autoencoder network is used to obtain a shared sparse code and a channel-specific sparse code for each channel in the redundancy group. The windowed input data of the channel is then reconstructed based on the shared dictionary and the channel-specific dictionary to obtain a reconstructed signal. The shared dictionary is used to characterize the common operating conditions of the redundancy group, and the channel-specific dictionary is used to characterize the unique deviation of the channel. The difference between the self-reconstructed signal and the windowed input data of the channel is calculated at the effective sampling points based on the time mask, and then normalized according to the number of effective sampling points in the window to obtain the self-reconstructed residual. Output the initial redundancy map, the shared sparse coding, the channel-private sparse coding, the self-reconstructed signal, and the self-reconstructed residual.

4. The method for identifying fault signals of automobiles based on sparse autoencoder networks according to claim 1, characterized in that, Step S3 is as follows: The initial redundancy map, the shared sparse coding of each channel, the channel-specific sparse coding of each channel, the self-reconstructed signal of each channel, and the self-reconstructed residual of each channel are used as inputs. For the edge connecting the i-th channel and the j-th channel in the initial redundant graph, bidirectional cross-channel mutual reconstruction is performed with the i-th channel as the source channel and the j-th channel as the target channel, and with the j-th channel as the source channel and the i-th channel as the target channel, respectively. The mutual reconstruction signal is generated only by the shared sparse coding and shared dictionary of the source channel. Based on the time mask corresponding to the target channel, the difference between the mutual reconstruction signal and the windowed input data of the target channel is calculated at the sampling positions marked as valid, and the directional mutual reconstruction residual is obtained. The bidirectional mutual reconstruction residual of the edge is obtained from the directional mutual reconstruction residuals of the two directions. Calculate the cross-channel normalized cross-correlation peak value of the edge, the discrete hysteresis corresponding to the cross-correlation peak value, and the edge-level missing ratio obtained by the time mask statistics of the i-th channel and the j-th channel; The bidirectional cross-reconstruction residual, the i-th channel self-reconstruction residual, the j-th channel self-reconstruction residual, the cross-channel normalized cross-correlation peak value, the discrete hysteresis corresponding to the cross-correlation peak value, and the edge-level missing ratio are used together as evidence. The edge-level evidence weights for the edges within the w-th time window are calculated using the following formula: ； in and They represent the first With the The passage is in the window The self-reconstruction residual, This represents the normalized cross-correlation peak value obtained after mask alignment and within the allowable hysteresis range, used to measure cross-channel linear correlation, and its value range is... This represents the discrete hysteresis corresponding to the relevant peak value, expressed in integer terms based on the number of sampling points. This represents the edge-level missing percentage obtained from the statistics of the two-channel mask, used to quantify data integrity. To map multi-source evidence to Monotonic fusion function Its parameter vector and its dimensions are set by the specific implementation. Indicates in window Top opposite edge The credibility of the evidence; Finally, the initial edge weights are weighted using evidence weights to obtain time-varying edge weights: ； in Representing an edge The initial edge weights, derived from the prior settings of the initial redundant graph, Display window Time-varying edge weights; A time-varying redundancy graph is generated using the time-varying edge weights of each edge; Output the time-varying redundancy graph and the edge-level evidence weights.

5. The method for identifying fault signals of automobiles based on sparse autoencoder networks according to claim 1, characterized in that, Step S4 is as follows: Using time-varying redundancy graphs and their edge-level evidence weights, as well as shared sparse coding, channel-private sparse coding, self-reconstructed signals, self-reconstructed residuals and mutual reconstruction residuals as inputs, a joint objective function is constructed and the sparse autoencoder network is iteratively updated. The joint objective function includes at least the weighted sum of the following constraints: a self-reconstruction constraint that measures the difference between the self-reconstructed signal and the windowed input data of the corresponding channel at the position marked as valid by the time mask; a mutual reconstruction constraint that uses only shared sparse coding and shared dictionary for cross-channel mutual reconstruction and calculates the mutual reconstruction residual at the valid sampling position of the reconstructed channel and weights it according to the edge-level evidence weight of the corresponding edge; a shared sparse coding consistency constraint that constructs a graph Laplacian operator based on the time-varying redundancy graph and weights it according to the edge-level evidence weight to ensure that the shared sparse coding of connected channels maintains consistency under the graph structure; and a sparse regularization constraint that imposes sparsity requirements on the shared sparse coding and the channel-specific sparse coding, so that common operating conditions are represented by shared sparse coding and channel-specific deviations are represented by channel-specific sparse coding. By minimizing the joint objective function, the parameters of the sparse autoencoder network, as well as the shared dictionary and the channel private dictionary, are iteratively updated, and the trained network parameters and the updated shared dictionary and channel private dictionary are output.

6. The method for identifying fault signals of automobiles based on sparse autoencoder networks according to claim 1, characterized in that, Step S5 is as follows: Given the trained network parameters and the updated shared dictionary and channel private dictionary, and using the initial redundancy graph as the graph structure reference, the new windowed input data and its corresponding time mask are input into the sparse autoencoder network. For each channel in the redundancy group, the shared sparse code, the channel private sparse code, and the self-reconstruction signal are obtained, and the self-reconstruction residual is calculated at the position marked as valid by the time mask. Based solely on shared sparse coding and a shared dictionary, cross-channel mutual reconstruction is performed in both directions of each edge of the initial redundant graph to obtain the mutual reconstruction signal, and the mutual reconstruction residual is calculated at the effective position marked by the time mask of the target channel. According to the edge-level evidence weight calculation method described above, the online edge-level evidence weights of each edge of the initial redundant graph are obtained based on the self-reconstruction residual, mutual reconstruction residual, cross-channel normalized cross-correlation peak value, discrete lag, and the missing ratio obtained based on time mask statistics. The corresponding edges of the initial redundant graph are weighted using the online edge-level evidence weights to generate an online time-varying redundant graph; Outputs online time-varying redundancy graph, online edge-level evidence weights, shared sparse coding, channel-private sparse coding, self-reconstruction residuals, and mutual reconstruction residuals.

7. The method for identifying automotive multi-sensor fault signals based on sparse autoencoder networks according to claim 1, characterized in that, Step S6 is as follows: Using online time-varying redundancy graph, online edge-level evidence weights, shared sparse coding, channel-private sparse coding, self-reconstruction residuals, and mutual reconstruction residuals as inputs, the following processing is performed on each channel within the redundancy group: Select each edge connected to the channel in the online time-varying redundancy graph, and sum the mutual reconstruction residuals obtained with the channel as the reconstructed end according to the online edge-level evidence weights of the corresponding edges to obtain the cross-channel inconsistency measure of the channel; The weighted summation result is superimposed with the channel-private sparse coding energy of the channel and the statistics of the channel's self-reconstruction residual within the window to obtain the consistency deviation of the channel. The adaptive threshold is calculated based on the sliding quantile of the consistency deviation of the channel across multiple historical windows. Output the consistency deviation of each channel and its corresponding adaptive threshold.

8. The method for identifying fault signals of automobiles based on sparse autoencoder networks according to claim 1, characterized in that, Step S7 is as follows: Using the consistency deviation of each channel and the corresponding adaptive threshold as input, the consistency deviation of each channel in the redundancy group is compared with the adaptive threshold of that channel. When the consistency deviation of any channel exceeds its adaptive threshold, a fault alarm is output and the abnormal channel identifier of that channel is given. When the consistency deviation of all channels in the redundancy group does not exceed their respective adaptive thresholds, the output determines that the redundancy group is in a consistent state. Furthermore, using the online edge-level evidence weight as an uncertainty measure, for any channel, when the online edge-level evidence weights corresponding to all edges connected to that channel in the online time-varying redundancy graph are all lower than a preset threshold, a security downgrade is triggered, switching to a mode that judges based solely on the self-reconstruction residual of that channel, while simultaneously outputting a high uncertainty state flag.

Images