Signal processing method, fault detection method, device, equipment and storage medium

Abstract

The present disclosure relates to a signal processing method, a fault detection method, an apparatus, a device and a storage medium, and relates to the technical field of signal processing. The method deeply fuses the low-rank-sparse double matrix decomposition mechanism of the traditional RPCA scheme with the time sequence context modeling and the local space feature extraction capability, which effectively makes up for the deficiency that the traditional RPCA scheme cannot depict the time sequence evolution law of the signal, and through the step-by-step feature extraction and reconstruction of the signal processing model, the low-rank matrix accurately fits the time sequence change characteristics of the sensor signal under the normal operation of the equipment, and the sparse matrix effectively represents the abnormal situation that the sensor signal deviates from the normal evolution mode, so that the effectiveness and convergence stability of the model training process are guaranteed. After the training is completed, the sparse matrix output by the signal processing model is used for equipment fault detection, which can significantly improve the accuracy of equipment fault detection.

Description

Technical Field

[0001] This disclosure relates to the field of signal processing technology, and in particular to a signal processing method, a fault detection method, an apparatus, a device, and a storage medium. Background Technology

[0002] In industrial equipment operation and maintenance scenarios, sensor signals with multiple characteristics, such as sound signals, vibration signals, and temperature signals, are collected to detect whether the equipment has malfunctioned.

[0003] In related technologies, Robust Principal Component Analysis (RPCA) is generally used to process these sensor signals for fault detection. This algorithm decomposes the original data matrix constructed from multiple sensor signals into low-rank and sparse matrices to represent the normal evolution and abnormal conditions of the sensor signals during equipment operation, respectively. However, it is essentially a matrix-level static decomposition method that does not fully utilize the temporal characteristics of sensor signals. That is, it does not consider the contextual relationships, periodic trends, and evolutionary patterns of sensor signals in the temporal dimension. This leads to insufficient accuracy in capturing fault features in the temporal dimension, resulting in inaccurate fault detection results. Summary of the Invention

[0004] This disclosure provides a signal processing method, a fault detection method, an apparatus, a device, and a storage medium. The technical solution of this disclosure is as follows.

[0005] According to one aspect of the embodiments of this disclosure, a method for training a signal processing model is provided, the method comprising: Multiple batches of training samples are acquired, each batch of training samples includes multiple training samples, and each training sample includes multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected from the same device. Based on the training samples of each batch, construct the original data matrix of the batch; For each batch of original data matrix, the original data matrix is ​​encoded by the convolutional encoder of the signal processing model to obtain encoded features. The encoded features are then extracted by the gated encoder of the signal processing model to obtain a first latent feature. The first latent feature is then restored by the gated decoder of the signal processing model to obtain a second latent feature. The second latent feature is then decoded by the convolutional decoder of the signal processing model to obtain the low-rank matrix of the batch. The first latent feature is used to characterize the evolution of the multiple sensor signals in the time dimension, and the low-rank matrix is ​​used to characterize the normal evolution of the multiple sensor signals in the time dimension when the device is operating normally. Based on the original data matrix and the low-rank matrix, a sparse matrix of the batch is obtained. The sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law. The signal processing model is trained based on the original data matrix, low-rank matrix, and sparse matrix of each batch.

[0006] In some embodiments, the convolutional encoder includes M convolutional layers, each convolutional layer including a convolutional unit, an activation unit, and a pooling unit; the input data of the first convolutional layer is the original data matrix, the input data of the i-th convolutional layer is the output data of the (i-1)-th convolutional layer, and the output data of the M-th convolutional layer is the encoded feature, where i is an integer greater than 1 and not greater than M; The process of encoding the original data matrix using a convolutional encoder based on a signal processing model to obtain encoded features includes: For the i-th convolutional layer, the input data of the convolutional layer is subjected to local feature extraction and feature dimension transformation by the convolutional unit of the i-th convolutional layer to obtain the first output data. The first output data is subjected to nonlinear transformation by the activation unit of the i-th convolutional layer to obtain the second output data. The second output data is subjected to temporal dimension downsampling by the pooling unit of the i-th convolutional layer to obtain the output data of the i-th convolutional layer.

[0007] In some embodiments, the convolutional decoder includes N transposed convolutional layers, the Nth transposed convolutional layer includes a transposed convolutional unit, and each transposed convolutional layer other than the Nth layer includes a transposed convolutional unit and an activation unit; the input data of the 1st transposed convolutional layer is the second latent feature, the input data of the jth transposed convolutional layer is the output data of the (j-1)th transposed convolutional layer, and the output data of the Nth transposed convolutional layer is the low-rank matrix, where j is an integer greater than 1 and not greater than N; The step of decoding the second latent feature using the convolutional decoder of the signal processing model to obtain the low-rank matrix of the batch includes: For the j-th transposed convolutional layer other than the N-th layer, the input data of the j-th transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the j-th transposed convolutional layer to obtain the third output data. The third output data is then subjected to nonlinear transformation through the activation unit of the j-th transposed convolutional layer to obtain the fourth output data. For the Nth transposed convolutional layer, the input data of the Nth transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the Nth transposed convolutional layer to obtain the low-rank matrix.

[0008] In some embodiments, the coding feature includes sub-coding features for each of multiple time steps within the time period, and the first hidden feature includes the sub-hidden features for each of the multiple time steps; The step of extracting latent features from the encoded features using the gated encoder of the signal processing model to obtain the first latent feature includes: For the m-th time step, the gated encoder extracts the latent features of the sub-latent features of the (m-1)-th time step and the sub-coded features of the m-th time step from the encoded features to obtain the sub-latent features of the m-th time step. The gated encoder is used to reduce the number of feature dimensions of the sub-coded features to a preset number.

[0009] In some embodiments, training the signal processing model based on the original data matrix, low-rank matrix, and sparse matrix for each batch includes: For each batch, based on the original data matrix, low-rank matrix, and sparse matrix of the batch, the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch are determined. The reconstruction loss is used to characterize the difference between the matrix reconstructed based on the low-rank matrix and the sparse matrix and the original data matrix. The low-rank regularization loss is used to characterize the low-rank degree and energy level of the low-rank matrix. The sparse regularization loss is used to characterize the overall energy level of the sparse matrix. Based on the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch, the loss function of the batch is determined; The model parameters of the signal processing model are adjusted with the goal of minimizing the loss function.

[0010] In some embodiments, determining the reconstruction loss for each batch based on the original data matrix, low-rank matrix, and sparse matrix of the batch includes: For each batch, a reconstructed data matrix is ​​obtained based on the low-rank matrix and sparse matrix of the batch. Based on the reconstructed data matrix and the original data matrix of the batch, the reconstruction loss of the batch is determined.

[0011] In some embodiments, the low-rank matrix of each batch includes sub-low-rank matrices of multiple training samples in the batch. For each batch, the low-rank regularization loss of the batch is determined based on the original data matrix, the low-rank matrix, and the sparse matrix of the batch, including: For each training sample in each batch, singular value decomposition is performed on the sub-low-rank matrix of the training sample to obtain multiple singular values; The nuclear norm of the training samples is obtained based on the multiple singular values; The low-rank regularized loss of the batch is obtained based on the nuclear norm of each of the multiple training samples in the batch.

[0012] In some embodiments, for each batch, based on the original data matrix, low-rank matrix, and sparse matrix of the batch, the sparse regularization loss of the batch is determined, including: For each batch, determine the sum of the absolute values ​​of multiple elements in the sparse matrix of that batch; The ratio of the sum to the number of the plurality of elements is determined as the sparse regularization loss of the batch.

[0013] In some embodiments, determining the loss function for the batch based on the batch's reconstruction loss, low-rank regularization loss, and sparse regularization loss includes: The loss function for the batch is obtained by weighted summation of the reconstruction loss, low-rank regularization loss, and sparse regularization loss.

[0014] According to another aspect of the embodiments of this disclosure, a fault detection method is provided, the method comprising: Multiple target samples of the target device are acquired, each target sample including multiple sensor signals in the same time period, the multiple sensor signals being different types of signals collected for the target device; Based on the multiple target samples, construct the target data matrix of the target device; The target data matrix is ​​encoded using a convolutional encoder of the signal processing model to obtain encoded features. Latent features are extracted from these encoded features using a gated encoder of the signal processing model to obtain a first latent feature. The first latent feature is then restored using a gated decoder of the signal processing model to obtain a second latent feature. Finally, the second latent feature is decoded using a convolutional decoder of the signal processing model to obtain a target low-rank matrix. The first latent feature characterizes the evolution of the multiple sensor signals in the temporal dimension, and the target low-rank matrix characterizes the normal evolution of the multiple sensor signals in the temporal dimension when the target device is operating normally. The signal processing model is a signal processing model trained using any of the above embodiments. Based on the target data matrix and the target low-rank matrix, a target sparse matrix is ​​obtained. The target sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law. Based on the target sparse matrix, fault detection is performed on the target device.

[0015] In some embodiments, the target sparse matrix includes anomalous signal values ​​of multiple feature dimensions of each target sample at multiple time steps; the fault detection of the target device based on the target sparse matrix includes: For each time step of each target sample, an anomaly score for that time step is determined based on the anomaly signal values ​​of the multiple feature dimensions at that time step. The anomaly score of the target device is determined based on the anomaly score of each target sample at each time step; If the anomaly score of the target device exceeds an anomaly threshold, the target device is determined to have malfunctioned. The anomaly threshold is determined based on the sensor signals when the target device is not malfunctioning.

[0016] In some embodiments, the method further includes: Acquire multiple training samples when the target device is not malfunctioning, each training sample including signal segments of multiple sensor signals of the target device in the same time period; Based on the signal processing model, a sparse matrix of the multiple training samples is obtained; Based on the sparse matrix, the anomaly scores of each training sample at multiple time steps are obtained; The anomaly threshold is obtained based on the mean and standard deviation of the anomaly scores of the multiple training samples at multiple time steps.

[0017] According to another aspect of the embodiments of this disclosure, a training apparatus for a signal processing model is provided, the apparatus comprising: The acquisition module is used to acquire multiple batches of training samples. Each batch of training samples includes multiple training samples, and each training sample includes multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected from the same device. A construction module is used to construct the original data matrix of each batch based on the training samples of each batch; The processing module is used to encode the original data matrix of each batch using a convolutional encoder of the signal processing model to obtain encoded features, extract latent features from the encoded features using a gated encoder of the signal processing model to obtain a first latent feature, restore the first latent feature using a gated decoder of the signal processing model to obtain a second latent feature, and decode the second latent feature using a convolutional decoder of the signal processing model to obtain a low-rank matrix of the batch. The first latent feature is used to characterize the evolution law of the multiple sensor signals in the time dimension, and the low-rank matrix is ​​used to characterize the normal evolution law of the multiple sensor signals in the time dimension when the device is operating normally. The determination module is used to obtain a sparse matrix of the batch based on the original data matrix and the low-rank matrix. The sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law. The training module is used to train the signal processing model based on the original data matrix, low-rank matrix, and sparse matrix for each batch.

[0018] In some embodiments, the convolutional encoder includes M convolutional layers, each convolutional layer including a convolutional unit, an activation unit, and a pooling unit; the input data of the first convolutional layer is the original data matrix, the input data of the i-th convolutional layer is the output data of the (i-1)-th convolutional layer, and the output data of the M-th convolutional layer is the encoded feature, where i is an integer greater than 1 and not greater than M; the processing module is used for: For the i-th convolutional layer, the input data of the convolutional layer is subjected to local feature extraction and feature dimension transformation by the convolutional unit of the i-th convolutional layer to obtain the first output data. The first output data is subjected to nonlinear transformation by the activation unit of the i-th convolutional layer to obtain the second output data. The second output data is subjected to temporal dimension downsampling by the pooling unit of the i-th convolutional layer to obtain the output data of the i-th convolutional layer.

[0019] In some embodiments, the convolutional decoder includes N transposed convolutional layers, the Nth transposed convolutional layer includes a transposed convolutional unit, and each transposed convolutional layer other than the Nth layer includes a transposed convolutional unit and an activation unit; the input data of the 1st transposed convolutional layer is the second latent feature, the input data of the jth transposed convolutional layer is the output data of the (j-1)th transposed convolutional layer, and the output data of the Nth transposed convolutional layer is the low-rank matrix, where j is an integer greater than 1 and not greater than N; the processing module is used for: For the j-th transposed convolutional layer other than the N-th layer, the input data of the j-th transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the j-th transposed convolutional layer to obtain the third output data. The third output data is then subjected to nonlinear transformation through the activation unit of the j-th transposed convolutional layer to obtain the fourth output data. For the Nth transposed convolutional layer, the input data of the Nth transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the Nth transposed convolutional layer to obtain the low-rank matrix.

[0020] In some embodiments, the coding feature includes sub-coding features for each of multiple time steps within the time period, and the first hidden feature includes the sub-hidden features for each of the multiple time steps; the processing module is configured to: For the m-th time step, the gated encoder extracts the latent features of the sub-latent features of the (m-1)-th time step and the sub-coded features of the m-th time step from the encoded features to obtain the sub-latent features of the m-th time step. The gated encoder is used to reduce the number of feature dimensions of the sub-coded features to a preset number.

[0021] In some embodiments, the training module is configured to: For each batch, based on the original data matrix, low-rank matrix, and sparse matrix of the batch, the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch are determined. The reconstruction loss is used to characterize the difference between the matrix reconstructed based on the low-rank matrix and the sparse matrix and the original data matrix. The low-rank regularization loss is used to characterize the low-rank degree and energy level of the low-rank matrix. The sparse regularization loss is used to characterize the overall energy level of the sparse matrix. Based on the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch, the loss function of the batch is determined; The model parameters of the signal processing model are adjusted with the goal of minimizing the loss function.

[0022] In some embodiments, the training module is configured to: For each batch, a reconstructed data matrix is ​​obtained based on the low-rank matrix and sparse matrix of the batch. Based on the reconstructed data matrix of the batch and the original data matrix, the reconstruction loss of the batch is determined.

[0023] In some embodiments, the low-rank matrix of each batch includes sub-low-rank matrices of multiple training samples in the batch, and the training module is configured to: For each training sample in each batch, singular value decomposition is performed on the sub-low-rank matrix of the training sample to obtain multiple singular values; The nuclear norm of the training samples is obtained based on the multiple singular values; The low-rank regularized loss of the batch is obtained based on the nuclear norm of each of the multiple training samples in the batch.

[0024] In some embodiments, the training module is configured to: For each batch, determine the sum of the absolute values ​​of multiple elements in the sparse matrix of that batch; The ratio of the sum to the number of the plurality of elements is determined as the sparse regularization loss of the batch.

[0025] In some embodiments, the training module is configured to: The loss function for the batch is obtained by weighted summation of the reconstruction loss, low-rank regularization loss, and sparse regularization loss.

[0026] According to another aspect of the present disclosure, a fault detection apparatus is provided, the apparatus comprising: The acquisition module is used to acquire multiple target samples of the target device. Each target sample includes signal segments of multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected for the target device. A construction module is used to construct a target data matrix for the target device based on the multiple target samples; The processing module is used to encode the target data matrix using a convolutional encoder of the signal processing model to obtain encoded features, extract latent features from the encoded features using a gated encoder of the signal processing model to obtain a first latent feature, restore the first latent feature using a gated decoder of the signal processing model to obtain a second latent feature, and decode the second latent feature using a convolutional decoder of the signal processing model to obtain a target low-rank matrix. The first latent feature is used to characterize the evolution of the multiple sensor signals in the time-series dimension, and the target low-rank matrix is ​​used to characterize the normal evolution of the multiple sensor signals in the time-series dimension when the target device is operating normally. The signal processing model is a signal processing model trained through any of the above embodiments. The determination module is used to obtain a target sparse matrix based on the target data matrix and the target low-rank matrix, wherein the target sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law; The detection module is used to perform fault detection on the target device based on the target sparse matrix.

[0027] In some embodiments, the target sparse matrix includes anomalous signal values ​​for multiple feature dimensions of each target sample at multiple time steps; the detection module is configured to: For each time step of each target sample, an anomaly score for that time step is determined based on the anomaly signal values ​​of the multiple feature dimensions at that time step. The anomaly score of the target device is determined based on the anomaly score of each target sample at each time step; If the anomaly score of the target device exceeds an anomaly threshold, it is determined that the target device has malfunctioned. The anomaly threshold is determined based on the sensor signals when the target device is not malfunctioning.

[0028] In some embodiments, the apparatus further includes a determining module for: Acquire multiple training samples when the target device is not malfunctioning, each training sample including signal segments of multiple sensor signals of the target device in the same time period; Based on the signal processing model, a sparse matrix of the multiple training samples is obtained; Based on the sparse matrix, the anomaly scores of each training sample at multiple time steps are obtained; The anomaly threshold is obtained based on the mean and standard deviation of the anomaly scores of the multiple training samples at multiple time steps.

[0029] According to another aspect of the embodiments of the present disclosure, a computer device is provided, the computer device comprising: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the training method or fault detection method of the above signal processing model.

[0030] According to another aspect of the present disclosure, a computer-readable storage medium is provided, wherein when the instructions in the computer-readable storage medium are executed by a processor of a computer device, the computer device is enabled to perform the training method or fault detection method of the above-described signal processing model.

[0031] According to another aspect of the present disclosure, a computer program product is provided, the computer program product including a computer program that, when executed by a processor, implements the training method or fault detection method of the above-described signal processing model.

[0032] This application provides a training method for a signal processing model. This method uses the original data matrix constructed from batch training samples as input to the signal processing model. Relying on the convolutional encoder, gated encoder, gated decoder, and convolutional decoder within the signal processing model, it constructs a complete closed-loop architecture for encoding, temporal feature restoration, and data reconstruction. It deeply integrates the low-rank-sparse dual matrix decomposition mechanism of the traditional RPCA scheme with temporal context modeling and local spatial feature extraction capabilities. This effectively compensates for the deficiency of the traditional RPCA scheme in characterizing the temporal evolution of signals. Through the step-by-step feature extraction and reconstruction of the signal processing model, the low-rank matrix accurately fits the temporal variation characteristics of sensor signals under normal equipment operation. Simultaneously, the sparse matrix effectively represents abnormal situations where sensor signals deviate from the normal evolution pattern. Therefore, this method achieves differentiated and collaborative learning of normal temporal signals and abnormal perturbation signals by the signal processing model, ensuring the effectiveness and convergence stability of the model training process and continuously optimizing the model's representation ability and generalization performance. After training, using the sparse matrix output by the signal processing model for equipment fault detection can significantly improve the accuracy of equipment fault detection.

[0033] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0035] Figure 1 This is a schematic diagram illustrating an implementation environment according to an exemplary embodiment.

[0036] Figure 2 This is a flowchart illustrating a training method for a signal processing model according to an exemplary embodiment.

[0037] Figure 3 This is a flowchart illustrating a training method for a signal processing model according to an exemplary embodiment.

[0038] Figure 4 This is a flowchart illustrating a training method for a signal processing model according to an exemplary embodiment.

[0039] Figure 5 This is a flowchart illustrating a fault detection method according to an exemplary embodiment.

[0040] Figure 6 This is a flowchart illustrating model training and fault detection according to an exemplary embodiment.

[0041] Figure 7 This is a block diagram illustrating a training apparatus for a signal processing model according to an exemplary embodiment.

[0042] Figure 8 This is a block diagram illustrating a fault detection device according to an exemplary embodiment.

[0043] Figure 9 This is a schematic diagram of the structure of a terminal according to an exemplary embodiment.

[0044] Figure 10 This is a schematic diagram of the structure of a server according to an exemplary embodiment. Detailed Implementation

[0045] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0046] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0047] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this disclosure are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the sensor signals involved in this disclosure were all obtained under full authorization.

[0048] The signal processing model training method provided in this application can be executed by a computer device, which can be at least one of a terminal and a server. The following is a schematic diagram illustrating the implementation environment of the signal processing model training method provided in this application. See also... Figure 1 , Figure 1 This is a schematic diagram of an implementation environment provided in an embodiment of this application. The implementation environment includes: terminal 101 and server 102.

[0049] In some embodiments, terminal 101 may run a client application with a target application, which provides the function of real-time detection of device faults based on sensor signals. This application does not limit the implementation form of the target application; for example, it may be an application that requires downloading and installation, a mini-program that does not require installation, a web application, etc.

[0050] In this embodiment, terminal 101 can be installed at multiple monitoring locations of the device under test to acquire multiple sensor signals from the device. Server 102 provides background services for the target application. After terminal 101 acquires multiple sensor signals from the device, it transmits these signals to server 102. Server 102 processes the multiple sensor signals using a signal processing model to obtain a low-rank matrix, which in turn generates a sparse matrix. Based on the sparse matrix, fault detection is performed on the device, and the detection results are sent to terminal 101 for output.

[0051] In other embodiments, the terminal 101 itself can also acquire multiple sensor signals from the device and process the multiple sensor signals through a signal processing model to obtain a low-rank matrix, and then obtain a sparse matrix. Based on the sparse matrix, the terminal can perform fault detection on the device and output the detection results.

[0052] Terminal 101 can be a computer device such as a vibration sensor, acoustic sensor, temperature sensor, mobile phone, tablet computer, multimedia playback device, PC (Personal Computer), wearable device, VR (Virtual Reality) device, AR (Augmented Reality) device, MR (Mixed Reality) device, etc., with signal processing and output display functions. Server 102 can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed file system. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Server 102 and terminal 101 are directly or indirectly connected via wired or wireless communication.

[0053] Figure 2 This is a flowchart illustrating a training method for a signal processing model according to an exemplary embodiment, such as... Figure 2 As shown, the method is performed by a computer device and includes at least one of the following steps.

[0054] 201. Obtain multiple batches of training samples. Each batch of training samples includes multiple training samples. Each training sample includes multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected from the same device.

[0055] In this embodiment, the sensor signals can be signals collected by devices in any scenario, such as industrial automation, medical equipment, aerospace, or outdoor monitoring. Multiple sensor signals are different types of physical detection signals, with each type corresponding to a feature dimension. The types of multiple sensor signals can be flexibly set according to requirements. For example, the multiple sensor signals can be vibration signals, temperature signals, and sound signals.

[0056] Each sensor signal comprises signal values ​​from multiple time steps. For a signal segment within a training sample, the number of time steps covered by that time segment is determined based on the signal's sampling rate. For example, a signal segment from 0 to 2 seconds could include signal values ​​from 100 time steps.

[0057] It should be noted that the number of training samples included in each batch may be the same or different. In this embodiment, the example of each batch including the same number of training samples is used for illustration. The training samples included in each batch may be for the same or different devices.

[0058] 202. Based on the training samples of each batch, construct the original data matrix for that batch.

[0059] Each batch includes multiple training samples, each training sample includes multiple signal segments from sensor signals within the same time period, and each signal segment includes signal values ​​at multiple time steps. Accordingly, the original data matrix can be represented as follows: .in, This represents the batch size, which is the number of training samples in that batch. This indicates the time step length, which is the number of time steps. This represents the number of feature dimensions (input_dim), which is also the number of sensor signals. An element in the original data matrix represents the signal value of one feature dimension of a training sample at one time step.

[0060] The batch size, time series length, and number of feature dimensions can all be set as needed. For example, if the number of sensor signals is 3, then... For example, For example, .

[0061] Optionally, the input format of the signal processing model may differ from the format of the original data matrix, in which case dimensionality permutation of the original data matrix is ​​required. For example, the format of the original data matrix is... Then, perform a dimension permutation on it. The original data matrix after dimension permutation is represented as follows: .in, This represents the original data matrix before dimension permutation. This represents the original data matrix after dimension permutation.

[0062] 203. For each batch of original data matrix, the original data matrix is ​​encoded by the convolutional encoder of the signal processing model to obtain encoded features. The encoded features are then extracted by the gated encoder of the signal processing model to obtain the first hidden feature. The first hidden feature is then restored by the gated decoder of the signal processing model to obtain the second hidden feature. The second hidden feature is then decoded by the convolutional decoder of the signal processing model to obtain the low-rank matrix of the batch. The first hidden feature is used to characterize the evolution of multiple sensor signals in the time dimension, and the low-rank matrix is ​​used to characterize the normal evolution of multiple sensor signals in the time dimension when the equipment is running normally.

[0063] In this embodiment, the encoded features, the first hidden feature, and the second hidden feature can all be matrices. The convolutional encoder and the gated encoder constitute the convolutional module of the signal processing model. The convolutional decoder and the gated decoder constitute the decoding module of the signal processing model.

[0064] Convolutional encoders are used to compress and reduce the dimensionality of the input raw data matrix. If the original data matrix contains a large amount of redundant information, a convolutional encoder, through operations such as convolution, activation, and pooling, transforms the high-dimensional raw data matrix into low-dimensional encoded features. This preserves the key information of the original data matrix while removing redundant information. Convolutional decoders are used to restore the dimensionality and reconstruct the details of abstract, deep latent features. Relying on operations such as transposed convolution and temporal upsampling, convolutional decoders progressively restore the feature dimension and temporal length, restoring the low-dimensional latent features to an output dimension that matches the original data matrix. This supplements and refines signal details, accurately reconstructing a low-rank matrix representing the normal evolution of the device.

[0065] The Gated Recurrent Unit (GRU) encoder extracts the global correlation features and dynamic evolution patterns of encoded features in the temporal dimension. Relying on a gating memory mechanism, it fuses historical hidden states with the encoded features of the current time step step by step, uncovering long-distance temporal dependencies between multi-source sensor signals, suppressing random noise interference, and outputting hidden features with strong temporal continuity. The Gated Decoder receives these temporal hidden features, performs inverse deduction of temporal logic and state reconstruction, adaptively reproduces the temporal change trends and periodic evolution characteristics of the signal, ensures the stable transmission of temporal context information during decoding, provides temporal constraint support for the accurate reconstruction of the convolutional decoder, and thus achieves effective decoupling between normal temporal components and abnormally sparse components.

[0066] 204. Based on the original data matrix and the low-rank matrix, the sparse matrix of this batch is obtained. The sparse matrix is ​​used to characterize the abnormal situation of multiple sensor signals relative to the normal evolution law.

[0067] For example, the residual between the original data matrix and the low-rank matrix is ​​determined as a sparse matrix. The original data matrix, the low-rank matrix, and the sparse matrix have the same size.

[0068] 205. Train the signal processing model based on the original data matrix, low-rank matrix, and sparse matrix of each batch.

[0069] In this embodiment, the loss function for each batch is determined based on the original data matrix, low-rank matrix, and sparse matrix, and the model parameters of the signal processing model are trained based on the loss function.

[0070] In this embodiment, the signal processing model is trained iteratively using multiple batches of training samples, meaning that each batch of training samples is used for one iteration.

[0071] Optionally, the signal processing model is trained iteratively based on multiple batches of training samples until an iteration stopping condition is met. The iteration stopping condition includes at least one of the following: the number of iterations reaches a preset number, or the loss function reaches a loss threshold.

[0072] This application provides a training method for a signal processing model. This method uses the original data matrix constructed from batch training samples as input to the signal processing model. Relying on the convolutional encoder, gated encoder, gated decoder, and convolutional decoder within the signal processing model, it constructs a complete closed-loop architecture for encoding, temporal feature restoration, and data reconstruction. It deeply integrates the low-rank-sparse dual matrix decomposition mechanism of the traditional RPCA scheme with temporal context modeling and local spatial feature extraction capabilities. This effectively compensates for the deficiency of the traditional RPCA scheme in characterizing the temporal evolution of signals. Through the step-by-step feature extraction and reconstruction of the signal processing model, the low-rank matrix accurately fits the temporal variation characteristics of sensor signals under normal equipment operation. Simultaneously, the sparse matrix effectively represents abnormal situations where sensor signals deviate from the normal evolution pattern. Therefore, this method achieves differentiated and collaborative learning of normal temporal signals and abnormal perturbation signals by the signal processing model, ensuring the effectiveness and convergence stability of the model training process and continuously optimizing the model's representation ability and generalization performance. After training, using the sparse matrix output by the signal processing model for equipment fault detection can significantly improve the accuracy of equipment fault detection.

[0073] The above Figure 2 The diagram shown is merely the basic process of this disclosure. The following section, based on a specific implementation method, further elaborates on the solution provided in this disclosure. See also... Figure 3 , Figure 3 This is a flowchart illustrating a training method for a signal processing model according to an exemplary embodiment. The method is performed by a computer device and includes at least one of the following steps.

[0074] 301. Obtain multiple batches of training samples. Each batch of training samples includes multiple training samples. Each training sample includes multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected from the same device.

[0075] In some embodiments, each training sample includes a signal segment based on preprocessing of the original signal segment. Preprocessing includes standardization or normalization. Preprocessing is used to eliminate dimensional differences between different sensor signals, ensuring uniform numerical ranges and distributions, and preventing one type of sensor signal from dominating model training due to excessively large values, while other types of sensor signals are overwhelmed.

[0076] It should be noted that for each sensor signal, since signal segments from different time periods can be used for different training samples, the entire sensor signal can be preprocessed uniformly to facilitate the direct extraction of signal segments later. In some other embodiments, preprocessing can be performed only on signal segments in the training samples; this is not specifically limited here.

[0077] The standardization algorithm can be set as needed. For example, the Z-score algorithm is used for standardization. The specific process of this algorithm is as follows: For each original signal segment, the mean and standard deviation of the original signal values ​​at multiple time steps in the original signal segment are determined; for the original signal value at each time step, the difference between the original signal value and the mean is determined, and the ratio of the difference to the standard deviation is determined as the standardized signal value at that time step. The standardized signal values ​​at multiple time steps constitute the preprocessed signal segment.

[0078] The normalization algorithm can be set as needed. For example, the Min-Max algorithm is used. The specific process of this algorithm is as follows: For each original signal segment, determine its maximum and minimum original signal values; for the original signal value at each time step, determine the first difference between the original signal value and the minimum original signal value, and determine the second difference between the maximum and minimum original signal values. The ratio of the first difference to the second difference is determined as the normalized signal value at that time step, with a value range of [0,1]. The normalized signal values ​​at multiple time steps constitute the preprocessed signal segment.

[0079] 302. Based on the training samples of each batch, construct the original data matrix of the batch.

[0080] This step is the same as step 202 above, and will not be repeated here.

[0081] 303. For each batch of original data matrix, the original data matrix is ​​encoded by the convolutional encoder of the signal processing model to obtain the encoded features.

[0082] The convolutional encoder includes M convolutional layers, each of which includes a convolutional unit, an activation unit, and a pooling unit. The input data of the first convolutional layer is the original data matrix, the input data of the i-th convolutional layer is the output data of the (i-1)-th convolutional layer, and the output data of the M-th convolutional layer is the encoded feature, where i is an integer greater than 1 and not greater than M.

[0083] The process described above, in which the original data matrix is ​​encoded by a convolutional encoder of a signal processing model to obtain encoded features, includes the following implementation: For the i-th convolutional layer, the input data of the convolutional layer is subjected to local feature extraction and feature dimension transformation by the convolutional unit of the i-th convolutional layer to obtain the first output data; the first output data is subjected to nonlinear transformation by the activation unit of the i-th convolutional layer to obtain the second output data; and the second output data is subjected to temporal dimension downsampling by the pooling unit of the i-th convolutional layer to obtain the output data of the i-th convolutional layer.

[0084] For example, taking a convolutional encoder consisting of two convolutional layers as an example, the convolutional unit ensures spatial consistency of the feature dimension through a 3×1 convolutional kernel and the same padding. The pooling unit performs downsampling with a stride of 2. After two downsampling steps, a 4x data compression is achieved, such as compressing 100 time steps of input data to 25 time steps, while simultaneously completing the dimensionality increase and decrease of the feature dimension (3→16→8). The specific process of obtaining the encoded features through two convolutional layers is shown in the following formula (1): , (1) in, Represents the original data matrix. This is the original data matrix after dimension permutation, adapted to the channel-first format of the convolutional unit. This represents the kernel parameters of the convolutional units in the first convolutional layer. This represents the bias parameter of the convolutional unit in the first convolutional layer. This represents the kernel parameters of the convolutional units in the second convolutional layer. This represents the bias parameter of the convolutional unit in the second convolutional layer. This represents a convolutional unit. This indicates the activation unit. This represents a pooling unit. This represents the output data of the first convolutional layer. This represents the output data of the second convolutional layer, i.e., the encoded features. 8 indicates the number of feature dimensions. This indicates the number of time steps after compression. Optionally, the feature shape of the original data matrix may change during the processing of the original data matrix by the convolutional encoder.

[0085] In this embodiment, convolutional units extract local features and transform feature dimensions from various sensor signals, effectively capturing local pattern features and providing high-dimensional structured features for subsequent time-series modeling. Activation units introduce nonlinear transformations to enhance the signal processing model's ability to express complex signal patterns. Pooling units compress redundant data through time-series downsampling, reducing the computational complexity of subsequent gating time-series modeling and providing high-quality feature input for the gated encoder to learn global time-series features. This convolutional encoder, through progressive processing across multiple convolutional layers, achieves gradual feature purification and enhancement, reduces noise interference, ensures the accuracy of subsequent low-rank matrix reconstruction of the device's normal time-series patterns, and provides a reliable foundation for fault detection.

[0086] 304. The gated encoder of the signal processing model extracts the latent features of the encoded features to obtain the first latent feature. The first latent feature is used to characterize the evolution of multiple sensor signals in the time dimension.

[0087] In this embodiment, the encoded features include sub-encoded features of each of the multiple time steps within the time period, and the first hidden feature includes sub-hidden features of each of the multiple time steps. The process of extracting hidden features from the encoded features using the gated encoder of the signal processing model to obtain the first hidden feature includes the following implementation: For the m-th time step, the gated encoder extracts hidden features from the sub-hidden features of the (m-1)-th time step and the sub-encoded features of the m-th time step in the encoded features to obtain the sub-hidden features of the m-th time step. The gated encoder is used to reduce the number of feature dimensions of the sub-encoded features to a preset number.

[0088] The preset number can be set as needed; this preset number is also the hidden feature dimension in the gate encoder. If the preset quantity is 2, then =2.

[0089] The process of obtaining the sub-hidden features at each time step through a gated encoder can be represented by the following formula (2): (2) in, Represents the dimensionality of the encoded feature after permutation. Sub-coded features at each time step, . Indicates the first Sub-hidden features at each time step . Indicates the first -1 time step of the hidden features. This represents a gated encoder. Based on the sub-hidden features of multiple time steps, the final output is the first hidden feature. H .

[0090] Optionally, for the first time step, the sub-hidden features of the corresponding previous time step are used. Set as a zero-value vector to complete the initial state configuration for time-series iterative computation.

[0091] In this embodiment, the gate encoder processes data only in ascending order of time (from the first time step to the last time step). The hidden state of each step depends only on the input of the current time step and the historical information of the previous time steps, and does not use the data of future time steps. Therefore, it is a one-way gate encoder. The gate decoder is the same and is a one-way gate decoder.

[0092] In this embodiment, the gated encoder uses a time-series traversal approach, combining the sub-encoded features of the current time step with the sub-latent features of the previous time step for joint calculation. This continuously preserves the correlation information, signal change trends, and periodic evolution patterns between consecutive time steps, overcoming the shortcomings of traditional static decomposition in failing to mine time-series features. This effectively mines the time-series correlation and dynamic evolution patterns of various sensor signals. At the same time, the feature dimensions are compressed and simplified to reduce feature redundancy and computational load, improve the purity of latent feature expression, and ensure the accuracy of subsequent reconstruction of the low-rank matrix of the normal time-series pattern.

[0093] 305. The first latent feature is restored by the gated decoder of the signal processing model to obtain the second latent feature.

[0094] In this embodiment, a gated decoder is connected before the convolutional decoder to restore the low-dimensional first latent feature to feature data with the same dimension as the output dimension of the convolutional encoder. Correspondingly, the second latent feature includes multiple restored sub-latent features for each time step. The process of restoring the first latent feature to obtain the second latent feature through the gated decoder of the signal processing model includes the following implementation: For the nth time step, the gated decoder restores the latent features of the restored sub-latent features of the (n-1)th time step and the sub-latent features of the first latent feature at the nth time step to obtain the restored sub-latent features of the nth time step. The gated decoder is used to restore the feature dimension of the sub-latent features.

[0095] Accordingly, the process of obtaining the second latent feature through the gated decoder can be represented by the following formula (3): (3) in, Indicates the output of the gated decoder. Restored sub-hidden features at each time step . Indicates the first Restored sub-hidden features at each time step. This represents the gated decoder. Based on the reconstructed sub-hidden features at multiple time steps, the final output is the second hidden feature. .

[0096] Optionally, for the first time step, the corresponding previous time step's reconstructed sub-hidden features are used. Set as a zero-value vector to complete the initial state configuration for time-series iterative computation.

[0097] 306. The second hidden feature is decoded by the convolutional decoder of the signal processing model to obtain the low-rank matrix of this batch. The low-rank matrix is ​​used to characterize the normal evolution of multiple sensor signals in the time dimension when the equipment is running normally.

[0098] The convolutional decoder includes N transposed convolutional layers. The Nth transposed convolutional layer includes transposed convolutional units, and each transposed convolutional layer other than the Nth layer includes transposed convolutional units and activation units. The input data of the first transposed convolutional layer is the second latent feature, the input data of the jth transposed convolutional layer is the output data of the (j-1)th transposed convolutional layer, and the output data of the Nth transposed convolutional layer is a low-rank matrix, where j is an integer greater than 1 and not greater than N.

[0099] Accordingly, the process of decoding the second latent features using the convolutional decoder of the signal processing model to obtain the low-rank matrix of this batch includes the following implementation: For the j-th transposed convolutional layer other than the N-th layer, the input data of the j-th transposed convolutional layer is restored in terms of feature dimension and temporal dimension by the transposed convolutional unit of the j-th transposed convolutional layer to obtain the third output data. The third output data is then subjected to a nonlinear transformation by the activation unit of the j-th transposed convolutional layer to obtain the fourth output data. For the N-th transposed convolutional layer, the input data of the N-th transposed convolutional layer is restored in terms of feature dimension and temporal dimension by the transposed convolutional unit of the N-th transposed convolutional layer to obtain the low-rank matrix.

[0100] For example, the convolutional decoder includes two transposed convolutional layers. The transposed decoder uses two one-dimensional transposed convolutional layers (ConvTranspose1d) that are symmetrical to the convolutional encoder to perform upsampling and remove unnecessary output padding to ensure that the data dimension after upsampling is completely matched with the original data matrix.

[0101] The specific process of obtaining low-rank features through two transposed convolutional layers is shown in the following formula (4): , (4) in, , is the tensor after permutation of the second latent feature dimension, adapted to the channel-first format of the convolutional unit. This represents the kernel parameters of the transposed convolutional units in the first transposed convolutional layer. This represents the bias parameter of the transposed convolutional unit in the first transposed convolutional layer. This represents the kernel parameters of the transposed convolutional units in the second transposed convolutional layer. This represents the bias parameter of the transposed convolutional unit in the second transposed convolutional layer. This represents the transposed convolution unit. This indicates the activation unit. This represents the output data of the first transposed convolutional layer. This represents the output data of the second transposed convolutional layer. Optionally, for Perform dimensional permutation, which yields a lower-order matrix. .

[0102] In this embodiment, the temporal continuity and feature dimension of the low-rank matrix are accurately restored by using a symmetrical progressive upsampling structure of convolutional decoder and convolutional encoder. At the same time, the distortion of the low-rank matrix is ​​avoided by using an inactive output layer, providing a high-quality low-rank matrix for subsequent sparse matrix extraction.

[0103] 307. Based on the original data matrix and the low-rank matrix, a sparse matrix for the batch is obtained. The sparse matrix is ​​used to characterize the abnormal situation of multiple sensor signals relative to the normal evolution law.

[0104] In this embodiment, the sparse matrix is ​​the residual between the original data matrix and the low-rank matrix. Based on the original data matrix and the low-rank matrix, the sparse matrix is ​​obtained, which can be expressed by the following formula (5): (5) in, Represents the original data matrix. Represents a low-rank matrix. This represents a sparse matrix.

[0105] In this embodiment, the core logic of the RPCA algorithm for bi-matrix factorization is implemented, ensuring data dimension matching throughout the process and eliminating the risk of dimension conflicts. Specifically, the second latent feature is restored to a low-rank matrix with the same input format as the original data matrix through a transposed convolutional layer. During the forward propagation process, the signal processing model first extracts the low-rank matrix using an encoder-decoder architecture, and then calculates the sparse matrix by the direct residual between the original data matrix and the low-rank matrix. This implements the core logic of RPCA bi-matrix factorization, ensuring data dimension matching throughout the process and eliminating the risk of dimension conflicts.

[0106] 308. For each batch, based on the original data matrix, low-rank matrix, and sparse matrix of that batch, determine the reconstruction loss, low-rank regularization loss, and sparse regularization loss of that batch. The reconstruction loss is used to characterize the difference between the matrix reconstructed based on the low-rank matrix and the sparse matrix and the original data matrix. The low-rank regularization loss is used to characterize the low-rank degree and energy level of the low-rank matrix. The sparse regularization loss is used to characterize the overall energy level of the sparse matrix.

[0107] In some embodiments, the process of determining the reconstruction loss of each batch based on the original data matrix, low-rank matrix, and sparse matrix of that batch includes the following implementation: for each batch, the reconstructed data matrix of that batch is obtained based on the low-rank matrix and sparse matrix of that batch, and the reconstruction loss of that batch is determined based on the reconstructed data matrix and the original data matrix of that batch.

[0108] Optionally, the reconstructed data matrix of the batch can be obtained based on the low-rank matrix and sparse matrix of the batch by summing the low-rank matrix and sparse matrix of the batch.

[0109] Optionally, based on the reconstructed data matrix and the original data matrix, the reconstruction loss for this batch is determined by taking the mean squared error (MSE) of the reconstructed data matrix and the original data matrix as the reconstruction loss.

[0110] Accordingly, the process of determining the reconstruction loss can be expressed by the following formula (6): (6) in, Indicates the reconstruction loss. Represents the first element in the original data matrix. The training sample of the th training sample The feature dimension at the th feature dimension Signal values ​​at each time step Represents the rank of the low-rank matrix. The training sample of the th training sample The feature dimension at the th feature dimension Signal values ​​at each time step Represents the th element in a sparse matrix. The training sample of the th training sample The feature dimension at the th feature dimension Signal values ​​at each time step Represents the first element in the reconstructed data matrix. The training sample of the th training sample The feature dimension at the th feature dimension The signal value at each time step.

[0111] In this embodiment, the reconstruction loss is calculated using mean squared error, which measures the difference between the input and output data, ensuring that the RPCA input data can be reconstructed from the superposition of low-rank and sparse matrices, thus fulfilling the equality constraint. Other algorithms can also be used to calculate the reconstruction loss; no specific limitation is made here.

[0112] In this embodiment, a reconstructed data matrix is ​​obtained by superimposing and fusing a low-rank matrix and a sparse matrix, and then the reconstruction loss is calculated. This can constrain the collaborative decomposition effect of the two components from an overall perspective, simultaneously optimize the preservation of normal time-series features and the separation of local abnormal information, ensure the overall accuracy of matrix restoration, and improve the rationality of signal decomposition and the reliability of fault detection.

[0113] In some embodiments, the low-rank matrix of each batch includes the sub-low-rank matrices of multiple training samples in that batch. For each batch, the process of determining the low-rank regularization loss of that batch based on the original data matrix, low-rank matrix, and sparse matrix of that batch includes the following implementation: for each training sample in each batch, perform singular value decomposition on the sub-low-rank matrix of the training sample to obtain multiple singular values; obtain the nuclear norm of the training sample based on the multiple singular values; and obtain the low-rank regularization loss of the batch based on the nuclear norm of each of the multiple training samples in that batch.

[0114] For example, the process of obtaining the nuclear norm of the training sample based on multiple singular values ​​is represented by the following formula (7): (7) in, Indicates the first The sub-low-rank matrix of each training sample. . express The A singular value. Indicates taking and The minimum value in. Indicates the first The nuclear norm of each training sample.

[0115] Optionally, the low-rank regularization loss of the batch is obtained based on the nuclear norm of each of the multiple training samples in the batch. That is, the mean of the nuclear norm of each of the multiple training samples in the batch is determined as the low-rank regularization loss of the batch. Accordingly, the process of determining the low-rank regularization loss can be represented by the following formula (8): (8) in, This indicates the low-rank regularized loss of this batch. This indicates the number of training samples in this batch.

[0116] In this embodiment, by performing singular value decomposition on the sub-low-rank matrix of each training sample, calculating the nuclear norm based on the singular value distribution characteristics, and using the nuclear norm as a convex relaxation constraint on the matrix rank, the size of the rank of the low-rank matrix can be effectively quantified and limited. This allows the common temporal patterns, periodic trends, and stable operating condition characteristics of stable equipment operation to converge to the low-rank matrix. Furthermore, by solving for the nuclear norm based on singular value decomposition, and using the sample-by-sample nuclear norm to jointly constrain the batch low-rank regularization loss, the rank characteristics and information complexity of the low-rank matrix are precisely controlled. This effectively strengthens the low-rank representation of normal time-series signals, suppresses noise and redundant interference, and ensures the accuracy and stability of decomposition of various sensor signals.

[0117] In some embodiments, the process of determining the sparse regularization loss for each batch based on the original data matrix, low-rank matrix, and sparse matrix of that batch includes the following implementation: for each batch, determining the sum of the absolute values ​​of multiple elements in the sparse matrix of that batch; and determining the ratio of the sum to the number of multiple elements as the sparse regularization loss for that batch.

[0118] Accordingly, the process of determining the sparse regularized loss can be represented by the following formula (9): (9) in, Indicates sparse regularized loss. Represents the th element in a sparse matrix. The training sample of the th training sample The feature dimension at the th feature dimension The abnormal signal value at each time step, that is, the value of a certain element in the sparse matrix.

[0119] In this embodiment, the total abnormal amplitude is quantified by summing the absolute values ​​of all elements in the sparse matrix. Utilizing the inherent sparsity-inducing properties of absolute value summation, the sparse matrix is ​​continuously constrained to have as few non-zero elements as possible and as small as possible in amplitude. This drives random noise and normal small fluctuations into the low-rank matrix, retaining only the true local anomalies caused by equipment failures. Furthermore, by combining the summation of the absolute values ​​of sparse matrix elements with global average normalization, sparsity regularization constraints on the abnormal components are achieved. This effectively suppresses irrelevant noise and normal disturbances, eliminates the influence of dimensionality and data units, and improves the anomaly separation effect of various sensor signals.

[0120] 309. Based on the reconstruction loss, low-rank regularization loss, and sparse regularization loss of this batch, determine the loss function for this batch.

[0121] In some embodiments, the process of determining the loss function of a batch based on the reconstruction loss, low-rank regularization loss, and sparse regularization loss of that batch includes the following implementation: weighted summation of the reconstruction loss, low-rank regularization loss, and sparse regularization loss of that batch to obtain the loss function of that batch.

[0122] The weights of the reconstruction loss, low-rank regularization loss, and sparse regularization loss can be set as needed. Optionally, the weight of the reconstruction loss remains fixed at 1, meaning that no additional weight needs to be set for the reconstruction loss.

[0123] For low-rank regularized loss and sparse regularized loss, the weights of the two can be set as needed. The weight of low-rank regularized loss is used to adjust the contribution of low-rank regularized loss to the loss function, and the weight of sparse regularized loss is used to adjust the contribution of sparse regularized loss to the loss function.

[0124] It should be noted that, in order to enable stable and coordinated optimization of reconstruction loss, low-rank regularization loss, and sparse regularization loss, batch mean normalization is used in the above embodiments to unify the kernel norm and Frobenius norm, which originally had large differences in magnitude, to the same scale. In this way, low-rank regularization loss and sparse regularization loss can use fixed weights to complete training directly without repeated parameter tuning, thereby reducing the parameter tuning cost for engineering implementation.

[0125] Optionally, the weight of the low-rank regularization loss is 0.01, allowing the signal processing model to focus more on learning reconstruction accuracy, while the low-rank constraint only plays a mild guiding role. The weight of the sparse regularization loss is 0.1, allowing the signal processing model to focus more on compressing outlier components to make them sparse.

[0126] For example, the loss function can be expressed by the following formula (10): (10) in, Represents the loss function. Indicates the reconstruction loss. This represents the low-rank regularization loss. Indicates sparse regularized loss. The weights represent the low-rank regularization loss. The weights represent the sparse regularization loss.

[0127] In this embodiment, by setting independent weights for reconstruction loss, low-rank regularization loss, and sparse regularization loss and then weighting and fusing them, the three optimization objectives of signal reconstruction accuracy, low-rank prior constraints, and sparse prior constraints can be flexibly balanced. This avoids a single loss dominating the training process and prevents the model from being overly biased towards any one of the data fitting, low-rank suppression, or sparsity constraints. This makes the signal decomposition logic more in line with the inherent characteristics of the device's time-series data and improves the signal decomposition effect.

[0128] 310. Adjust the model parameters of the signal processing model with the goal of minimizing the loss function.

[0129] Optionally, the model parameters are initialized before training the signal processing model to provide an initial state for subsequent gradient descent optimization. During each iteration based on each batch of training samples, the gradient of the loss function with respect to the model parameters is calculated according to the chain rule, and backpropagation is performed. The Adam optimizer updates the model parameters based on the calculated gradient, causing the loss function to gradually decrease, achieving model parameter adjustment with the goal of minimizing the loss function. Optionally, the initial learning rate of the Adam optimizer is set to... When the iteration stopping condition is met, it is determined that the signal processing model has converged and learned the timing patterns of normal device signals, thus completing the training of the signal processing model.

[0130] In this embodiment of the application, the process of training a signal processing model based on the original data matrix, low-rank matrix and sparse matrix of each batch is realized through the above steps 308-310. It should be noted that this implementation method is only an optional implementation method of the process. The process can also be implemented in other ways, such as adjusting the model parameters based on only one or two of the three losses.

[0131] In this embodiment, the reconstruction loss quantifies the difference between the reconstructed matrix and the original data matrix, constraining the overall model to fit the actual sensor data distribution of the device. This ensures that the decoded and reconstructed low-rank matrix can closely match the temporal variation trends of multiple sensor signals, preventing feature reconstruction distortion and providing a reliable benchmark for anomaly residual calculation. The low-rank regularization loss continuously constrains the rank characteristics and energy distribution of the low-rank matrix, prompting the signal processing model to converge the common temporal features of stable device operation to the low-rank matrix, suppressing normal random fluctuations from mixing into the sparse matrix, and conforming to the characteristics of long-term stable device operation. The sparse regularization loss limits the overall energy level of the sparse matrix, driving the anomalous components to maintain sparse distribution characteristics, retaining only real fault features such as local mutations and short-term deviations, and reducing invalid residual interference. Thus, by combining the reconstruction loss with the joint constraint mechanism of low-rank and sparse dual regularization losses, the signal processing model optimization is jointly supervised from multiple dimensions, including reconstruction accuracy, inherent matrix properties, and anomaly distribution characteristics. This significantly improves signal decomposition quality and thus enhances model training efficiency.

[0132] For example, see Figure 4 , Figure 4This is a flowchart illustrating the training process of a signal processing model according to an exemplary embodiment. The model inputs signal segments with three feature dimensions: sound, vibration, and temperature. After standardization preprocessing, the signal segments are input into the signal processing model. First, they enter the convolution module, where the convolutional units of the convolutional encoder perform two convolutional and activation processes on the input data. The pooling units of the convolutional encoder then perform two pooling processes on the input data, achieving a 4x compression of the time step to obtain coded features. Next, a one-way gated encoder extracts low-rank latent features from the coded features to obtain the first latent feature, which is then input into the decoding module for processing. The decoding module includes a one-way gated decoder and a convolutional decoder. After the first latent feature is restored by the one-way gated decoder, the second latent feature is obtained. The second latent feature is then upsampled twice by the transposed convolutional layer of the convolutional decoder to obtain a low-rank matrix. The residual between the original data matrix and the low-rank matrix is ​​calculated to obtain a sparse matrix, achieving a bimatrix decomposition of the original data matrix. Finally, based on the determined reconstruction loss, low-rank regularization loss, and sparse regularization loss, a loss function is obtained to optimize the model parameters of the signal processing model.

[0133] This application provides a training method for a signal processing model. This method uses the original data matrix constructed from batch training samples as input to the signal processing model. Relying on the convolutional encoder, gated encoder, gated decoder, and convolutional decoder within the signal processing model, it constructs a complete closed-loop architecture for encoding, temporal feature restoration, and data reconstruction. It deeply integrates the low-rank-sparse dual matrix decomposition mechanism of the traditional RPCA scheme with temporal context modeling and local spatial feature extraction capabilities. This effectively compensates for the deficiency of the traditional RPCA scheme in characterizing the temporal evolution of signals. Through the step-by-step feature extraction and reconstruction of the signal processing model, the low-rank matrix accurately fits the temporal variation characteristics of sensor signals under normal equipment operation. Simultaneously, the sparse matrix effectively represents abnormal situations where sensor signals deviate from the normal evolution pattern. Therefore, this method achieves differentiated and collaborative learning of normal temporal signals and abnormal perturbation signals by the signal processing model, ensuring the effectiveness and convergence stability of the model training process and continuously optimizing the model's representation ability and generalization performance. After training, using the sparse matrix output by the signal processing model for equipment fault detection can significantly improve the accuracy of equipment fault detection.

[0134] In this embodiment, fault detection is performed using a signal processing model trained through any of the above embodiments. The fault detection process is described in [reference needed]. Figure 5 , Figure 5 This is a flowchart of a fault detection method provided in an embodiment of this application, which includes at least one of the following steps.

[0135] 501. Acquire multiple target samples of the target device, each target sample including multiple sensor signals in the same time period, wherein the multiple sensor signals are different types of signals collected for the target device.

[0136] The sensor signals in this step are the same as those in the above embodiments, and will not be described again here.

[0137] 502. Based on multiple target samples, construct the target data matrix of the target device.

[0138] In this step, the process of constructing the target data matrix based on multiple target samples is the same as the process of constructing the original data matrix based on multiple training samples in step 202, and will not be repeated here.

[0139] 503. The target data matrix is ​​encoded using a convolutional encoder of the signal processing model to obtain encoded features. Latent features are extracted from the encoded features using a gated encoder of the signal processing model to obtain the first latent feature. The first latent feature is restored using a gated decoder of the signal processing model to obtain the second latent feature. The second latent feature is decoded using a convolutional decoder of the signal processing model to obtain the target low-rank matrix. The first latent feature is used to characterize the evolution of multiple sensor signals in the time dimension. The target low-rank matrix is ​​used to characterize the normal evolution of multiple sensor signals in the time dimension when the target device is operating normally.

[0140] In this step, the processing of the target data matrix using the signal processing model is the same as the processing of the original data matrix using the signal processing model in the above embodiment, and will not be repeated here.

[0141] 504. Based on the target data matrix and the target low-rank matrix, the target sparse matrix is ​​obtained. The target sparse matrix is ​​used to characterize the abnormal situation of multiple sensor signals relative to the normal evolution law.

[0142] In this step, the process of obtaining the target sparse matrix based on the target data matrix and the target low-rank matrix is ​​the same as the process of obtaining the sparse matrix based on the original data matrix and the low-rank matrix in step 204 above, and will not be repeated here.

[0143] 505. Based on the target sparse matrix, perform fault detection on the target equipment.

[0144] In some embodiments, the target sparse matrix includes the abnormal signal values ​​of multiple feature dimensions of each target sample at multiple time steps; then the above-mentioned process of fault detection of the target device based on the target sparse matrix includes the following implementation: for each time step of each target sample, an abnormal score for that time step is determined based on the abnormal signal values ​​of multiple feature dimensions at that time step; an abnormal score for the target device is determined based on the abnormal score for each time step of each target sample; if the abnormal score of the target device exceeds an abnormal threshold, it is determined that the target device has failed, and the abnormal threshold is determined based on the sensor signal when the target device has not failed.

[0145] Optionally, for each time step, the anomaly score for that time step is determined based on the anomaly signal values ​​of multiple feature dimensions at that time step. That is, the mean of the absolute values ​​of the anomaly signal values ​​of multiple feature dimensions at that time step is determined as the anomaly score for that time step. The process of obtaining the anomaly score can be expressed by the following formula (11): (11) in, Indicates the first The target sample at the th Abnormal scores at each time step Indicates the first The first target sample The feature dimension at the th feature dimension Abnormal signal values ​​at each time step This indicates the number of feature dimensions.

[0146] In this embodiment, the abnormal score is jointly calculated by combining the abnormal signal values ​​of multiple sensors such as sound, vibration and temperature at the same time step. By integrating the deviation characteristics of multi-dimensional sensing, the problem of misjudgment by a single type of sensor signal is avoided. This comprehensively reflects the degree of abnormal deviation of the overall operating status of the equipment. Then, the abnormal threshold calibrated by normal operating data is used for fault identification. This not only enables accurate fault location in the time dimension and comprehensive capture of abnormal features, but also effectively distinguishes between normal fluctuations and real faults, greatly improving the accuracy and real-time performance of equipment fault detection.

[0147] In some embodiments, the process of determining the above-mentioned anomaly threshold includes the following implementation: acquiring multiple training samples when the target device is not malfunctioning, each training sample including signal segments of multiple sensor signals of the target device at the same time period; obtaining a sparse matrix of multiple training samples based on a signal processing model; obtaining the anomaly score of each training sample at multiple time steps based on the sparse matrix; and obtaining the anomaly threshold based on the mean and standard deviation of the anomaly scores of multiple training samples at multiple time steps.

[0148] For example, the process of obtaining the anomaly threshold based on the mean and standard deviation of the anomaly scores of multiple training samples at multiple time steps includes the following implementation: the sum of the mean and three times the standard deviation is determined as the anomaly threshold, which can be referred to in the following formula (12): (12) in, This indicates the abnormal threshold. This represents the mean. , Indicates standard deviation. .

[0149] In this embodiment, only historical sensor signals of the target device under fault-free conditions are used as training samples. Statistical calculations are performed based on the sparse abnormal distribution pattern under normal operating conditions throughout the process. This avoids interference from fault data and abnormal operating conditions in setting abnormal thresholds, ensuring that the abnormal thresholds fully match the normal fluctuation range of the device's steady-state operation. The judgment benchmark is more objective and in line with actual operating conditions, which can effectively distinguish between normal time-series fluctuations and fault-type anomalies, significantly improving the accuracy and stability of fault judgment.

[0150] The fault detection method provided in this application embodiment can perform real-time fault detection or offline detection of equipment.

[0151] For example, see Figure 6 , Figure 6 This is a flowchart illustrating model training and fault detection according to an exemplary embodiment. Three types of sensor signals, D1, D2, and D3, are used as training samples, and the signal processing model is trained iteratively for 100 rounds using the Adam optimizer. During the inference, i.e., usage phase, of the signal processing model, 3... The algorithm calculates a dynamic anomaly threshold, obtains a sparse matrix from the low-rank matrix output by the signal processing model, calculates anomaly scores based on the sparse matrix, determines whether a fault exists based on the anomaly scores, and then outputs the original anomaly data matrix, along with the corresponding low-rank and sparse matrices. Therefore, this signal processing model can detect equipment operating status and provide fault early warning.

[0152] This application deeply integrates the core logic of low-rank sparse dual matrix factorization in traditional robust principal component analysis with the local feature extraction capability of convolutional neural networks and the temporal dependency modeling capability of gated recurrent units, constructing a TS-RPCA-NN time-series signal processing model with an encoder-decoder architecture. This model accurately constrains low-rank components using the kernel norm, induces sparse components using the L1 norm, and completes fault determination using a 3σ dynamic threshold based on normal operating condition statistics. This not only compensates for the lack of temporal context modeling capability in traditional RPCA but also solves the problems of high engineering implementation costs and difficult parameter tuning for complex deep learning models. It achieves accurate extraction of normal operating characteristics and effective identification of abnormal fault signals, providing reliable technical support for industrial equipment condition monitoring and fault early warning. This results in a dual improvement in the recognition performance and engineering feasibility of anomaly detection for various time-series sensor data (sound, vibration, temperature, etc.) in industrial equipment.

[0153] This application provides a fault detection method. This method relies on a convolutional encoder, gated encoder, gated decoder, and convolutional decoder in a signal processing model to construct a complete closed-loop architecture for encoding, temporal feature restoration, and data reconstruction. It deeply integrates the low-rank-sparse dual-matrix decomposition mechanism of the traditional RPCA scheme with temporal context modeling and local spatial feature extraction capabilities. This effectively compensates for the shortcomings of the traditional RPCA scheme in characterizing the temporal evolution of signals. Through the step-by-step feature extraction and reconstruction of the signal processing model, the low-rank matrix accurately fits the temporal variation characteristics of sensor signals under normal equipment operation. Simultaneously, the sparse matrix effectively characterizes abnormal situations where sensor signals deviate from the normal evolution pattern. Therefore, equipment fault detection based on the sparse matrix output by the signal processing model can significantly improve the accuracy of equipment fault detection.

[0154] Figure 7 This is a block diagram illustrating a training apparatus for a signal processing model according to an exemplary embodiment. (Refer to...) Figure 7 The device includes: The acquisition module 701 is used to acquire multiple batches of training samples. Each batch of training samples includes multiple training samples, and each training sample includes multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected from the same device. Module 702 is used to construct the original data matrix of each batch based on the training samples of each batch; The processing module 703 is used to encode the original data matrix of each batch through the convolutional encoder of the signal processing model to obtain the encoded features, extract the latent features of the encoded features through the gated encoder of the signal processing model to obtain the first latent feature, restore the first latent feature through the gated decoder of the signal processing model to obtain the second latent feature, and decode the second latent feature through the convolutional decoder of the signal processing model to obtain the low-rank matrix of the batch. The first latent feature is used to characterize the evolution law of multiple sensor signals in the time dimension, and the low-rank matrix is ​​used to characterize the normal evolution law of multiple sensor signals in the time dimension when the equipment is running normally. The determination module 704 is used to obtain a batch of sparse matrices based on the original data matrix and the low-rank matrix. The sparse matrix is ​​used to characterize the abnormal situation of multiple sensor signals relative to the normal evolution law. Training module 705 is used to train the signal processing model based on the original data matrix, low-rank matrix, and sparse matrix for each batch.

[0155] In some embodiments, the convolutional encoder includes M convolutional layers, each convolutional layer including a convolutional unit, an activation unit, and a pooling unit; the input data of the first convolutional layer is the original data matrix, the input data of the i-th convolutional layer is the output data of the (i-1)-th convolutional layer, and the output data of the M-th convolutional layer is the encoded feature, where i is an integer greater than 1 and not greater than M; the processing module 703 is used for: For the i-th convolutional layer, the convolutional unit of the i-th convolutional layer performs local feature extraction and feature dimension transformation on the input data of the convolutional layer to obtain the first output data. The activation unit of the i-th convolutional layer performs nonlinear transformation on the first output data to obtain the second output data. The pooling unit of the i-th convolutional layer performs temporal dimension downsampling on the second output data to obtain the output data of the i-th convolutional layer.

[0156] In some embodiments, the convolutional decoder includes N transposed convolutional layers, the Nth transposed convolutional layer includes a transposed convolutional unit, and each transposed convolutional layer other than the Nth layer includes a transposed convolutional unit and an activation unit; the input data of the first transposed convolutional layer is a second latent feature, the input data of the jth transposed convolutional layer is the output data of the (j-1)th transposed convolutional layer, and the output data of the Nth transposed convolutional layer is a low-rank matrix, where j is an integer greater than 1 and not greater than N; the processing module 703 is used for: For the j-th transposed convolutional layer other than the N-th layer, the input data of the j-th transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the j-th transposed convolutional layer to obtain the third output data. The third output data is then transformed nonlinearly through the activation unit of the j-th transposed convolutional layer to obtain the fourth output data. For the Nth transposed convolutional layer, the input data of the Nth transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the Nth transposed convolutional layer to obtain a low-rank matrix.

[0157] In some embodiments, the coding features include sub-coding features for each of multiple time steps within a time period, and the first hidden feature includes sub-hidden features for each of multiple time steps; the processing module 703 is configured to: For the m-th time step, the hidden features of the m-th time step are extracted from the sub-hidden features and the sub-coded features of the m-th time step by the gate encoder, so as to obtain the sub-hidden features of the m-th time step. The gate encoder is used to reduce the number of feature dimensions of the sub-coded features to a preset number.

[0158] In some embodiments, the training module 705 is used for: For each batch, based on the original data matrix, low-rank matrix, and sparse matrix of the batch, the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch are determined. The reconstruction loss is used to characterize the difference between the matrix reconstructed based on the low-rank matrix and sparse matrix and the original data matrix. The low-rank regularization loss is used to characterize the low-rank degree and energy level of the low-rank matrix. The sparse regularization loss is used to characterize the overall energy level of the sparse matrix. The batch loss function is determined based on the batch reconstruction loss, low-rank regularization loss, and sparse regularization loss. The model parameters of the signal processing model are adjusted with the goal of minimizing the loss function.

[0159] In some embodiments, the training module 705 is used for: For each batch, the reconstructed data matrix of the batch is obtained based on the low-rank matrix and sparse matrix of the batch. Based on the reconstructed data matrix of the batch and the original data matrix, the reconstruction loss of the batch is determined.

[0160] In some embodiments, the low-rank matrix of each batch includes sub-low-rank matrices of multiple training samples in the batch, and the training module 705 is used for: For each training sample in each batch, singular value decomposition is performed on the sub-low-rank matrix of the training sample to obtain multiple singular values; The nuclear norm of the training samples is obtained based on multiple singular values; Based on the nuclear norm of each of the multiple training samples in a batch, the low-rank regularized loss of the batch is obtained.

[0161] In some embodiments, the training module 705 is used for: For each batch, determine the sum of the absolute values ​​of multiple elements in the sparse matrix of the batch; The ratio of the sum to the number of multiple elements is used to determine the sparse regularization loss of the batch.

[0162] In some embodiments, the training module 705 is used for: The batch loss function is obtained by weighted summation of the reconstruction loss, low-rank regularization loss, and sparse regularization loss.

[0163] Figure 8 This is a block diagram illustrating a fault detection device according to an exemplary embodiment. (Refer to...) Figure 8 The device includes: The acquisition module 801 is used to acquire multiple target samples of the target device. Each target sample includes signal segments of multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected for the target device. Module 802 is used to construct the target data matrix of the target device based on multiple target samples; The processing module 803 is used to encode the target data matrix through the convolutional encoder of the signal processing model to obtain encoded features, extract latent features from the encoded features through the gated encoder of the signal processing model to obtain a first latent feature, restore the first latent feature through the gated decoder of the signal processing model to obtain a second latent feature, and decode the second latent feature through the convolutional decoder of the signal processing model to obtain a target low-rank matrix. The first latent feature is used to characterize the evolution law of multiple sensor signals in the time dimension, and the target low-rank matrix is ​​used to characterize the normal evolution law of multiple sensor signals in the time dimension when the target device is operating normally. The signal processing model is a signal processing model trained through any of the above embodiments. The determination module 804 is used to obtain the target sparse matrix based on the target data matrix and the target low-rank matrix. The target sparse matrix is ​​used to characterize the abnormal situation of multiple sensor signals relative to the normal evolution law. The detection module 805 is used to perform fault detection on the target device based on the target sparse matrix.

[0164] In some embodiments, the target sparse matrix includes the abnormal signal values ​​of multiple feature dimensions of each target sample at multiple time steps; the detection module 805 is configured to: For each time step of each target sample, an anomaly score for that time step is determined based on the anomaly signal values ​​of multiple feature dimensions at that time step. Based on the anomaly score of each target sample at each time step, determine the anomaly score of the target device; If the anomaly score of the target device exceeds the anomaly threshold, the target device is determined to be faulty. The anomaly threshold is determined based on the sensor signals when the target device is not faulty.

[0165] In some embodiments, the apparatus further includes a determining module 804, configured to: Acquire multiple training samples when the target device is not malfunctioning. Each training sample includes signal segments from multiple sensor signals of the target device at the same time period. Based on the signal processing model, a sparse matrix of multiple training samples is obtained; Based on the sparse matrix, the anomaly scores of each training sample at multiple time steps are obtained; Anomaly thresholds are obtained based on the mean and standard deviation of the anomaly scores of multiple training samples at multiple time steps.

[0166] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0167] In some embodiments, the computer device is provided as a terminal. Figure 9 A schematic diagram of a terminal 900 provided in an exemplary embodiment of this disclosure is shown. The terminal 900 may be a smartphone, tablet computer, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), laptop computer, or desktop computer. The terminal 900 may also be referred to as a user device, portable terminal, laptop terminal, desktop terminal, or other names.

[0168] Typically, terminal 900 includes a processor 901 and a memory 902.

[0169] Processor 901 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 901 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 901 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 901 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 901 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0170] The memory 902 may include one or more computer-readable storage media, which may be non-transitory. The memory 902 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 902 are used to store at least one program code, which is executed by the processor 901 to implement the signal processing model training method or fault detection method provided in the method embodiments of this disclosure.

[0171] In some embodiments, the terminal 900 may also optionally include a peripheral device interface 903 and at least one peripheral device. The processor 901, memory 902, and peripheral device interface 903 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 903 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of the following: a radio frequency circuit 904, a display screen 905, a camera assembly 906, an audio circuit 907, and a power supply 908.

[0172] Peripheral device interface 903 can be used to connect at least one I / O (Input / Output) related peripheral device to processor 901 and memory 902. In some embodiments, processor 901, memory 902 and peripheral device interface 903 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 901, memory 902 and peripheral device interface 903 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0173] The radio frequency (RF) circuit 904 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 904 communicates with communication networks and other communication devices via electromagnetic signals. The RF circuit 904 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. Optionally, the RF circuit 904 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The RF circuit 904 can communicate with other terminals via at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: metropolitan area networks (MANs), various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks (WLANs), and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 904 may also include circuitry related to NFC (Near Field Communication), which is not limited in this disclosure.

[0174] Display screen 905 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 905 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 901 for processing. In this case, display screen 905 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 905, which serves as the front panel of terminal 900; in other embodiments, there may be at least two display screens 905, respectively disposed on different surfaces of terminal 900 or in a folded design; in still other embodiments, display screen 905 may be a flexible display screen, disposed on a curved or folded surface of terminal 900. Furthermore, display screen 905 may be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. Display screen 905 may be made of materials such as LCD (Liquid Crystal Display) or OLED (Organic Light-Emitting Diode).

[0175] The camera assembly 906 is used to acquire images or videos. Optionally, the camera assembly 906 includes a front-facing camera and a rear-facing camera. Typically, the front-facing camera is located on the front panel of the terminal, and the rear-facing camera is located on the back of the terminal. In some embodiments, there are at least two rear-facing cameras, which are any one of a main camera, a depth-sensing camera, a wide-angle camera, and a telephoto camera, to achieve background blurring by fusion of the main camera and the depth-sensing camera, panoramic shooting by fusion of the main camera and the wide-angle camera, VR (Virtual Reality) shooting, or other fusion shooting functions. In some embodiments, the camera assembly 906 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm-light flash and a cool-light flash, which can be used for light compensation at different color temperatures.

[0176] The audio circuit 907 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting them into electrical signals that are input to the processor 901 for processing, or to the radio frequency circuit 904 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each positioned at a different location on the terminal 900. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert electrical signals from the processor 901 or the radio frequency circuit 904 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 907 may also include a headphone jack.

[0177] Power supply 908 is used to power the various components in terminal 900. Power supply 908 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 908 includes a rechargeable battery, the rechargeable battery can support wired or wireless charging. The rechargeable battery can also be used to support fast charging technology.

[0178] Those skilled in the art will understand that Figure 9 The structure shown does not constitute a limitation on terminal 900 and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0179] Figure 10 This is a schematic diagram of a server structure according to an embodiment of this application. The server 1000 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 1001 and one or more memories 1002. The memories 1002 are used to store executable program code, and the processors 1001 are configured to execute the executable program code to implement the signal processing model training method or fault detection method provided in the various method embodiments described above. Of course, the server may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server may also include other components for implementing device functions, which will not be elaborated here.

[0180] In an exemplary embodiment, a computer-readable storage medium is also provided, which, when executed by a processor of a computer device, enables the computer device to perform the training method for the signal processing model described above. Optionally, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc.

[0181] In an exemplary embodiment, a computer program product is also provided, the computer program product including a computer program that, when executed by a processor, implements the training method or fault detection method of the above-described signal processing model.

[0182] In some embodiments, the computer program product involved in this disclosure may be deployed and executed on a computer device, or on multiple computer devices located in one location, or on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0183] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims. All the above-described optional technical solutions can be combined in any way to form optional embodiments of this application, and will not be elaborated upon here.

[0184] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A training method for a signal processing model, characterized in that, The method includes: Multiple batches of training samples are acquired, each batch of training samples includes multiple training samples, and each training sample includes multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected from the same device. Based on the training samples of each batch, construct the original data matrix of the batch; For each batch of original data matrix, the original data matrix is ​​encoded by the convolutional encoder of the signal processing model to obtain encoded features. The encoded features are then extracted by the gated encoder of the signal processing model to obtain a first latent feature. The first latent feature is then restored by the gated decoder of the signal processing model to obtain a second latent feature. The second latent feature is then decoded by the convolutional decoder of the signal processing model to obtain the low-rank matrix of the batch. The first latent feature is used to characterize the evolution of the multiple sensor signals in the time dimension, and the low-rank matrix is ​​used to characterize the normal evolution of the multiple sensor signals in the time dimension when the device is operating normally. Based on the original data matrix and the low-rank matrix, a sparse matrix of the batch is obtained. The sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law. The signal processing model is trained based on the original data matrix, low-rank matrix, and sparse matrix of each batch.

2. The method according to claim 1, characterized in that, The convolutional encoder includes M convolutional layers, each of which includes a convolutional unit, an activation unit, and a pooling unit; the input data of the first convolutional layer is the original data matrix, the input data of the i-th convolutional layer is the output data of the (i-1)-th convolutional layer, and the output data of the M-th convolutional layer is the encoded feature, where i is an integer greater than 1 and not greater than M. The process of encoding the original data matrix using a convolutional encoder based on a signal processing model to obtain encoded features includes: For the i-th convolutional layer, the input data of the convolutional layer is subjected to local feature extraction and feature dimension transformation by the convolutional unit of the i-th convolutional layer to obtain the first output data. The first output data is subjected to nonlinear transformation by the activation unit of the i-th convolutional layer to obtain the second output data. The second output data is subjected to temporal dimension downsampling by the pooling unit of the i-th convolutional layer to obtain the output data of the i-th convolutional layer.

3. The method according to claim 1 or 2, characterized in that, The convolutional decoder includes N transposed convolutional layers. The Nth transposed convolutional layer includes a transposed convolutional unit, and each transposed convolutional layer other than the Nth layer includes a transposed convolutional unit and an activation unit. The input data of the 1st transposed convolutional layer is the second latent feature, the input data of the jth transposed convolutional layer is the output data of the (j-1)th transposed convolutional layer, and the output data of the Nth transposed convolutional layer is the low-rank matrix, where j is an integer greater than 1 and not greater than N. The step of decoding the second latent feature using the convolutional decoder of the signal processing model to obtain the low-rank matrix of the batch includes: For the j-th transposed convolutional layer other than the N-th layer, the input data of the j-th transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the j-th transposed convolutional layer to obtain the third output data. The third output data is then subjected to nonlinear transformation through the activation unit of the j-th transposed convolutional layer to obtain the fourth output data. For the Nth transposed convolutional layer, the input data of the Nth transposed convolutional layer is restored in terms of feature dimension and temporal dimension through the transposed convolutional unit of the Nth transposed convolutional layer to obtain the low-rank matrix.

4. The method according to claim 1, characterized in that, The coding feature includes sub-coding features for each of multiple time steps within the time period, and the first hidden feature includes the sub-hidden features for each of the multiple time steps. The step of extracting latent features from the encoded features using the gated encoder of the signal processing model to obtain the first latent feature includes: For the m-th time step, the gated encoder extracts the latent features of the sub-latent features of the (m-1)-th time step and the sub-coded features of the m-th time step from the encoded features to obtain the sub-latent features of the m-th time step. The gated encoder is used to reduce the number of feature dimensions of the sub-coded features to a preset number.

5. The method according to claim 1, characterized in that, The training of the signal processing model based on the original data matrix, low-rank matrix, and sparse matrix of each batch includes: For each batch, based on the original data matrix, low-rank matrix, and sparse matrix of the batch, the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch are determined. The reconstruction loss is used to characterize the difference between the matrix reconstructed based on the low-rank matrix and the sparse matrix and the original data matrix. The low-rank regularization loss is used to characterize the low-rank degree and energy level of the low-rank matrix. The sparse regularization loss is used to characterize the overall energy level of the sparse matrix. Based on the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch, the loss function of the batch is determined; The model parameters of the signal processing model are adjusted with the goal of minimizing the loss function.

6. The method according to claim 5, characterized in that, For each batch, the reconstruction loss is determined based on the original data matrix, low-rank matrix, and sparse matrix of the batch, including: For each batch, a reconstructed data matrix is ​​obtained based on the low-rank matrix and sparse matrix of the batch. Based on the reconstructed data matrix and the original data matrix of the batch, the reconstruction loss of the batch is determined.

7. The method according to claim 5, characterized in that, The low-rank matrix for each batch includes sub-low-rank matrices for each of the multiple training samples in that batch. For each batch, based on the original data matrix, low-rank matrix, and sparse matrix of that batch, the low-rank regularization loss for that batch is determined, including: For each training sample in each batch, singular value decomposition is performed on the sub-low-rank matrix of the training sample to obtain multiple singular values; The nuclear norm of the training samples is obtained based on the multiple singular values; The low-rank regularized loss of the batch is obtained based on the nuclear norm of each of the multiple training samples in the batch.

8. The method according to claim 5, characterized in that, For each batch, based on the original data matrix, low-rank matrix, and sparse matrix of the batch, the sparse regularization loss of the batch is determined, including: For each batch, determine the sum of the absolute values ​​of multiple elements in the sparse matrix of that batch; The ratio of the sum to the number of the plurality of elements is determined as the sparse regularization loss of the batch.

9. The method according to claim 5, characterized in that, The step of determining the loss function for the batch based on the reconstruction loss, low-rank regularization loss, and sparse regularization loss of the batch includes: The loss function for the batch is obtained by weighted summation of the reconstruction loss, low-rank regularization loss, and sparse regularization loss.

10. A fault detection method, characterized in that, The method includes: Multiple target samples of the target device are acquired, each target sample including multiple sensor signals in the same time period, the multiple sensor signals being different types of signals collected for the target device; Based on the multiple target samples, construct the target data matrix of the target device; The target data matrix is ​​encoded using a convolutional encoder of the signal processing model to obtain encoded features. Latent features are extracted from the encoded features using a gated encoder of the signal processing model to obtain a first latent feature. The first latent feature is then restored using a gated decoder of the signal processing model to obtain a second latent feature. The second latent feature is then decoded using a convolutional decoder of the signal processing model to obtain a target low-rank matrix. The first latent feature is used to characterize the evolution of the multiple sensor signals in the temporal dimension. The target low-rank matrix is ​​used to characterize the normal evolution of the multiple sensor signals in the temporal dimension when the target device is operating normally. The signal processing model is a signal processing model trained using any one of claims 1-9. Based on the target data matrix and the target low-rank matrix, a target sparse matrix is ​​obtained. The target sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law. Based on the target sparse matrix, fault detection is performed on the target device.

11. The method according to claim 10, characterized in that, The target sparse matrix includes the anomalous signal values ​​of multiple feature dimensions of each target sample at multiple time steps; The fault detection of the target device based on the target sparse matrix includes: For each time step of each target sample, an anomaly score for that time step is determined based on the anomaly signal values ​​of the multiple feature dimensions at that time step. The anomaly score of the target device is determined based on the anomaly score of each target sample at each time step; If the anomaly score of the target device exceeds an anomaly threshold, the target device is determined to have malfunctioned. The anomaly threshold is determined based on the sensor signals when the target device is not malfunctioning.

12. The method according to claim 11, characterized in that, The method further includes: Acquire multiple training samples when the target device is not malfunctioning, each training sample including signal segments of multiple sensor signals of the target device in the same time period; Based on the signal processing model, a sparse matrix of the multiple training samples is obtained; Based on the sparse matrix, the anomaly scores of each training sample at multiple time steps are obtained; The anomaly threshold is obtained based on the mean and standard deviation of the anomaly scores of the multiple training samples at multiple time steps.

13. A training device for a signal processing model, characterized in that, The device includes: The acquisition module is used to acquire multiple batches of training samples. Each batch of training samples includes multiple training samples, and each training sample includes multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected from the same device. A construction module is used to construct the original data matrix of each batch based on the training samples of each batch; The processing module is used to encode the original data matrix of each batch using a convolutional encoder of the signal processing model to obtain encoded features, extract latent features from the encoded features using a gated encoder of the signal processing model to obtain a first latent feature, restore the first latent feature using a gated decoder of the signal processing model to obtain a second latent feature, and decode the second latent feature using a convolutional decoder of the signal processing model to obtain a low-rank matrix of the batch. The first latent feature is used to characterize the evolution law of the multiple sensor signals in the time dimension, and the low-rank matrix is ​​used to characterize the normal evolution law of the multiple sensor signals in the time dimension when the device is operating normally. The determination module is used to obtain a sparse matrix of the batch based on the original data matrix and the low-rank matrix. The sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law. The training module is used to train the signal processing model based on the original data matrix, low-rank matrix, and sparse matrix for each batch.

14. A fault detection device, characterized in that, The device includes: The acquisition module is used to acquire multiple target samples of the target device. Each target sample includes signal segments of multiple sensor signals in the same time period. The multiple sensor signals are different types of signals collected for the target device. A construction module is used to construct a target data matrix for the target device based on the multiple target samples; The processing module is configured to encode the target data matrix using a convolutional encoder of the signal processing model to obtain encoded features, extract latent features from the encoded features using a gated encoder of the signal processing model to obtain a first latent feature, restore the first latent feature using a gated decoder of the signal processing model to obtain a second latent feature, and decode the second latent feature using a convolutional decoder of the signal processing model to obtain a target low-rank matrix. The first latent feature is used to characterize the evolution of the multiple sensor signals in the temporal dimension, and the target low-rank matrix is ​​used to characterize the normal evolution of the multiple sensor signals in the temporal dimension when the target device is operating normally. The signal processing model is a signal processing model trained by any one of claims 1-9. The determination module is used to obtain a target sparse matrix based on the target data matrix and the target low-rank matrix, wherein the target sparse matrix is ​​used to characterize the abnormal situation of the multiple sensor signals relative to the normal evolution law; The detection module is used to perform fault detection on the target device based on the target sparse matrix.

15. A computer device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the training method of the signal processing model as described in any one of claims 1 to 9 or the fault detection method as described in any one of claims 10 to 12.

16. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of a computer device, the computer device is able to perform the training method of the signal processing model according to any one of claims 1 to 9 or the fault detection method according to any one of claims 10 to 12.

17. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the training method of the signal processing model according to any one of claims 1 to 9 or the fault detection method according to any one of claims 10 to 12.

Images