System fault detection method, device, mobile terminal and storage medium

Abstract

The present invention discloses a system fault detection method, device, mobile terminal, and storage medium. The method comprises: obtaining power fault data of a system to be detected; inputting the power fault data into a stacked sparse autoencoder, so that the stacked sparse autoencoder extracts features based on the power fault data and outputs reduced-dimensional feature data; inputting the feature data into a clustering model, so that the clustering model classifies the feature data and outputs the corresponding fault type, thereby completing fault detection in the detection system. The embodiments of the present invention improve detection efficiency.

Description

Technical Field

[0001] The present invention relates to the field of fault detection technology, and in particular to a system fault detection method, device, mobile terminal and storage medium. Background Art

[0002] Existing fault detection methods can be categorized into two main categories: qualitative analysis and quantitative analysis. Qualitative analysis methods primarily rely on the work experience of metering operators and use subjective judgment to analyze the nature, characteristics, and development patterns of power metering automation system data. Fault detection is performed by establishing an empirical knowledge base and designing a set of algorithms or logical reasoning diagrams to simulate the reasoning and decision-making processes of human experts. For fault detection in power metering automation systems, metering operators and maintenance personnel often use empirical detection methods, establishing detection algorithm processes based on the empirical knowledge base. However, the accuracy of these methods is subject to subjective influences and relies on the operator's data analysis capabilities.

[0003] Furthermore, data-driven detection methods are increasingly being applied to fault detection in power metering automation systems. These methods utilize data mining techniques to identify data features, establish mathematical models, and implement fault detection. This method requires a large amount of labeled data to build the model, and data quality can affect the accuracy of detection results. Data collected from industrial sites is unlabeled, making analyzing and labeling large amounts of unlabeled industrial data a labor-intensive task for operations and maintenance personnel. This is not only slow, costly, and inefficient, but also subject to subjective factors.

[0004] As can be seen from the above, existing system fault detection methods rely on manual intervention and have the problem of low detection efficiency. Summary of the Invention

[0005] The embodiments of the present invention provide a system fault detection method, device, mobile terminal and storage medium, which improve the detection efficiency.

[0006] A first aspect of an embodiment of the present application provides a system fault detection method, including:

[0007] Obtain power fault data of the system to be detected;

[0008] Inputting the power fault data into a stacked sparse autoencoder, so that the stacked sparse autoencoder extracts features based on the power fault data and outputs feature data after dimensionality reduction;

[0009] The characteristic data is input into the clustering model so that the clustering model can classify according to the characteristic data, output the corresponding fault type, and complete the fault detection of the detection system.

[0010] In a possible implementation of the first aspect, the generation process of the stacked sparse autoencoder is specifically as follows:

[0011] Adding sparse constraints to the loss function of the autoencoder to generate the first loss function;

[0012] Iteratively updating the network parameters according to the back-propagation algorithm to generate a sparse autoencoder after the first loss function converges;

[0013] Multiple sparse autoencoders are stacked to generate a stacked sparse autoencoder.

[0014] In a possible implementation of the first aspect, the clustering model training process is specifically as follows:

[0015] Obtain the initial cluster center according to the density of sample feature data;

[0016] Initialize the membership matrix; the membership matrix represents the degree to which a certain sample feature data belongs to different fault types;

[0017] By iterating the membership matrix and the initial cluster center until the iteration termination condition is met, the iteration is stopped and the clustering model is output.

[0018] In a possible implementation of the first aspect, satisfying an iteration termination condition is specifically:

[0019] The membership matrix is ​​calculated iteratively. When the absolute value of the difference between two calculation results obtained by two consecutive iterative calculations is less than a preset termination threshold, it is determined that the iteration termination condition is met.

[0020] In a possible implementation of the first aspect, obtaining power fault data of the system to be detected is specifically:

[0021] Acquire initial fault data of the system to be detected, filter data that does not meet quality requirements in the initial fault data, and generate first data;

[0022] Normalization is performed on the first data to generate and obtain power fault data of the system to be detected.

[0023] A second aspect of an embodiment of the present application provides a system fault detection device, comprising: an acquisition module, an extraction module, and a detection module;

[0024] The acquisition module is used to obtain the power fault data of the system to be detected;

[0025] The extraction module is used to input the power fault data into the stacked sparse autoencoder, so that the stacked sparse autoencoder performs feature extraction based on the power fault data and outputs feature data after dimensionality reduction;

[0026] The detection module is used to input feature data into the clustering model so that the clustering model can classify according to the feature data, output the corresponding fault type, and complete the fault detection of the detection system.

[0027] In a possible implementation of the second aspect, the generation process of the stacked sparse autoencoder is specifically as follows:

[0028] Adding sparse constraints to the loss function of the autoencoder to generate the first loss function;

[0029] Iteratively updating the network parameters according to the back-propagation algorithm to generate a sparse autoencoder after the first loss function converges;

[0030] Multiple sparse autoencoders are stacked to generate a stacked sparse autoencoder.

[0031] In a possible implementation of the second aspect, the clustering model training process is specifically as follows:

[0032] Obtain the initial cluster center according to the density of sample feature data;

[0033] Initialize the membership matrix; the membership matrix represents the degree to which a certain sample feature data belongs to different fault types;

[0034] By iterating the membership matrix and the initial cluster center until the iteration termination condition is met, the iteration is stopped and the clustering model is output.

[0035] A third aspect of an embodiment of the present application provides a mobile terminal, including a processor and a memory, wherein the memory stores computer-readable program code, and when the processor executes the computer-readable program code, the steps of the above-mentioned system fault detection method are implemented.

[0036] A fourth aspect of an embodiment of the present application provides a storage medium, which stores computer-readable program code. When the computer-readable program code is executed, the steps of the above-mentioned system fault detection method are implemented.

[0037] Compared with the prior art, the embodiments of the present invention provide a system fault detection method, device, mobile terminal and storage medium, the method comprising: obtaining power fault data of the system to be detected; inputting the power fault data into a stacked sparse autoencoder so that the stacked sparse autoencoder performs feature extraction based on the power fault data and outputs feature data after dimensionality reduction; inputting the feature data into a clustering model so that the clustering model performs classification based on the feature data and outputs the corresponding fault type, thereby completing fault detection of the detection system.

[0038] Its beneficial effects are as follows: the embodiment of the present invention inputs power fault data into a stacked sparse autoencoder to obtain feature data after dimensionality reduction; the feature data is input into a clustering model to obtain the corresponding fault type, thereby completing the fault detection of the detection system. The embodiment of the present invention combines the stacked sparse autoencoder with the clustering model, uses the stacked sparse autoencoder to extract highly abstract features layer by layer to obtain an effective dimensionality reduction feature representation, can directly extract features from unlabeled data, and then input the feature data into the clustering model, which can realize classification detection of a large amount of unlabeled data, avoiding the problem of low detection efficiency caused by the need to manually label a large amount of data before classification detection in the prior art, and greatly improving the efficiency of detection.

[0039] Furthermore, the embodiment of the present invention proposes a two-stage unsupervised fault clustering algorithm, namely a deep adaptive fuzzy clustering algorithm, for the problem of fault detection in the electric power metering automation system. It combines the stacked sparse autoencoder with the adaptive Gath-Geva clustering (a fuzzy clustering instance) to form an unsupervised fault detection framework for clustering analysis of unlabeled big data. The stacked sparse autoencoder is used to extract highly abstract features of the original data and fine-tune the network using different unsupervised strategies; the adaptive Gath-Geva clustering can adaptively obtain the optimal clustering results without pre-setting the number of clusters. With the further improvement of detection accuracy, it can be applied to the electric power metering automation system to realize all-round automated real-time monitoring and achieve the purpose of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of a system fault detection method provided by one embodiment of the present invention;

[0041] Figure 2 1 is a schematic diagram of the structure of an automatic encoder provided by an embodiment of the present invention;

[0042] Figure 3 is a schematic diagram of the structure of a stacked sparse autoencoder provided by one embodiment of the present invention;

[0043] Figure 4 1 is a flow chart of an SAE training method provided by one embodiment of the present invention;

[0044] Figure 5 1 is a flow chart of an SSAE fine-tuning optimization method provided by one embodiment of the present invention;

[0045] Figure 6 1 is a schematic diagram of the training process of the two-stage model provided by one embodiment of the present invention;

[0046] Figure 71 is a flow chart of an adaptive Gath-Geva clustering model training method provided by one embodiment of the present invention;

[0047] Figure 8 It is a structural diagram of a system fault detection device provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0049] Reference Figure 1 , is a flow chart of a system fault detection method provided by an embodiment of the present invention, including S101-S103:

[0050] S101: Acquire power fault data of a system to be detected.

[0051] Wherein, the system to be detected is an electric power metering automation system to be detected.

[0052] In this embodiment, the acquisition of power fault data of the system to be detected is specifically as follows:

[0053] Acquiring initial fault data of the system to be detected, filtering data that does not meet quality conditions in the initial fault data, and generating first data;

[0054] Normalization is performed on the first data to generate and acquire power fault data of the system to be detected.

[0055] Furthermore, data that does not meet quality requirements includes: data with missing features, data with abnormal collection, and duplicate data. Normalizing the data can enhance the robustness of the trained model.

[0056] S102: Inputting the power fault data into a stacked sparse autoencoder, so that the stacked sparse autoencoder performs feature extraction based on the power fault data and outputs feature data after dimensionality reduction.

[0057] Among them, the stacked sparse autoencoder is used to extract highly abstract features layer by layer to obtain an effective dimensionality reduction feature representation, which can directly extract features from unlabeled data and obtain the approximate distribution of data features. It does not require prior knowledge of data processing technology and professional knowledge related to fault detection. Therefore, it does not require human intervention, does not rely on the experience knowledge base, and eliminates the work of manual feature selection and recording labels, and is not affected by subjective factors.

[0058] In this embodiment, the generation process of the stacked sparse autoencoder is specifically as follows:

[0059] Adding sparse constraints to the loss function of the autoencoder to generate the first loss function;

[0060] Iteratively updating network parameters according to a back-propagation algorithm to generate a sparse autoencoder after the first loss function converges;

[0061] The stacked sparse autoencoder is generated by stacking a plurality of the sparse autoencoders.

[0062] S103: Input the characteristic data into the clustering model so that the clustering model performs classification according to the characteristic data and outputs the corresponding fault type, thereby completing the fault detection of the detection system.

[0063] In this embodiment, the training process of the clustering model is specifically as follows:

[0064] Obtain the initial cluster center according to the density of sample feature data;

[0065] Initializing a membership matrix; wherein the membership matrix represents the degree to which a certain sample feature data belongs to different fault types;

[0066] The membership matrix and the initial cluster center are iterated until an iteration termination condition is met, and then the iteration is stopped and the cluster model is output.

[0067] In a specific embodiment, the iteration termination condition is satisfied as follows:

[0068] The membership matrix is ​​iteratively calculated, and when the absolute value of the difference between two calculation results obtained by two consecutive iterative calculations is less than a preset termination threshold, it is determined that the iteration termination condition is met.

[0069] To further illustrate the generation process of stacked sparse autoencoders, please refer to Figure 2-Figure 3 . Figure 2 is a schematic structural diagram of an automatic encoder provided by an embodiment of the present invention, Figure 3 2 is a schematic diagram of the structure of a stacked sparse autoencoder provided by one embodiment of the present invention.

[0070] An autoencoder (AE) is a neural network with the same input and output, such as Figure 2 As shown, the autoencoder structure consists of two parts: encoder and decoder. The original data X∈R nxd There are n samples, each sample contains d features, which are input to the encoding layer with N neurons. Z∈R is calculated by formula (1) Nxn :

[0071] Z=f(W1X+b1); (1)

[0072] Softplus(x)=log(I+e x ); (2)

[0073] Among them, W1∈R dxN is the weight matrix, f(·) is the nonlinear activation function (as shown in formula (2)), b1∈R Nx1 is the bias vector of the encoding layer. Taking Z as the input of the decoding layer, the reconstruction of the original data is obtained, as shown in the following formula:

[0074] Y=g(W2Z+b2); (3)

[0075]

[0076] Among them, g(·) is the sigmoid loss function, which is used to activate the output data of the output layer. In order to reduce the number of parameters in model training, let b2 is the bias vector of the decoding layer. Through training, the weight matrix and bias vector are adjusted to minimize the loss function:

[0077]

[0078] The first term in J(W,b) is the total squared error between the input and output data, the second term is the weight decay term used to prevent overfitting of the AE, and λ is the decay weight. After training is completed, the output of the hidden layer is a valid dimensionality-reduced feature representation of the input data.

[0079] The Sparse AutoEncoder (SAE) is obtained by adding some sparse constraints to the AE. In the embodiment of the present invention, an additional penalty term is added to the loss function (5): Kullback-Leibler divergence (6) to achieve sparsity. The Kullback-Leibler divergence is as follows:

[0080]

[0081]

[0082] Among them, ρ is the sparsity parameter, and the average activation of the jth neuron in the hidden layer is As defined in formula (7), a j (x (i) ) represents the jth neuron in the hidden layer under the given input x (i) At this time, the first loss function formed by using KL divergence to improve the loss function (5) is:

[0083]

[0084] Among them, β is the weight of the sparsity penalty term.

[0085] The back propagation algorithm is used to update the iterative network parameters with the first loss function (8) until it converges, thereby realizing the unsupervised feature learning of SAE. Please refer to the training process Figure 4 , Figure 4 FIG. 1 is a flow chart of the SAE training method provided by one embodiment of the present invention. Figure 4 As shown, the output Y and input X of SAE are input into the loss function J sparse (W, b), so that J sparse (W, b) Update the parameters and input them into SAE.

[0086] After all SAEs are trained, they are stacked to form a deep neural feature extractor (Stacked Sparse AutoEncoder, SSAE), which is then fine-tuned using training samples. In this case, the present invention uses (9) as the new loss function:

[0087]

[0088] Among them, KL divergence is calculated by the last SAE. Please refer to the fine-tuning process Figure 5 , Figure 5 FIG. 1 is a flow chart of the SSAE fine-tuning optimization method provided by one embodiment of the present invention. Figure 5 As shown, the output Y and input X of SSAE are input to the loss function J SSAE (W, b), so that J SSAE (W, b) Update the parameters and input them into SSAE.

[0089] The features extracted by SSAE are used as the input of the second stage adaptive Gath-Geva clustering (Adaptive GG, AGG) of the model for training. Please refer to the training process. Figure 6 , Figure 6 1 is a schematic diagram of the training process of the two-stage model provided by one embodiment of the present invention.

[0090] First, the initial cluster center is selected based on the data density, and the distance between all samples is calculated using formula (10), and the average value is taken:

[0091] d(x i ,x j )=||x i -x j ||; (10)

[0092]

[0093] Where n is the number of samples, and x represents the samples in the data set and all have the same number of features. The density of sample x is defined as follows:

[0094]

[0095]

[0096] Among them, x i are other samples in the dataset. Find the sample with the largest density as the first initial cluster center, remove all samples whose distance to it is less than the mean value and repeat this process until a sufficient number (the number of clusters K in the dataset) of initial cluster centers V are obtained.

[0097] Initialize the membership matrix U = (u ik ) cxn , where u ik represents the membership of the kth sample to the ith category (i.e., the degree to which a certain sample characteristic data belongs to different fault types). By iteratively optimizing and adjusting the membership matrix U and the initial cluster center V, the minimization objective function (14) is expressed by the following formula:

[0098]

[0099] Among them, m>1 is a weighted index, generally let m=2, D is u ik The corresponding fuzzy maximum likelihood estimation distance is expressed by the following formula:

[0100]

[0101] in, is the i-th cluster center of the l-th iteration, and its iteration formula is shown in (16); F i is the covariance matrix of the ith cluster, and its iterative formula is shown in (17); p i is the prior probability of the i-th cluster, and its iterative formula is shown in (18):

[0102]

[0103]

[0104]

[0105] Membership matrix U=(u ik ) cxn Updated from (19):

[0106]

[0107] K is the number of clusters in the data set. In the actual training process of the model, the present invention adopts (20) as the iteration termination condition:

[0108] |U l -U l-1 |<0.01; (20)

[0109] The membership matrix U and cluster center V are iteratively adjusted until the membership matrix U satisfies (20) (i.e., when the absolute value of the difference between the two calculation results obtained by two consecutive iterative calculations of the membership matrix is ​​less than the preset termination threshold, the iteration termination condition is determined to be satisfied).

[0110] In order to adaptively obtain the optimal number of clusters, the present invention needs to traverse K during model training. The overall algorithm process is as follows: Figure 7 As shown, Figure 7 FIG. 1 is a flow chart of an adaptive Gath-Geva clustering model training method provided by an embodiment of the present invention. Figure 7 As shown, the K range is set to n is the number of samples in the data set, and the clustering evaluation index PBMF commonly used in fuzzy clustering is used as the evaluation criterion:

[0111]

[0112] in:

[0113]

[0114] E1=J m | K＝1 ; (twenty three)

[0115]

[0116] Among them, K is the number of clusters, and similarly, m=2, z k is the kth cluster center.

[0117] In the second stage of model training, K is set from 2 to Select K initial cluster centers, use (16) and (19) to iteratively update the membership matrix U and cluster centers V until (20) is satisfied, calculate the PBMF corresponding to the number of clusters at this time, and after all training is completed, record K when the PBMF reaches the maximum, and adaptively obtain the optimal number of clusters and clustering results without presetting the number of clusters.

[0118] The trained two-stage model is used to detect faults on input data. First, the data is normalized to the range [0, 1]. This preprocessed data serves as the input for the SSAE. After several encoder passes, highly abstract features are extracted layer by layer to obtain an effective reduced-dimensional feature representation. The second stage, AGG clustering, is used to classify these reduced-dimensional features and determine the fault category most similar to the feature, completing the fault detection process.

[0119] For further explanation of the system fault detection device, please refer to Figure 8 , Figure 8 80 is a schematic structural diagram of a system fault detection device provided by an embodiment of the present invention, comprising: an acquisition module 801, an extraction module 802 and a detection module 803;

[0120] The acquisition module 801 is used to acquire power fault data of the system to be detected;

[0121] The extraction module 802 is configured to input the power fault data into a stacked sparse autoencoder, so that the stacked sparse autoencoder performs feature extraction based on the power fault data and outputs feature data after dimensionality reduction;

[0122] The detection module 803 is used to input the feature data into a clustering model, so that the clustering model performs classification according to the feature data, outputs corresponding fault types, and completes the fault detection of the detection system.

[0123] In this embodiment, the generation process of the stacked sparse autoencoder is specifically as follows:

[0124] Adding sparse constraints to the loss function of the autoencoder to generate the first loss function;

[0125] Iteratively updating network parameters according to a back-propagation algorithm to generate a sparse autoencoder after the first loss function converges;

[0126] The stacked sparse autoencoder is generated by stacking a plurality of the sparse autoencoders.

[0127] In this embodiment, the training process of the clustering model is specifically as follows:

[0128] Obtain the initial cluster center according to the density of sample feature data;

[0129] Initializing a membership matrix; wherein the membership matrix represents the degree to which a certain sample feature data belongs to different fault types;

[0130] The membership matrix and the initial cluster center are iterated until an iteration termination condition is met, and then the iteration is stopped and the cluster model is output.

[0131] A specific embodiment of the present invention provides a mobile terminal, including a processor and a memory, wherein the memory stores computer-readable program code, and the processor implements the steps of the above-mentioned system fault detection method when executing the computer-readable program code.

[0132] A specific embodiment of the present invention provides a storage medium, wherein the storage medium stores computer-readable program code, and when the computer-readable program code is executed, the steps of the above-mentioned system fault detection method are implemented.

[0133] In the embodiment of the present invention, the acquisition module 801 acquires the power fault data of the system to be detected; the extraction module 802 inputs the power fault data into the stacked sparse autoencoder, so that the stacked sparse autoencoder performs feature extraction based on the power fault data and outputs feature data after dimensionality reduction; the detection module 803 inputs the feature data into the clustering model, so that the clustering model performs classification based on the feature data and outputs the corresponding fault type, thereby completing the fault detection of the detection system.

[0134] In an embodiment of the present invention, power fault data is input into a stacked sparse autoencoder to obtain feature data after dimensionality reduction; the feature data is input into a clustering model to obtain the corresponding fault type, thereby completing the fault detection of the detection system. In an embodiment of the present invention, a stacked sparse autoencoder is combined with a clustering model, and the stacked sparse autoencoder is used to extract highly abstract features layer by layer to obtain an effective dimensionality reduction feature representation. Features can be directly extracted from unlabeled data, and then the feature data is input into the clustering model. This enables classification detection of large amounts of unlabeled data, avoiding the problem of low detection efficiency caused by the need to manually label large amounts of data before classification detection in the prior art, thereby greatly improving detection efficiency.

[0135] Furthermore, the embodiment of the present invention proposes a two-stage unsupervised fault clustering algorithm, namely a deep adaptive fuzzy clustering algorithm, for the problem of fault detection in the electric power metering automation system. It combines the stacked sparse autoencoder with the adaptive Gath-Geva clustering (a fuzzy clustering instance) to form an unsupervised fault detection framework for clustering analysis of unlabeled big data. The stacked sparse autoencoder is used to extract highly abstract features of the original data and fine-tune the network using different unsupervised strategies; the adaptive Gath-Geva clustering can adaptively obtain the optimal clustering results without pre-setting the number of clusters. With the further improvement of detection accuracy, it can be applied to the electric power metering automation system to realize all-round automated real-time monitoring and achieve the purpose of fault detection.

[0136] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A system fault detection method, characterized in that: include: Obtain power fault data of the system to be detected; Inputting the power fault data into a stacked sparse autoencoder, so that the stacked sparse autoencoder performs feature extraction based on the power fault data and outputs feature data after dimensionality reduction; The characteristic data is input into a clustering model so that the clustering model classifies the characteristic data and outputs corresponding fault types, thereby completing fault detection of the detection system. The clustering model is an adaptive Gath-Geva clustering model. The clustering model determines the number of fault categories by taking the number of cluster centers that reaches the maximum value of the PBMF index within a preset number of cluster centers as the number of fault categories. The training process of the clustering model is specifically as follows: Obtaining initial cluster centers based on the density of sample feature data; wherein the sample feature data is extracted by a trained sparse autoencoder; Initializing a membership matrix; wherein the membership matrix represents the degree to which a certain sample feature data belongs to different fault types; The membership matrix and the initial cluster center are iterated until an iteration termination condition is met, and then the iteration is stopped and the cluster model is output.

2. A system fault detection method according to claim 1, characterized in that: The generation process of the stacked sparse autoencoder is specifically as follows: Adding sparse constraints to the loss function of the autoencoder to generate the first loss function; Iteratively updating network parameters according to a back-propagation algorithm to generate a sparse autoencoder after the first loss function converges; The stacked sparse autoencoder is generated by stacking a plurality of the sparse autoencoders.

3. A system fault detection method according to claim 2, characterized in that: The iteration termination condition is specifically satisfied as follows: The membership matrix is ​​iteratively calculated, and when the absolute value of the difference between two calculation results obtained by two consecutive iterative calculations is less than a preset termination threshold, it is determined that the iteration termination condition is met.

4. A system fault detection method according to claim 1, characterized in that: The obtaining of power fault data of the system to be detected is specifically as follows: Acquiring initial fault data of the system to be detected, filtering data that does not meet quality conditions in the initial fault data, and generating first data; Normalization is performed on the first data to generate and acquire power fault data of the system to be detected.

5. A system fault detection device, characterized in that: include: Acquisition module, extraction module and detection module; Wherein, the acquisition module is used to acquire power fault data of the system to be detected; The extraction module is used to input the power fault data into a stacked sparse autoencoder, so that the stacked sparse autoencoder performs feature extraction based on the power fault data and outputs feature data after dimensionality reduction; The detection module is configured to input the feature data into a clustering model so that the clustering model classifies the feature data and outputs the corresponding fault type, thereby completing fault detection of the detection system. The clustering model is an adaptive Gath-Geva clustering model. The clustering model determines the number of fault categories by taking the number of cluster centers where the PBMF index reaches a maximum value within a preset range of cluster centers as the number of fault categories. The training process of the clustering model is specifically as follows: Obtaining initial cluster centers based on the density of sample feature data; wherein the sample feature data is extracted by a trained sparse autoencoder; Initializing a membership matrix; wherein the membership matrix represents the degree to which a certain sample feature data belongs to different fault types; The membership matrix and the initial cluster center are iterated until an iteration termination condition is met, and then the iteration is stopped and the cluster model is output.

6. A system fault detection device according to claim 5, characterized in that: The generation process of the stacked sparse autoencoder is specifically as follows: Adding sparse constraints to the loss function of the autoencoder to generate the first loss function; Iteratively updating network parameters according to a back-propagation algorithm to generate a sparse autoencoder after the first loss function converges; The stacked sparse autoencoder is generated by stacking a plurality of the sparse autoencoders.

7. A mobile terminal, characterized in that: The system comprises a processor and a memory, wherein the memory stores a computer-readable program code, and when the processor executes the computer-readable program code, the steps of a system fault detection method according to any one of claims 1 to 4 are implemented.

8. A storage medium, characterized in that: The storage medium stores computer-readable program codes, and when the computer-readable program codes are executed, the steps of a system fault detection method according to any one of claims 1 to 4 are implemented.

Images