Fault detection model training method and device, computer device, and storage medium

Abstract

The application relates to a fault detection model training method and device, computer equipment, a storage medium and a computer program product. The method comprises: performing time sequence correlation feature processing based on a plurality of running state index data of a plurality of time points corresponding to first sample data of a normal category, to obtain a target time sequence correlation feature map; performing encoding and decoding processing on the target time sequence correlation feature map to obtain a target reconstruction time sequence correlation feature map corresponding to the target time sequence correlation feature map; and training a fault detection model to be trained based on the target time sequence correlation feature map and the target reconstruction time sequence correlation feature map. By using the method, model training and fault detection are performed in the form of a self-encoder, the similarity between original features and reconstruction features is used to detect faults, the accuracy of model training can be improved in the case of non-ideal data distribution, and the workload of model training is reduced.

Description

Technical Field

[0001] This application relates to the field of neural network technology, and in particular to a training method, apparatus, computer device, storage medium, and computer program product for a fault detection model. Background Technology

[0002] Air conditioning and other cooling equipment are crucial infrastructure for ensuring the efficient and reasonable operation of data center servers within normal temperature ranges. However, during operation, air conditioning systems inevitably experience various malfunctions due to improper operation or equipment aging, leading to problems such as cooling failures in data centers.

[0003] The training of fault diagnosis models in related technologies is carried out under ideal data distribution conditions. Since faults are occasional events, the ratio of fault data to normal data is relatively unbalanced. Due to the imbalance of training data, the prediction accuracy of the trained model is poor. Summary of the Invention

[0004] Therefore, it is necessary to provide a training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for a fault detection model that can improve the accuracy of fault detection, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for training a fault detection model. The method includes:

[0006] Acquire time-series data of the target device and determine the time-series features corresponding to the time-series data. The time-series data includes at least a first sample data of the normal category, and the first sample data includes multiple operating status index data at multiple time points.

[0007] Based on the preset time window and the time-series features corresponding to the multiple operating status index data at the multiple time points, time-series related feature processing is performed to obtain the target time-series related feature map;

[0008] The target temporal correlation feature map is encoded and decoded to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map;

[0009] Based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, the fault detection model to be trained is trained to obtain the trained target fault detection model.

[0010] In one embodiment, the time-series data further includes second sample data of the fault category;

[0011] The step of training the fault detection model to be trained based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map to obtain a trained target fault detection model includes:

[0012] Based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, the fault detection model to be trained is trained to obtain the trained initial fault detection model.

[0013] Based on a preset analysis algorithm, the second sample data of the fault category is subjected to dimensionality reduction processing to obtain target fault sample data, and based on a preset time-domain feature algorithm, the target time-domain feature set corresponding to the target fault sample data is determined.

[0014] Based on the target time-domain feature set and the initial fault detection model, the target fault detection model is determined.

[0015] In one embodiment, time-series correlation feature processing is performed on the time-series features corresponding to multiple operating status indicator data at multiple time points based on a preset time window to obtain a target time-series correlation feature map, including:

[0016] Based on a preset time window, extract from the time-series features corresponding to multiple operational status index data at multiple time points to obtain multiple target time-series features, and calculate the correlation features between each target time-series feature;

[0017] Based on the correlation characteristics among the target time-series features, a target time-series correlation feature map is determined.

[0018] In one embodiment, the step of encoding and decoding the target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map includes:

[0019] The temporal correlation feature map is convolved to obtain a convolved temporal correlation feature map, which includes a target temporal correlation feature map with multiple dimensions.

[0020] Based on a convolutional long short-term memory network, the target temporal correlation feature maps of the multiple dimensions are processed respectively to obtain multiple temporal pattern information;

[0021] Based on the target time-series related feature maps of the multiple dimensions and the multiple time pattern information, feature reconstruction processing is performed to obtain the target reconstructed time-series related feature map.

[0022] In one embodiment, the multiple dimensions of temporal correlation features include at least a first-dimensional target temporal correlation feature map, a second-dimensional target temporal correlation feature map, and a third-dimensional target temporal correlation feature map; the convolutional processing of the temporal correlation feature maps to obtain convolutionally processed temporal correlation feature maps includes:

[0023] The temporal correlation feature map is convolved by the first convolutional layer to obtain the target temporal correlation feature map of the first dimension.

[0024] The second convolutional layer is used to convolve the target temporal correlation feature map of the first dimension to obtain the target temporal correlation feature map of the second dimension.

[0025] The target temporal correlation feature map of the second dimension is processed by convolutional layer to obtain the target temporal correlation feature map of the third dimension.

[0026] In one embodiment, the convolutional long short-term memory network processes the target temporal-related feature maps of the multiple dimensions to obtain multiple temporal pattern information, including:

[0027] Temporal analysis is performed on the target temporal correlation feature map of the first dimension based on the first convolutional long short-term memory sub-network to obtain the first time pattern information;

[0028] Temporal analysis is performed on the target temporal correlation feature map of the second dimension based on the second convolutional long short-term memory sub-network to obtain the second temporal pattern information;

[0029] Temporal analysis is performed on the target temporal correlation feature map of the third dimension based on the third convolutional long short-term memory sub-network to obtain the third temporal pattern information.

[0030] In one embodiment, the feature reconstruction process based on the target time-series correlated feature map of the multiple dimensions and the multiple time pattern information to obtain the reconstructed time-series correlated feature map includes:

[0031] The first deconvolutional layer is used to perform feature reconstruction processing on the superposition result of the first temporal pattern information and the first dimension target temporal correlation feature map to obtain the first reconstructed temporal correlation feature map corresponding to the first dimension target temporal correlation feature map.

[0032] The second deconvolutional layer performs feature reconstruction processing on the superposition result of the first reconstructed temporal correlation feature map, the second temporal pattern information and the second dimension target temporal correlation feature map to obtain the second reconstructed temporal correlation feature map corresponding to the second dimension target temporal correlation feature map.

[0033] The third deconvolutional layer performs feature reconstruction processing on the superposition result of the second reconstructed temporal correlation feature map, the third temporal pattern information, and the third-dimensional target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map.

[0034] In one embodiment, the method further includes:

[0035] Acquire the timing data to be detected from the target device, and determine the timing features to be detected corresponding to the timing data to be detected;

[0036] Based on the time series features to be detected, a time series correlation feature map to be detected is determined, and the time series correlation feature map to be detected is input into the target fault detection model to obtain the reconstructed time series correlation feature map to be detected;

[0037] Based on the time-series correlation feature map to be detected and the reconstructed time-series correlation feature map to be detected, the fault detection result corresponding to the time-series data to be detected of the target device is determined.

[0038] In one embodiment, determining the fault detection result corresponding to the time series data to be detected of the target device based on the time series correlation feature map to be detected and the reconstructed time series correlation feature map to be detected includes:

[0039] If the timing-related feature map to be detected is inconsistent with the reconstructed timing-related feature map to be detected, the fault detection result corresponding to the timing data to be detected of the target device is determined to be a fault result.

[0040] The temporal correlation feature map to be detected and the reconstructed temporal correlation feature to be detected Figure One In cases where the fault detection result corresponding to the timing data to be tested of the target device is determined to be a non-fault result.

[0041] In one embodiment, determining the fault detection result corresponding to the timing data to be detected of the target device as a fault result when the timing correlation feature map to be detected is inconsistent with the reconstructed timing correlation feature map to be detected includes:

[0042] If the time-series correlation feature map to be detected is inconsistent with the reconstructed time-series correlation feature map to be detected, the time-series data to be detected of the target device is subjected to dimensionality reduction processing to obtain the target time-series data to be detected.

[0043] Based on a preset time-domain feature algorithm, the time-domain features to be detected corresponding to the target time-series data to be detected are determined;

[0044] If the time-domain feature to be detected and the target time-domain feature set satisfy the second similarity condition, then the fault detection result corresponding to the time-series data to be detected of the target device is determined to be a fault result.

[0045] In one embodiment, the step of determining the fault detection result corresponding to the time-series data to be detected of the target device as a fault result if the time-domain feature to be detected and the target time-domain feature set satisfy a second similarity condition includes:

[0046] Calculate the similarity between the time-domain feature to be detected and each target time-domain feature contained in the target time-domain feature set;

[0047] If there is a target time-domain feature with a similarity greater than a preset similarity threshold, then the fault detection result corresponding to the time-series data to be detected of the target device is determined to be a fault result.

[0048] In one embodiment, the method further includes:

[0049] Among the similarities greater than a preset similarity threshold, the largest similarity is determined, and the target temporal feature corresponding to the largest similarity is determined.

[0050] Based on the preset correspondence between time-domain features and fault types, the target fault type corresponding to the target time-domain feature is determined, and the fault type corresponding to the time-series data to be detected of the target device is determined as the target fault type.

[0051] Secondly, this application also provides a training device for a fault detection model. The device includes:

[0052] The first acquisition module is used to acquire time-series data of the target device and determine the time-series features corresponding to the time-series data. The time-series data includes at least first sample data of the normal category, and the first sample data includes multiple operating status indicator data at multiple time points.

[0053] The first determining module is used to perform time-series related feature processing on the time-series features corresponding to multiple operating status index data at multiple time points based on a preset time window, and to obtain a target time-series related feature map.

[0054] The encoding / decoding module is used to encode and decode the target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map.

[0055] The training module is used to train the fault detection model to be trained based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, so as to obtain the trained target fault detection model.

[0056] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0057] Acquire time-series data of the target device and determine the time-series features corresponding to the time-series data. The time-series data includes at least a first sample data of the normal category, and the first sample data includes multiple operating status index data at multiple time points.

[0058] Based on the preset time window and the time-series features corresponding to the multiple operating status index data at the multiple time points, time-series related feature processing is performed to obtain the target time-series related feature map;

[0059] The target temporal correlation feature map is encoded and decoded to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map;

[0060] Based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, the fault detection model to be trained is trained to obtain the trained target fault detection model.

[0061] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0062] Acquire time-series data of the target device and determine the time-series features corresponding to the time-series data. The time-series data includes at least a first sample data of the normal category, and the first sample data includes multiple operating status index data at multiple time points.

[0063] Based on the preset time window and the time-series features corresponding to the multiple operating status index data at the multiple time points, time-series related feature processing is performed to obtain the target time-series related feature map;

[0064] The target temporal correlation feature map is encoded and decoded to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map;

[0065] Based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, the fault detection model to be trained is trained to obtain the trained target fault detection model.

[0066] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0067] Acquire time-series data of the target device and determine the time-series features corresponding to the time-series data. The time-series data includes at least a first sample data of the normal category, and the first sample data includes multiple operating status index data at multiple time points.

[0068] Based on the preset time window and the time-series features corresponding to the multiple operating status index data at the multiple time points, time-series related feature processing is performed to obtain the target time-series related feature map;

[0069] The target temporal correlation feature map is encoded and decoded to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map;

[0070] Based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, the fault detection model to be trained is trained to obtain the trained target fault detection model.

[0071] The aforementioned training method, apparatus, computer equipment, storage medium, and computer program product for the fault detection model include the following steps: First, based on multiple operational status indicator data at multiple time points corresponding to first sample data containing the normal category, perform time-series correlation feature processing to obtain a target time-series correlation feature map; second, encode and decode the target time-series correlation feature map to obtain a target reconstructed time-series correlation feature map corresponding to the target time-series correlation feature map; third, train the fault detection model to be trained based on the target time-series correlation feature map and the target reconstructed time-series correlation feature map. By employing this method, encoding and feature reconstruction processing can be performed on data of the normal category to obtain a trained fault detection model. Model training and fault detection are performed using an autoencoder, enabling fault detection based on the similarity between the original features and the reconstructed features. This method can improve model training accuracy even with suboptimal data distribution and reduces the workload of model training. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating the training method of a fault detection model in one embodiment;

[0073] Figure 2 This is a flowchart illustrating the process of determining training steps in one embodiment;

[0074] Figure 3 This is a flowchart illustrating the steps for determining the target temporal correlation feature map in one embodiment;

[0075] Figure 4 This is a flowchart illustrating the steps for calculating the temporal correlation feature map of the target reconstruction in one embodiment;

[0076] Figure 5 This is a flowchart illustrating the convolution step in one embodiment;

[0077] Figure 6 This is a flowchart illustrating the long short-term memory processing steps in one embodiment;

[0078] Figure 7 This is a flowchart illustrating the deconvolution step in one embodiment;

[0079] Figure 8 This is a flowchart illustrating the fault detection steps in one embodiment;

[0080] Figure 9 This is a flowchart illustrating the fault detection steps in one embodiment;

[0081] Figure 10 This is a flowchart illustrating the fault detection steps in one embodiment;

[0082] Figure 11 This is a flowchart illustrating the fault detection and processing steps in one embodiment;

[0083] Figure 12 This is a flowchart illustrating the fault detection steps in one embodiment;

[0084] Figure 13 This is a schematic diagram of the structure of a fault detection model in one embodiment;

[0085] Figure 14 This is a schematic diagram of the structure of a fault detection model in one embodiment;

[0086] Figure 15 This is a structural block diagram of a training device for a fault detection model in one embodiment;

[0087] Figure 16 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0089] In one embodiment, such as Figure 1As shown, a training method for a fault detection model is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, etc., and the server can be a standalone server or a server cluster composed of multiple servers. In this embodiment, the training method for the fault detection model includes the following steps:

[0090] Step 102: Obtain the timing data of the target device and determine the timing characteristics corresponding to the timing data.

[0091] The time-series data includes at least the first sample data of the normal category, which includes multiple operational status indicators at multiple time points. The target equipment can be infrastructure that ensures the server equipment in the data center operates within a normal temperature range; it can be cooling equipment, such as air conditioning equipment. The time-series data of the target equipment includes at least the operational status indicator data of the air conditioning equipment under normal operating conditions. The multiple time points can be multiple data collection time points corresponding to the data collection period, which can be the time period during which the air conditioning equipment operates normally. The multiple operational status indicators can include, respectively, rack temperature, water valve opening, air conditioning return air temperature, air conditioning fan speed, air conditioning set fan speed, air conditioning minimum set fan speed, etc. Time-series features are feature vector data obtained after feature extraction from the time-series data; for example, they can be feature vector data obtained after feature normalization of the time-series data.

[0092] Specifically, the terminal can collect time-series data from multiple target devices during their normal operating periods, and perform feature extraction processing on the time-series data corresponding to each of the multiple target devices to obtain the time-series features corresponding to each time-series data. In one example, the terminal can perform feature extraction and normalization processing on the collected time-series data from multiple target devices during their normal operating periods to obtain the normalized time-series features corresponding to each time-series data.

[0093] Step 104: Perform time-series correlation feature processing on the time-series features corresponding to multiple operating status indicator data at multiple time points within a preset time window to obtain the target time-series correlation feature map.

[0094] The preset time window can be a sliding time pane corresponding to different time steps, and the target time-series correlation feature map is a multi-dimensional feature correlation map time series used to represent the state information at different time steps. For example, the target time-series correlation feature map can be a feature correlation map representing the operation state indicator features corresponding to the various operation state indicator data contained in the time-series features.

[0095] Specifically, the terminal can extract multiple target time-series features corresponding to the step size of the preset time window from the time-series features corresponding to multiple operation status indicator data at multiple time points, and calculate the feature correlation between the operation status indicator features corresponding to each operation status indicator data contained in each target time-series feature. Based on this, the terminal can obtain a target time-series correlation feature map based on the calculated feature correlation between the operation status indicator features corresponding to each operation status indicator data contained in each pair of target time-series features.

[0096] Step 106: Encode and decode the target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map.

[0097] The encoding and decoding process can include encoding through convolutional layers and feature decoding through deconvolutional layers, or feature reconstruction through deconvolutional layers; the target reconstruction temporal correlation feature map can be the target reconstruction temporal correlation feature map obtained after at least performing feature encoding and feature decoding on the target temporal correlation feature map.

[0098] Specifically, after determining the target temporal correlation feature map, the terminal can encode and decode the target temporal correlation feature map through multiple convolutional layers, multiple image convolutional long short-term memory networks, and multiple deconvolutional layers to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map.

[0099] Step 108: Based on the target temporal correlation feature map and the target reconstruction temporal correlation feature map, train the fault detection model to be trained to obtain the trained target fault detection model.

[0100] The fault detection model to be trained can be a neural network model, such as an unsupervised neural network model. It can be trained based on sample data of normal categories to obtain a fault detection model that meets the preset training completion conditions.

[0101] Specifically, the target time-series correlation feature map can be feature map data calculated by the terminal based on the time-series data contained in the first sample data. The terminal can also encode and decode the target time-series correlation feature map to obtain the target reconstructed time-series correlation feature map corresponding to the target time-series correlation feature map. Based on this, the terminal can calculate the loss function corresponding to the fault detection model to be trained based on the target time-series correlation feature map and the target reconstructed time-series correlation feature map, and update the model parameters corresponding to the fault detection model to be trained based on the loss function. That is, the terminal can calculate the feature map difference between the target time-series correlation feature map and the target reconstructed time-series correlation feature map, obtain the loss function based on the feature map difference, update the model parameters of the fault detection model that has not met the training completion condition based on the loss function, obtain the updated model, and re-execute the steps of acquiring the time-series data of the target device and determining the time-series features corresponding to the time-series data based on the updated model until the preset training completion condition is met, and obtain the trained target fault detection model.

[0102] In one example, if the terminal determines that the current number of training iterations has met the preset training iteration threshold, the terminal can determine that the preset training completion condition has been met. In another example, if the terminal determines that the loss value corresponding to the currently calculated loss function has met the preset convergence condition, the terminal can determine that the preset training completion condition has been met. The preset convergence condition may be that the loss value corresponding to the loss function has not changed, or that the loss value has reached the minimum loss value threshold, etc.

[0103] In the training method of the aforementioned fault detection model, time-series correlation feature processing is performed on multiple operational status index data at multiple time points corresponding to the first sample data containing the normal category to obtain a target time-series correlation feature map. The target time-series correlation feature map is then encoded and decoded to obtain a target reconstructed time-series correlation feature map. Based on the target time-series correlation feature map and the target reconstructed time-series correlation feature map, the fault detection model to be trained is trained. By adopting this method, encoding and feature reconstruction processing can be performed on data of the normal category to obtain a trained fault detection model. Model training and fault detection are performed using an autoencoder, enabling fault detection based on the similarity between the original and reconstructed features. This method can improve model training accuracy even with suboptimal data distribution and reduces the workload of model training.

[0104] In one embodiment, the time-series data further includes second sample data of the fault category; specifically, the time-series data may also include second sample data of the fault category, the second sample data including multiple operating status indicator data at multiple time points; the second sample data of the fault category included in the time-series data of the target device may be operating status indicator data collected by the terminal when the target device is in a fault operating state, the multiple time points may be the collection time periods corresponding to the target device being in a fault operating state, and the multiple operating status indicator data respectively include cabinet measuring point temperature, water valve opening, air conditioner return air temperature, air conditioner fan speed, air conditioner set fan speed, air conditioner minimum set fan speed, etc.

[0105] Accordingly, such as Figure 2 As shown, the specific processing steps of step 108, "training the fault detection model to be trained based on the target temporal correlation feature map and the target reconstruction temporal correlation feature map, to obtain the trained target fault detection model," include:

[0106] Step 202: Based on the target temporal correlation feature map and the target reconstruction temporal correlation feature map, train the fault detection model to be trained to obtain the trained initial fault detection model.

[0107] Specifically, the terminal can calculate the loss function corresponding to the fault detection model to be trained based on the target time-series correlation feature map and the target reconstructed time-series correlation feature map, and update the model parameters corresponding to the fault detection model to be trained based on the loss function to obtain the updated model. Based on this, the terminal can re-execute the steps of acquiring the time-series data of the target device and determining the time-series features corresponding to the time-series data, until the terminal determines that the fault detection model has met the preset training completion conditions under the current situation. In this way, the terminal can obtain the trained initial fault detection model.

[0108] Step 204: Based on the preset analysis algorithm, the second sample data of the fault category is subjected to dimensionality reduction processing to obtain the target fault sample data, and based on the preset time domain feature algorithm, the target time domain feature set corresponding to the target fault sample data is determined.

[0109] The target fault detection model may include a fault detection module and a fault identification module. The fault detection module may be an initial fault detection model; the fault identification module may be an identification module containing the target time-domain feature set; the preset analysis algorithm may be the kernel PCA principal component analysis algorithm, used to perform dimensionality reduction processing on the second sample data of the fault category; the preset time-domain feature algorithm is used to extract the time-domain features after data extraction. The time-domain features may include sample variance, mean square margin, kurtosis, impulse factor, etc., and may also include skewness, kurtosis, margin factor, shape factor, etc.

[0110] Specifically, the terminal can perform dimensionality reduction processing on the second sample data using a preset analysis algorithm to obtain the target fault sample data. In other words, the terminal can use a preset analysis algorithm to perform dimensionality reduction processing on the fault data corresponding to the collection time period when the target device is in a faulty operating state, obtaining the target fault sample data, which can be one-dimensional data. Based on this, the terminal can also extract time-domain features from the target fault sample data using a preset sliding time window and a preset time-domain feature algorithm, obtaining the time-domain features corresponding to the target fault sample data, forming a target time-domain feature set.

[0111] In one example, the terminal can determine the fault identification module based on the target time-domain feature set, which contains the time-domain features corresponding to the operating status index data of the target device in a fault operating state.

[0112] Step 206: Determine the target fault detection model based on the target time-domain feature set and the initial fault detection model.

[0113] Specifically, the terminal can obtain a target fault detection model based on the initial fault detection model and a fault identification module containing the target time-domain feature set. This target fault detection model can realize real-time fault detection of the target device.

[0114] In this embodiment, by performing time-domain feature analysis on the sample data of fault categories, real-time monitoring and fault pre-monitoring of target devices in the data center can be realized. It can also realize multiple determinations of whether the target device has a fault, thereby improving the accuracy of fault detection.

[0115] In one embodiment, such as Figure 3 As shown, the specific processing steps of step 104, "based on the time-series features corresponding to multiple operational status indicator data at multiple time points within a preset time window, perform time-series correlation feature processing to obtain the target time-series correlation feature map," include:

[0116] Step 302: Based on a preset time window, extract time-series features from the time-series features corresponding to multiple operational status indicator data at multiple time points to obtain multiple target time-series features, and calculate the correlation features between each target time-series feature;

[0117] Among them, the correlation features between the time series features of each target can be the correlation features between the operational status indicator features corresponding to the multiple operational status indicator data contained in each target time series.

[0118] Specifically, the terminal can extract multiple time-series features from the time-series data based on a preset time window, which are then used as target time-series features. Each target time-series feature contains multiple operational status indicator features corresponding to different operational status indicator data, and the terminal can calculate the correlation features between each operational status indicator feature.

[0119] In one example, the terminal can calculate the correlation features between each target time series feature in pairs. The calculation of each target time series feature in pairs can be a first target time series feature and a second target time series feature. The specific calculation process can include: the first target time series feature can include multiple first operating state indicator features, and the second target time series feature can include multiple second operating state indicator features. The terminal can calculate the first correlation between each first operating state indicator feature, and calculate the second correlation between each second operating state indicator feature. It can also calculate the third correlation between the first operating state indicator feature and the second operating state indicator feature, and calculate the fourth correlation between the second operating state indicator feature and the first operating state indicator feature. Based on each first correlation, each second correlation, each third correlation, and each fourth correlation, the terminal can obtain the correlation feature between the first target time series feature and the second target time series feature.

[0120] Step 304: Based on the correlation characteristics between the time series features of each target, determine the target time series correlation feature map.

[0121] Specifically, after obtaining multiple target time-series features, the terminal can calculate the correlation features between each pair of target time-series features. The terminal can combine the correlation features between each target time-series feature to obtain a target time-series correlation feature map. The target time-series correlation feature map can be a correlation feature map matrix.

[0122] In this embodiment, by performing feature extraction processing on time series data and calculating the correlation graph features, a multi-dimensional feature correlation graph time series can be constructed when facing multivariate time series feature attributes. This accurately extracts the state information represented in the time series data at different time steps, providing a stable data foundation for subsequent processing.

[0123] In one embodiment, such as Figure 4 As shown, the specific processing steps of step 106, "encoding and decoding the target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map," include:

[0124] Step 402: Perform convolution processing on the target temporal correlation feature map to obtain the convolution-processed target temporal correlation feature map.

[0125] The convolutionally processed target temporal correlation feature map includes multiple target temporal correlation feature maps of different dimensions. The fault detection model (initial fault model, or fault detection module) may include a convolutional module, which is used to encode the target temporal correlation feature map.

[0126] Specifically, the terminal can encode the target temporal-related feature map through a convolutional module to obtain the target temporal-related feature map after convolution. The convolutional module can contain multiple convolutional layers of different scales. For example, the convolutional module can include a first convolutional layer of a first scale, a second convolutional layer of a second scale, and a third convolutional layer of a third scale.

[0127] In one example, the first scale can be smaller than the second scale, and the second scale can be smaller than the third scale. Specifically, the terminal can encode the target temporal correlation feature map through the first convolutional layer at the first scale to obtain the target temporal correlation feature map of the first dimension corresponding to the first scale. The terminal can then input the target temporal correlation feature map of the first dimension corresponding to the first scale into the second convolutional layer at the second scale, and encode the target temporal correlation feature map of the first dimension corresponding to the first scale through the second convolutional layer at the second scale to obtain the target temporal correlation feature map of the second dimension corresponding to the second scale. Similarly, the terminal can input the target temporal correlation feature map of the second dimension corresponding to the second scale into the third convolutional layer for encoding to obtain the target temporal correlation feature map of the third dimension corresponding to the third scale. The target temporal correlation feature map of the first dimension, the target temporal correlation feature map of the second dimension, and the target temporal correlation feature map of the third dimension are used as the target temporal correlation feature map after convolution.

[0128] Step 404: Based on the convolutional long short-term memory network, the target temporal-related feature maps of multiple different dimensions are processed to obtain multiple temporal pattern information.

[0129] The convolutional long short-term memory (SLSTM) network can be an image SLSTM network used for temporal analysis of the target temporally relevant feature map. The fault detection model (initial fault model, or fault detection module) can also include a SLSTM network, which may include multiple SLSTM networks. For example, in one example, it may include a first SLSTM sub-network, a second SLSTM sub-network, and a third SLSTM sub-network. The number of SLSTM sub-networks included in the SLSTM network can be consistent with the number of convolutional layers included in the convolutional module.

[0130] Specifically, the terminal can process multiple target time-related feature maps of different dimensions based on the multiple convolutional long short-term memory sub-networks contained in the convolutional long short-term memory network, and extract the time pattern information in each target time-related feature map as the output result of the corresponding convolutional long short-term memory sub-network.

[0131] In one example, the terminal can process the first-dimensional target temporal correlation feature map through the first convolutional long short-term memory sub-network to obtain the first temporal pattern information corresponding to the first-dimensional target temporal correlation feature map. Similarly, the terminal can obtain the second temporal pattern information corresponding to the second-dimensional target temporal correlation feature map and the third temporal pattern information corresponding to the third-dimensional target temporal correlation feature map.

[0132] Step 406: Based on multiple target time-series related feature maps of different dimensions and multiple time pattern information, feature reconstruction processing is performed to obtain the target reconstruction time-series related feature map.

[0133] The target reconstruction temporal correlation feature map is the temporal correlation feature map obtained by the terminal after encoding and decoding the target temporal correlation feature map. The fault detection model (initial fault model, or fault detection module) may also include a deconvolution module, which may contain multiple deconvolution layers of different scales. The number of deconvolution layers in the deconvolution module may be the same as the number of convolution layers in the convolution module.

[0134] Specifically, the terminal can perform superposition processing based on the target temporal-related feature maps of different dimensions and the time pattern information corresponding to the target temporal-related feature maps of that dimension to obtain the superposition results corresponding to each dimension. Then, based on multiple superposition results and the multiple deconvolution layers contained in the deconvolution module, the terminal performs decoding processing to obtain the target reconstruction temporal-related feature map. In other words, the terminal can perform feature reconstruction processing based on multiple superposition results and the multiple deconvolution layers contained in the deconvolution module to obtain the target reconstruction temporal-related feature map.

[0135] In this embodiment, unsupervised learning of the model can be achieved through convolutional layers, deconvolutional layers, and an autoencoder of an image convolutional long short-term memory network, avoiding a large amount of labeling work on sample data. Encoding the target temporal-related feature map through convolutional layers can remove noise information in the target temporal-related feature map and achieve data compression, ensuring data accuracy while avoiding data redundancy.

[0136] In one embodiment, the multi-dimensional temporal correlation features include at least a first-dimensional target temporal correlation feature map, a second-dimensional target temporal correlation feature map, and a third-dimensional target temporal correlation feature map. Specifically, the target temporal correlation feature maps of each dimension are output by the convolutional module. The convolutional module may include a first convolutional layer at a first scale, a second convolutional layer at a second scale, and a third convolutional layer at a third scale. The first scale may be smaller than the second scale, the second scale may be smaller than the third scale, the dimension corresponding to the convolutional layer at the first scale may be 32 dimensions, the dimension corresponding to the convolutional layer at the second scale may be 64 dimensions, and the dimension corresponding to the convolutional layer at the third scale may be 128 dimensions.

[0137] In one example, the stride of the first convolutional layer can be 1*1, the stride of the second convolutional layer can be 2*2, and the stride of the third convolutional layer can be 2*2.

[0138] Accordingly, such as Figure 5 As shown, the specific processing steps of step 402, "convolution processing the temporal correlation feature map to obtain the convolutionally processed temporal correlation feature map," include:

[0139] Step 502: Perform convolution processing on the target temporal correlation feature map through the first convolutional layer to obtain the target temporal correlation feature map of the first dimension.

[0140] Specifically, the terminal can encode the target temporal correlation feature map through the first convolutional layer. That is, the terminal can input the target temporal correlation feature map into the first convolutional layer, so that the first convolutional layer encodes the target temporal correlation feature map within its first convolutional layer, obtaining the target temporal correlation feature map of the first dimension corresponding to the first scale. In one example, the target temporal correlation feature map of the first dimension output by the first convolutional layer obtained by the terminal can be a 32-dimensional target temporal correlation feature map.

[0141] Step 504: Perform convolution processing on the first-dimensional target temporal correlation feature map through the second convolutional layer to obtain the second-dimensional target temporal correlation feature map.

[0142] Specifically, the terminal can encode the first-dimensional target temporal correlation feature map using a second convolutional layer. In other words, the terminal can input the first-dimensional target temporal correlation feature map into the second convolutional layer, which then encodes it to obtain the second-dimensional target temporal correlation feature map corresponding to the second scale. In one example, the second-dimensional target temporal correlation feature map output by the second convolutional layer can be a 64-dimensional target temporal correlation feature map.

[0143] Step 506: Perform convolution processing on the second-dimensional target temporal correlation feature map through the third convolutional layer to obtain the third-dimensional target temporal correlation feature map.

[0144] Specifically, the terminal can encode the second-dimensional target temporal correlation feature map using a third convolutional layer. In other words, the terminal can input the second-dimensional target temporal correlation feature map into the third convolutional layer, allowing the third convolutional layer to encode it and obtain the third-dimensional target temporal correlation feature map corresponding to the third scale. In one example, the third-dimensional target temporal correlation feature map output by the third convolutional layer can be a 128-dimensional target temporal correlation feature map.

[0145] In this embodiment, the target temporal-related feature map is encoded by multiple convolutional layers of different scales. This removes noise information from the target temporal-related feature map and also compresses it. At the same time, the convolutional-convolutional long short-term memory network autoencoder model in the deep learning autoencoder can transform the data into a temporal map. This allows the model to monitor multiple temporal variables of the target device, as well as the temporal variables of the target device and its neighboring target devices. This makes the status monitoring information of the target device more comprehensive and complete, providing an accurate data foundation. Furthermore, the fault detection model trained based on this data can be used to determine the accuracy of fault detection of the target device.

[0146] In one embodiment, such as Figure 6 As shown, the specific processing steps of step 404, "based on a convolutional long short-term memory network, process the target temporal correlation feature maps of multiple dimensions to obtain multiple temporal pattern information," include:

[0147] Step 602: Perform temporal analysis on the target temporal correlation feature map of the first dimension based on the first convolutional long short-term memory sub-network to obtain the first time pattern information.

[0148] The first convolutional long short-term memory subnetwork can be a Conv LSTM image convolutional long short-term memory network.

[0149] Specifically, the terminal can perform temporal analysis on the target temporal correlation feature map of the first dimension through the first convolutional long short-term memory sub-network, that is, obtain the temporal correlation information between the compressed temporal correlation feature maps of multiple steps, and determine the obtained temporal correlation information as the first time pattern information.

[0150] Step 604: Perform temporal analysis on the target temporal correlation feature map of the second dimension based on the second convolutional long short-term memory sub-network to obtain the second temporal pattern information.

[0151] The second convolutional long short-term memory subnetwork can be a Conv LSTM image convolutional long short-term memory network.

[0152] Specifically, the terminal can perform temporal analysis on the target temporal correlation feature map of the second dimension through the second convolutional long short-term memory sub-network, that is, to obtain the temporal correlation information between the compressed temporal correlation feature maps of multiple steps, and to determine the obtained temporal correlation information as the second time pattern information.

[0153] Step 606: Perform temporal analysis on the target temporal correlation feature map of the third dimension based on the third convolutional long short-term memory sub-network to obtain the third temporal pattern information.

[0154] The third convolutional long short-term memory subnetwork can be a Conv LSTM image convolutional long short-term memory network.

[0155] Specifically, the terminal can perform temporal analysis on the target temporal correlation feature map of the third dimension through the third convolutional long short-term memory sub-network, that is, to obtain the temporal correlation information between the compressed temporal correlation feature maps of multiple steps, and to determine the obtained temporal correlation information as the third temporal pattern information.

[0156] In this embodiment, by using a convolutional long short-term memory network to process the temporal correlation information of the target dimension's temporal correlation feature map, the temporal data can be transformed into a temporal correlation feature map. This allows the fault detection model to simultaneously monitor multiple temporal variables of the target device, as well as the temporal variables of the target device and neighboring target devices. This makes the status monitoring information of the target device more comprehensive and complete, providing an accurate data foundation.

[0157] In one embodiment, such as Figure 7 As shown, the specific processing steps of step 406, "performing feature reconstruction based on multi-dimensional target temporal correlation feature maps and multiple time pattern information to obtain target reconstruction temporal correlation feature maps," include:

[0158] Step 702: Through the first deconvolution layer, the superposition result of the first time pattern information and the first dimension target time-related feature map is processed to perform feature reconstruction to obtain the first reconstructed time-related feature map corresponding to the first dimension target time-related feature map.

[0159] The temporal-related feature maps of the target in each dimension are output by the convolutional module. The convolutional module may include a first deconvolutional layer at the fourth scale, a second deconvolutional layer at the fifth scale, and a third deconvolutional layer at the sixth scale.

[0160] Specifically, the terminal can overlay the first temporal pattern information and the first-dimensional target temporal correlation feature map to obtain the overlay result of the first temporal pattern information and the first-dimensional target temporal correlation feature map. The terminal can then decode this overlay result through a first deconvolutional layer. In other words, the terminal can input the overlay result of the first temporal pattern information and the first-dimensional target temporal correlation feature map into the first deconvolutional layer, so that the first deconvolutional layer performs feature reconstruction processing on the overlay result of the first temporal pattern information and the first-dimensional target temporal correlation feature map within the first deconvolutional layer, obtaining the first reconstructed temporal correlation feature map corresponding to the first-dimensional target temporal correlation feature map. In one example, the first reconstructed temporal correlation feature map corresponding to the first-dimensional target temporal correlation feature map obtained by the terminal can be a 64-dimensional target temporal correlation feature map.

[0161] Step 704: Through the second deconvolution layer, feature reconstruction processing is performed on the superposition result of the first reconstructed temporal correlation feature map, the second time pattern information, and the second-dimensional target temporal correlation feature map to obtain the second reconstructed temporal correlation feature map corresponding to the second-dimensional target temporal correlation feature map.

[0162] Specifically, the terminal can overlay the second time pattern information and the second-dimensional target time-series correlation feature map to obtain a first overlay result of the second time pattern information and the second-dimensional target time-series correlation feature map. Simultaneously, the terminal can also overlay the first overlay result with the first reconstructed time-series correlation feature map output by the first deconvolution layer to obtain a second overlay result. Based on this, the terminal can decode the second overlay result through the second deconvolution layer. That is, the terminal can overlay the first overlay result of the second time pattern information and the second-dimensional target time-series correlation feature map with the first reconstructed time-series correlation feature map output by the first deconvolution layer, and input the resulting second overlay result into the second deconvolution layer. The second deconvolution layer then performs feature reconstruction processing on the second overlay result within the second deconvolution layer to obtain the second reconstructed time-series correlation feature map corresponding to the second-dimensional target time-series correlation feature map. In one example, the second overlay result obtained by the terminal can be a 128-dimensional target time-series correlation feature map, and the second reconstructed time-series correlation feature map output by the second deconvolution layer can be a 32-dimensional target time-series correlation feature map.

[0163] Step 706: Through the third deconvolution layer, feature reconstruction processing is performed on the superposition result of the second reconstructed temporal correlation feature map, the third temporal pattern information and the third dimension target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map.

[0164] Specifically, the terminal can overlay the third temporal pattern information and the third-dimensional target temporal correlation feature map to obtain a third overlay result. Simultaneously, the terminal can overlay this third overlay result with the second reconstructed temporal correlation feature map output from the second deconvolution layer to obtain a fourth overlay result. Based on this, the terminal can decode this fourth overlay result through the third deconvolution layer. In other words, the terminal can overlay the third overlay result of the third temporal pattern information and the third-dimensional target temporal correlation feature map with the second reconstructed temporal correlation feature map output from the second deconvolution layer to obtain a fourth overlay result. This fourth overlay result is then input into the third deconvolution layer, whereby the third deconvolution layer performs feature reconstruction processing on the fourth overlay result to obtain the target reconstructed temporal correlation feature map corresponding to the third-dimensional target temporal correlation feature map.

[0165] In this embodiment, feature extraction and feature reconstruction are performed using a self-written decoder in unsupervised mode, which can accurately extract features of normal category data and achieve accurate fault detection of the target device.

[0166] In one embodiment, such as Figure 8 As shown, the training method for this fault detection model also includes:

[0167] Step 802: Obtain the timing data to be tested from the target device, and determine the timing features to be tested corresponding to the timing data to be tested.

[0168] The time-series data to be tested for the target device can be the operational status index data of the target device within a preset time period, and the preset time period can be the data collection period determined based on the actual application scenario.

[0169] Specifically, for the target device to be detected, the terminal can detect the operating status index data of the target device in real time within a preset time period, and perform feature extraction processing on the time sequence data to be detected to obtain the time sequence features corresponding to the time sequence data to be detected.

[0170] Step 804: Determine the time-series related feature map to be detected based on the time-series features to be detected, and input the time-series related feature map to be detected into the target fault detection model to obtain the reconstructed time-series related feature map to be detected.

[0171] Specifically, the terminal can extract multiple target time-series features corresponding to the step size of the preset time window from the multiple operational status index data at multiple time points contained in the time-series data to be detected, based on the preset time window. Then, it can calculate the feature correlation between the operational status index features corresponding to each operational status index data contained in each target time-series feature. Based on this, the terminal can obtain the time-series correlation feature map to be detected based on the calculated feature correlation between the operational status index features corresponding to each operational status index data contained in each pair of target time-series features.

[0172] In this way, the terminal can input the temporal correlation feature map to be detected into the trained target fault detection model. The target fault detection model can use its convolutional modules to perform convolution processing on the temporal correlation feature map to obtain the convolutionally processed target temporal correlation feature map. The terminal can also process the convolutionally processed target temporal correlation feature map through the convolutional long short-term memory network included in the target fault detection model to obtain temporal pattern information. Based on this, the terminal can use the deconvolutional modules included in the target fault detection model to decode the convolutionally processed target temporal correlation feature map and the temporal pattern information to obtain the reconstructed temporal correlation feature map to be detected.

[0173] Step 806: Based on the timing-related feature map to be detected and the reconstructed timing-related feature map to be detected, determine the fault detection result corresponding to the timing data to be detected of the target device.

[0174] Specifically, after determining the reconstruction time-series correlation feature map to be detected, the terminal can compare the reconstruction time-series correlation feature map to be detected with the corresponding time-series correlation feature map of the time-series data to be detected. The terminal can determine the fault detection result corresponding to the time-series data to be detected, i.e., the fault detection result of the target device, based on the comparison result between the reconstruction time-series correlation feature map to be detected and the corresponding time-series correlation feature map of the target fault detection model. This result can indicate either a fault or no fault. Simultaneously, the terminal can also output the time periods during which the target device experienced a fault and the time periods during which no fault occurred.

[0175] In one example, based on the target fault detection model, the terminal can calculate the feature map difference between the reconstructed time-series related feature map to be detected and the detected time-series related feature map. If the terminal determines that the calculated feature map difference is greater than or equal to a preset feature map difference threshold, the terminal can determine that the fault detection result for the time period corresponding to the time-series data to be detected is a fault detection result; if the terminal determines that the calculated feature map difference is less than the preset feature map difference threshold, the terminal can determine that the fault detection result for the time period corresponding to the time-series data to be detected is a fault-free detection result.

[0176] In another example, based on the target fault detection model, the terminal can calculate the feature map difference between the reconstructed temporal correlation feature map to be detected and the detected temporal correlation feature map, and count the first time period when the feature map difference is greater than or equal to a preset feature map difference threshold, and the second time period when the feature map difference is less than the preset feature map difference threshold. Based on this, the fault detection result output by the terminal through the target fault detection model can be that the target device has a fault in the first time period and has not a fault in the second time period.

[0177] In this embodiment, by using a self-written decoder in unsupervised mode to perform fault detection on the timing data of the target device, accurate fault detection results can be obtained.

[0178] In one embodiment, such as Figure 9 As shown, the specific processing steps of step 806, "determining the fault detection result corresponding to the timing data of the target device based on the timing correlation feature map to be detected and the reconstructed timing correlation feature map to be detected," include:

[0179] Step 902: If the timing-related feature map to be detected and the reconstructed timing-related feature map to be detected do not meet the first similarity condition, determine that the fault detection result corresponding to the timing data to be detected of the target device is a fault result.

[0180] The first similarity condition can be that the feature map difference between the detected temporal correlation feature map and the reconstructed temporal correlation feature map is greater than or equal to a preset feature map difference threshold. The fault detection result can characterize whether the target device has malfunctioned, is in an abnormal operating state, etc.

[0181] Specifically, the terminal can determine the fault detection result corresponding to the time series data to be detected by comparing the reconstructed time series correlation feature map to be detected with the time series correlation feature map to be detected using the target fault detection model. For example, based on the target fault detection model, the terminal can calculate the feature map difference between the reconstructed time series correlation feature map to be detected and the time series correlation feature map to be detected. If the terminal determines that the calculated feature map difference is greater than or equal to a preset feature map difference threshold, then the terminal can determine that the fault detection result for the time period corresponding to the time series data to be detected is a fault detection result.

[0182] Step 904: If the timing-related feature map to be detected and the reconstructed timing-related feature map to be detected satisfy the first similarity condition, determine that the fault detection result corresponding to the timing data to be detected of the target device is a non-fault result.

[0183] Specifically, the terminal can determine the fault detection result corresponding to the time series data to be detected by comparing the reconstructed time series correlation feature map to be detected with the time series correlation feature map to be detected using the target fault detection model. For example, based on the target fault detection model, the terminal can calculate the feature map difference between the reconstructed time series correlation feature map to be detected and the time series correlation feature map to be detected. If the terminal determines that the calculated feature map difference is less than a preset feature map difference threshold, the terminal can determine that the fault detection result for the time period corresponding to the time series data to be detected is a detection result of no fault occurring.

[0184] In this embodiment, by using a self-written decoder in unsupervised mode to perform fault detection on the timing data of the target device, accurate fault detection results can be obtained.

[0185] In one embodiment, such as Figure 10 As shown, the specific processing procedure for step 902, "when the timing-related feature map to be detected is inconsistent with the reconstructed timing-related feature map to be detected, determine that the fault detection result corresponding to the timing data to be detected of the target device is a fault result," includes:

[0186] Step 1002: If the time-series correlation feature map to be detected and the reconstructed time-series correlation feature map to be detected do not meet the first similarity condition, the time-series data to be detected of the target device is subjected to dimensionality reduction processing to obtain the target time-series data to be detected.

[0187] Specifically, based on the target fault detection model, the terminal can calculate the feature map difference between the reconstructed time-series related feature map to be detected and the time-series related feature map to be detected. If the terminal determines that the calculated feature map difference is greater than or equal to the preset feature map difference threshold, the terminal can determine that the time-series related feature map to be detected and the reconstructed time-series related feature map to be detected do not meet the first similarity condition. Based on this, the terminal can input the time-series data to be detected into the fault identification module included in the target fault detection model, and perform dimensionality reduction processing on the time-series data to be detected of the target device through the preset analysis algorithm in the fault identification module to obtain the target time-series data to be detected.

[0188] Step 1004: Based on the preset time-domain feature algorithm, determine the time-domain features to be detected corresponding to the target time-series data to be detected.

[0189] Specifically, the terminal can use the preset time-domain feature algorithm in the fault identification module to extract time-domain features from the target time-series data to be detected, thereby obtaining the time-domain features to be detected corresponding to the target time-series data.

[0190] Step 1006: If the time-domain feature to be detected and the target time-domain feature set satisfy the second similarity condition, then the fault detection result corresponding to the time-series data to be detected of the target device is determined to be the fault result.

[0191] Specifically, the terminal can perform fault detection on the time-domain features to be detected based on the target time-domain feature set contained in the fault identification module. The specific detection process can be as follows: based on multiple target time-domain features contained in the target time-domain feature set, the terminal can calculate the similarity between each target time-domain feature and the time-domain feature to be detected based on a preset similarity algorithm; filtering is performed on the similarity between each target time-domain feature and the time-domain feature to be detected, and if there is a similarity greater than or equal to a preset similarity threshold, the terminal can determine that the time-domain feature to be detected and the target time-domain feature set meet the second similarity condition, that is, the terminal can determine that the fault detection result corresponding to the time-series data to be detected of the target device is a fault result.

[0192] In other words, if the terminal determines that the initial fault detection result corresponding to the timing data to be detected of the target device is a fault result based on the fault detection module in the target fault detection model, the terminal can input the timing data to be detected of the target device into the fault identification module in the target fault detection model to perform a second judgment on the timing data to be detected of the target device and obtain the fault detection result.

[0193] In this embodiment, by performing multiple judgments on the timing data to be detected, the accuracy of fault detection of the target device can be guaranteed.

[0194] In one embodiment, such as Figure 11As shown, the specific processing procedure of step 1006, "If the time-domain feature to be detected and the target time-domain feature set satisfy the second similarity condition, then determine that the fault detection result corresponding to the time-series data to be detected of the target device is a fault result," includes:

[0195] Step 1102: Calculate the similarity between the time-domain feature to be detected and each target time-domain feature contained in the target time-domain feature set.

[0196] Specifically, the terminal can calculate the similarity between each target time-domain feature and the time-domain feature to be detected based on multiple target time-domain features contained in the target time-domain feature set, using a preset similarity algorithm. The preset similarity algorithm can be a similarity distance algorithm, such as a cosine distance algorithm, etc.

[0197] Step 1104: If there is a target time-domain feature with a similarity greater than the preset similarity threshold, then the fault detection result corresponding to the time-series data to be detected of the target device is determined to be a fault result.

[0198] Specifically, the terminal can filter among the similarities between each target time-domain feature and the time-domain feature to be detected. If there is a similarity greater than or equal to the preset similarity threshold, the terminal can determine that the time-domain feature to be detected and the target time-domain feature set satisfy the second similarity condition. That is, the terminal can determine that the fault detection result corresponding to the time-series data to be detected of the target device is the fault result.

[0199] In other words, if the terminal determines that the initial fault detection result corresponding to the timing data to be detected of the target device is a fault result based on the fault detection module in the target fault detection model, the terminal can input the timing data to be detected of the target device into the fault identification module in the target fault detection model to perform a second judgment on the timing data to be detected of the target device and obtain the fault detection result.

[0200] In this embodiment, by performing multiple judgments on the timing data to be detected, the accuracy of fault detection of the target device can be guaranteed.

[0201] In one embodiment, such as Figure 12 As shown, the training method for this fault detection model also includes:

[0202] Step 1202: Filter among the similarities that are greater than or equal to the preset similarity threshold, determine the largest similarity, and determine the target time-domain feature corresponding to the largest similarity.

[0203] Specifically, after calculating the similarity between each time-domain feature contained in the target time-domain feature set and the time-domain feature to be detected, the terminal can filter among the similarities that are greater than or equal to the preset similarity threshold and determine the target time-domain feature corresponding to the largest similarity value.

[0204] Step 1204: Based on the preset correspondence between time-domain features and fault types, determine the target fault type corresponding to the target time-domain features, and determine the fault type corresponding to the time-series data to be tested of the target device as the target fault type.

[0205] Specifically, the terminal can determine the target fault type corresponding to the target time domain feature based on the preset correspondence between time domain features and fault types; in this way, the terminal can determine that the fault detection result of the target device is that a fault has occurred, and the fault type is the target fault type corresponding to the target time domain feature.

[0206] In this embodiment, by performing multiple judgments on the timing data to be detected, the accuracy of fault detection of the target device can be guaranteed, and the type of fault can be accurately determined.

[0207] The following describes in detail the specific execution process of the training method for the above-mentioned fault detection model, with reference to a specific embodiment:

[0208] In related technologies, air conditioning and other cooling equipment can experience various malfunctions during operation, leading to cooling failures in data centers, resulting in server overheating alarms or even system crashes, and indirectly causing a surge in data center energy consumption and electricity costs. Therefore, establishing an effective fault diagnosis and isolation mechanism for data center air conditioning systems is of great significance.

[0209] Methods for fault diagnosis of cooling equipment in data centers include model-driven and data-driven approaches. Model-driven methods start from the design mechanism of the air conditioning system, mathematically model the cooling control logic, and compare the model results with monitored variables in real time to achieve fault diagnosis. However, air conditioning systems, especially water-cooled air conditioners, are large-scale and have complex design mechanisms, making model building and diagnostic implementation difficult. Data-driven methods mainly rely on the massive amounts of data generated during operation, automatically mining safety operation and maintenance knowledge and patterns to achieve equipment fault diagnosis. Especially with the rapid development of artificial intelligence technologies such as machine learning and deep learning, diagnostic methods based on various intelligent models can autonomously form diagnostic models from massive amounts of data and continuously improve themselves, thereby increasing the accuracy of fault diagnosis.

[0210] Currently, using artificial intelligence models for fault detection in refrigeration equipment has become a common method. However, several problems remain. For example, existing fault diagnosis models are trained under ideal data distribution conditions. On the one hand, this requires a balanced ratio of fault data to normal data. However, faults are sporadic events, leading to a severe imbalance between fault and normal data, resulting in high model diagnostic errors. On the other hand, model training requires clear feature distinctions between fault and normal data to facilitate learning fault patterns. However, in reality, due to noise and other factors, the two may be highly similar, making model building difficult and resulting in frequent false alarms. Furthermore, most existing intelligent model-based air conditioning fault detection methods employ supervised learning strategies, meaning that extensive manual annotation of data using domain knowledge is required before model training, increasing labor costs.

[0211] The training method for a fault detection model provided in this application embodiment can be a data center air conditioning fault detection method based on deep learning autoencoders and temporal feature analysis, such as... Figure 13 The diagram shown may be a structural schematic of the target fault detection model provided in the embodiments of this application, as well as an application flowchart.

[0212] The target fault detection model may include a fault detection module and a fault identification module, and may also include a data processing module. The application process of the data processing module may include: acquiring time-series data of the computer room cooling equipment and Z-Score data standardization; the application process of the fault detection module may include: acquiring normal class data (i.e., the first sample data of the normal category) output by the data processing module, constructing a time-series correlation graph, convolutional encoder feature encoding, convolutional-long short-term memory time-series analysis, convolutional decoder decoding, and acquiring an autoencoder model;

[0213] The application process of the fault identification module may include: fault class data, Kernel-PCA data dimensionality reduction, and construction of a time-domain feature judgment set for air conditioning faults.

[0214] The application flowchart can be for fault detection and judgment: online data, real-time detection of the air conditioner status of the self-encoder, whether it is abnormal, in the case of abnormality, Kernel-PCA data dimensionality reduction and calculation of time domain features, and K-neighbor fault identification.

[0215] Specifically, the fault detection module is mainly implemented through an autoencoder based on a Convolutional Long Short-Term Memory (CNN-LSTM) network. When encountering abnormal data, it calculates the difference between the encoder's encoding result and the encoding result of normal data to initially determine the abnormal state and time period of the device. Simultaneously, the terminal can perform specific fault determination through the fault determination module, which mainly consists of Kernel-PCA principal component analysis, sliding window temporal feature calculation, and a K-nearest neighbor (K-neighbor) classifier. Its basic principle is as follows: first, the data initially determined to be from an abnormal period is subjected to Kernel-PCA dimensionality reduction to obtain key feature information; second, the temporal feature information of the data is calculated, including mean, variance, margin, kurtosis, etc.; finally, the temporal features are compared with various fault features using K-neighbor to determine whether a fault exists and its category.

[0216] like Figure 14 As shown, in one example, it could be a model structure diagram of the target fault detection model, or it could be a structure diagram of the fault detection module in the target fault detection model.

[0217] Step 1: The terminal can construct a time-series correlation graph. Based on the multivariate time-series feature attributes of the target device, a multi-dimensional feature correlation graph time series can be constructed to characterize the system's state information at different time steps. In other words, for multiple operating state indicator data of the target device, a target time-series correlation feature graph can be obtained based on the multiple operating state indicator data of the target device.

[0218] In one example, the time series characteristics of the normal state are: The size of the sliding pane (preset time window) is set to W=[5,10,15]. A specific sliding pane... Any two time series features are denoted as: and The resulting correlation graph matrix is:

[0219]

[0220] Where x can be the time series feature corresponding to the time series data, T is the data acquisition period, x1 can be the time series feature corresponding to the time series data acquired at the first acquisition time point within the data acquisition period, x2 can be the time series feature corresponding to the time series data acquired at the second acquisition time point within the data acquisition period, and so on; t can be the duration of the sliding window, and tw can be the time point corresponding to the first sliding step in the sliding window. express and The target temporal correlation feature map between them It is the scaling factor.

[0221] Step 2, convolutional encoder feature encoding; the terminal can encode the temporal correlation map (target temporal correlation feature map) through the convolutional encoder, remove noise information in the feature correlation matrix, and compress the data at the same time.

[0222] In the training method of the fault detection model provided in this application embodiment, three convolutional encodings can be performed. The size of the convolutional kernel in each layer can be as follows: the kernel size of the first convolutional layer conv1 is: stride: 1*1, channel: 32; the kernel size of the second convolutional layer conv2 is: stride: 2*2, channel: 64; and the kernel size of the third convolutional layer conv3 is: stride: 2*2, channel: 128. The stride can be h (which can be denoted as h-steps).

[0223] Specifically, the time-series correlation graph sequence determined by the terminal can be denoted as: The relevant graph at any given time is: The result after any convolutional encoding layer is denoted as:

[0224]

[0225] in These are convolution weights. This is a convolution operation.

[0226] Step 3, Convolutional Long Short-Term Memory Temporal Analysis: The terminal can capture temporal pattern information in the target temporal-related feature map sequence through the Convolutional Long Short-Term Memory (ConvLSTM) network.

[0227] Long Short-Term Memory (LSTM) analysis is performed on the encoded features obtained from three convolutions (which can be denoted as the first-dimensional target temporal correlation feature map, the second-dimensional target temporal correlation feature map, and the third-dimensional target temporal correlation feature map). Assuming the ConvLSTM allows an input data time step of [value missing], [the analysis is not provided in the original text]. For each set of encoded features, the output of the ConvLSTM network is:

[0228] ,

[0229] ,

[0230] ,

[0231] ,

[0232] ,

[0233] .

[0234] Step 4: Feature Decoding by the Convolutional Decoder. The terminal can reconstruct the feature matrix using the convolutional decoder based on the compressed feature correlations and temporal information. The convolutional decoder includes a first deconvolutional layer (DeConv3), a second deconvolutional layer (DeConv2), and a third deconvolutional layer (DeConv1). The terminal can sequentially deconvolve the encoded results after the three convolutions. Before feature reconstruction at each deconvolutional layer, the LSTM output is superimposed with the temporally correlated feature maps of different dimensions output by each convolutional layer. The superimposed result is then input into the deconvolutional layer. The result is denoted as:

[0235]

[0236] in, This is a deconvolution operation.

[0237] Specifically, the terminal can calculate the loss function based on the reconstructed temporal feature matrix (target reconstructed temporal correlation feature map) and the target temporal correlation feature map, and update the model parameters of the fault detection model based on the calculated loss function until the preset training completion conditions are met, thus obtaining the trained fault detection model.

[0238] In one example, the target fault detection model includes a fault identification module, which can analyze abnormal alarm data of time-series fault detection to obtain patterns in the fault data. The details of the fault identification module are as follows:

[0239] Step 1: KernelPCA data dimensionality reduction. The data for the time period of autoencoder alarm anomalies is reduced to 1D features to extract the main information representing the equipment fault state. For the original multidimensional features, an orthogonal space is reconstructed so that the data points are distributed on one axis of the new coordinate system, while preserving as much of the main information of the original data distribution as possible.

[0240]

[0241] Step 2, Temporal Feature Analysis: The terminal can perform temporal feature analysis on the dimensionality-reduced data, extracting key temporal features using a dynamic sliding pane, including:

[0242] Sample variance:

[0243] Mean square value:

[0244] Margin:

[0245] kurtosis:

[0246] Pulse factor:

[0247] In diagnostic tasks, other time-domain features, such as skewness, kurtosis, margin factor, and shape factor, can also be used according to actual needs.

[0248] Step 3, Fault determination and classification: Using the K-neighbor algorithm, the abnormal time-domain feature data is compared with the time-domain feature judgment set to determine whether there is a fault and the specific fault type.

[0249] The execution process of the training method for the above fault detection model will be described in detail with another example:

[0250] Phase 1: Model Construction and Training

[0251] Step 1: Obtain the timing data of the computer room air conditioning equipment, including: rack measuring point temperature, water valve opening, air conditioning return air temperature, air conditioning fan speed, air conditioning set fan speed, air conditioning minimum set fan speed, etc.

[0252] Step 2: Perform feature normalization on the time-series data of the computer room.

[0253] Step 3: Based on the time-series data of the computer room after feature normalization, that is, using the normal state data of the cooling equipment, construct a CNN+ConvLSTM autoencoder fault detection model.

[0254] Step 3-1: Based on the data center time series data after feature normalization, determine the high-dimensional time series features (target time series features), calculate the correlation between them pairwise, and construct the target time series correlation feature map.

[0255] Step 3-2: Through three layers of CNN convolution, the target temporal correlation feature map is compressed and denoised in sequence to obtain the first target temporal correlation feature map, the second target temporal correlation feature map and the third target temporal correlation feature map of different dimensions.

[0256] Step 3-3: By using the ConvLSTM image convolutional long short-term memory network, the temporal correlation information in the compressed target temporal correlation feature map with h steps can be captured, thereby obtaining the first temporal mode information, the second mode information, and the third mode information.

[0257] Steps 3-4 involve reconstructing the temporal-related feature maps of the first, second, and third targets from different dimensions using three layers of DCNN deconvolution. Before each reconstruction layer, the results from the LSTM are superimposed with the unreconstructed results before deconvolution.

[0258] Steps 3-5: Calculate the difference between the reconstructed time series graph and the original data graph, and use gradient descent to complete model training.

[0259] Step 4: Use KernelPCA principal component analysis to reduce the dimensionality of the high-dimensional outlier data (second sample data) to 1-dimensional data.

[0260] Step 5: Using a sliding pane method, calculate the time-domain features of the dimensionality-reduced time-series data sequentially to form a time-domain feature judgment set for refrigeration equipment failure, i.e., the target time-domain feature set.

[0261] Phase Two: Fault Detection and Judgment

[0262] Step 1: Obtain the timing data of the computer room cooling equipment, including: rack measuring point temperature, water valve opening, air conditioner return air temperature, air conditioner fan speed, air conditioner set fan speed, air conditioner minimum set fan speed, etc.

[0263] Step 2: Perform feature normalization on the time-series data of the computer room.

[0264] Step 3: Inject the time series data into the CNN+ConvLSTM fault detection model. When the feature map difference between the reconstructed target time series related feature map and the original target time series related feature map is greater than the preset feature map difference threshold, mark the time series data corresponding to this point as a data anomaly point and record the abnormal time to obtain the abnormal time period.

[0265] Step 4: Through kernel PCA principal component analysis, the dimensionality of the refrigeration equipment status data during the abnormal time period is reduced to 1-dimensional data.

[0266] Step 5: Using a sliding pane approach, calculate the temporal characteristics of the dimensionality-reduced outlier data sequentially.

[0267] Step 6: Using the K-neighbor algorithm, the abnormal time-domain feature data is compared with the time-domain feature judgment set to determine whether there is a fault and the specific fault type.

[0268] The fault detection model training method provided in this embodiment combines deep learning autoencoders with temporal analysis. A convolutional-long short-term memory (LSTM) network autoencoder enables real-time monitoring and fault pre-detection of data center air conditioning status. Kernel-PCA combined with temporal analysis achieves accurate determination of equipment faults and specific fault types. Furthermore, by combining kernel-PCA with temporal analysis, a temporal feature set for air conditioning faults is formed, capturing the data pattern state under various air conditioning faults. This method can also generally solve the problem of high similarity between fault and non-fault modes in fault diagnosis. The fault detection model training method provided in this embodiment can perform fault detection through deep learning autoencoders, avoiding data labeling work during model training, reducing the workload of model training, and simultaneously achieving real-time monitoring and pre-detection of air conditioning equipment.

[0269] The fault detection model training method provided in this embodiment enables fault detection and identification for data center air conditioning systems, proposing a fault diagnosis method that combines fault detection and fault identification. It includes a fault detection module based on a convolutional-long short-term memory network autoencoder, realizing real-time monitoring of data center air conditioning status and fault pre-detection. Simultaneously, a fault identification module based on kernel-PCA combined with time-domain analysis is designed to further analyze the pre-detection results, achieving accurate determination of equipment faults and specific fault types. This method also addresses the problem of low accuracy in diagnostic models due to a limited number of fault examples in actual diagnosis, improving the practicality and accuracy of air conditioning fault diagnosis methods. Furthermore, the autoencoder strategy avoids the need for extensive manual data annotation during fault detection model training. It also solves the problem in actual air conditioning fault diagnosis where noise and other factors cause indistinguishable fault data features from some normal data features, making model training difficult and prone to false alarms.

[0270] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0271] Based on the same inventive concept, this application also provides a training device for a fault detection model to implement the training method for the fault detection model described above. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more training device embodiments for fault detection models provided below can be found in the limitations of the training method for fault detection models described above, and will not be repeated here.

[0272] In one embodiment, such as Figure 15 As shown, a training device 1500 for a fault detection model is provided, comprising: a first acquisition module 1502, a first determination module 1504, an encoding / decoding module 1506, and a training module 1508, wherein:

[0273] The first acquisition module 1502 is used to acquire time-series data of the target device and determine the time-series characteristics corresponding to the time-series data. The time-series data includes at least the first sample data of the normal category, and the first sample data includes multiple operating status indicator data at multiple time points.

[0274] The first determining module 1504 is used to perform time-series related feature processing on the time-series features corresponding to multiple operating status indicator data at multiple time points based on a preset time window, so as to obtain a target time-series related feature map.

[0275] The encoding / decoding module 1506 is used to encode and decode the target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map.

[0276] Training module 1508 is used to train the fault detection model to be trained based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, so as to obtain the trained target fault detection model.

[0277] Each module in the training device for the aforementioned fault detection model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0278] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 16As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data related to the target device and fault detection models. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a training method for a fault detection model.

[0279] Those skilled in the art will understand that Figure 16 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0280] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0281] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0282] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0283] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0284] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0285] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0286] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A training method for a fault detection model, characterized in that, The method includes: Acquire time-series data of the target device and determine the time-series features corresponding to the time-series data. The time-series data includes at least first sample data of the normal category. The first sample data includes multiple operating status index data of the target device at multiple time points in the normal operating state. Based on a preset time window, multiple target time-series features are extracted from the time-series features corresponding to multiple operational status indicator data at multiple time points to obtain multiple target time-series features, and the correlation features between each target time-series feature are calculated; based on the correlation features between each target time-series feature, a target time-series correlation feature map is determined. The target temporal correlation feature map is encoded and decoded to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map; Based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, the fault detection model to be trained is trained to obtain the trained target fault detection model. The step of encoding and decoding the target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map includes: The target temporal correlation feature map is convolved by a first convolutional layer to obtain a first-dimensional target temporal correlation feature map; the first-dimensional target temporal correlation feature map is convolved by a second convolutional layer to obtain a second-dimensional target temporal correlation feature map; and the second-dimensional target temporal correlation feature map is convolved by a third convolutional layer to obtain a third-dimensional target temporal correlation feature map. Temporal analysis is performed on the target temporal correlation feature map of the first dimension based on the first long short-term memory subnetwork to obtain first temporal pattern information; temporal analysis is performed on the target temporal correlation feature map of the second dimension based on the second long short-term memory subnetwork to obtain second temporal pattern information; temporal analysis is performed on the target temporal correlation feature map of the third dimension based on the third long short-term memory subnetwork to obtain third temporal pattern information. The first deconvolutional layer performs feature reconstruction processing on the superposition result of the first temporal pattern information and the first-dimensional target temporal correlation feature map to obtain the first reconstructed temporal correlation feature map corresponding to the first-dimensional target temporal correlation feature map. The second deconvolutional layer performs feature reconstruction processing on the superposition result of the first reconstructed temporal correlation feature map, the second temporal pattern information, and the second-dimensional target temporal correlation feature map to obtain the second reconstructed temporal correlation feature map corresponding to the second-dimensional target temporal correlation feature map. The third deconvolutional layer performs feature reconstruction processing on the superposition result of the second reconstructed temporal correlation feature map, the third temporal pattern information, and the third-dimensional target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map.

2. The method according to claim 1, characterized in that, The time-series data also includes second sample data for fault categories; The step of training the fault detection model to be trained based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map to obtain a trained target fault detection model includes: Based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, the fault detection model to be trained is trained to obtain the trained initial fault detection model. Based on a preset analysis algorithm, the second sample data of the fault category is subjected to dimensionality reduction processing to obtain target fault sample data, and based on a preset time-domain feature algorithm, the target time-domain feature set corresponding to the target fault sample data is determined. Based on the target time-domain feature set and the initial fault detection model, the target fault detection model is determined.

3. The method according to claim 2, characterized in that, The method further includes: Acquire the timing data to be detected from the target device, and determine the timing features to be detected corresponding to the timing data to be detected; Based on the time series features to be detected, a time series correlation feature map to be detected is determined, and the time series correlation feature map to be detected is input into the target fault detection model to obtain the reconstructed time series correlation feature map to be detected; Based on the time-series correlation feature map to be detected and the reconstructed time-series correlation feature map to be detected, the fault detection result corresponding to the time-series data to be detected of the target device is determined.

4. The method according to claim 3, characterized in that, The step of determining the fault detection result corresponding to the time series data to be detected of the target device based on the time series correlation feature map to be detected and the reconstructed time series correlation feature map to be detected includes: If the time-series correlation feature map to be detected and the reconstructed time-series correlation feature map to be detected do not meet the first similarity condition, the fault detection result corresponding to the time-series data to be detected of the target device is determined to be a fault result. If the timing-related feature map to be detected and the reconstructed timing-related feature map to be detected satisfy the first similarity condition, the fault detection result corresponding to the timing data to be detected of the target device is determined to be a non-fault result.

5. The method according to claim 4, characterized in that, In the case where the time-series correlation feature map to be detected is inconsistent with the reconstructed time-series correlation feature map to be detected, determining the fault detection result corresponding to the time-series data to be detected of the target device as a fault result includes: If the time-series correlation feature map to be detected is inconsistent with the reconstructed time-series correlation feature map to be detected, the time-series data to be detected of the target device is subjected to dimensionality reduction processing to obtain the target time-series data to be detected. Based on a preset time-domain feature algorithm, the time-domain features to be detected corresponding to the target time-series data to be detected are determined; If the time-domain feature to be detected and the target time-domain feature set satisfy the second similarity condition, then the fault detection result corresponding to the time-series data to be detected of the target device is determined to be a fault result.

6. The method according to claim 5, characterized in that, If the time-domain feature to be detected and the target time-domain feature set satisfy the second similarity condition, then the fault detection result corresponding to the time-series data to be detected of the target device is determined to be a fault result, including: Calculate the similarity between the time-domain feature to be detected and each target time-domain feature contained in the target time-domain feature set; If there is a target time-domain feature with a similarity greater than a preset similarity threshold, then the fault detection result corresponding to the time-series data to be detected of the target device is determined to be a fault result.

7. The method according to claim 6, characterized in that, The method further includes: Among the similarities greater than a preset similarity threshold, the largest similarity is determined, and the target temporal feature corresponding to the largest similarity is determined. Based on the preset correspondence between time-domain features and fault types, the target fault type corresponding to the target time-domain feature is determined, and the fault type corresponding to the time-series data to be detected of the target device is determined as the target fault type.

8. A training device for a fault detection model, characterized in that, The device includes: The first acquisition module is used to acquire time-series data of the target device and determine the time-series features corresponding to the time-series data. The time-series data includes at least first sample data of the normal category. The first sample data includes multiple operating status index data of the target device at multiple time points in the normal operating state. The first determining module is used to extract multiple target time-series features from the time-series features corresponding to multiple operating status indicator data at multiple time points based on a preset time window, and to calculate the correlation features between each target time-series feature; and to determine a target time-series correlation feature map based on the correlation features between each target time-series feature. The encoding / decoding module is used to encode and decode the target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map corresponding to the target temporal correlation feature map. The training module is used to train the fault detection model to be trained based on the target temporal correlation feature map and the target reconstructed temporal correlation feature map, so as to obtain the trained target fault detection model. The encoding / decoding module is specifically used to perform convolution processing on the target temporal correlation feature map through a first convolutional layer to obtain a first-dimensional target temporal correlation feature map; to perform convolution processing on the first-dimensional target temporal correlation feature map through a second convolutional layer to obtain a second-dimensional target temporal correlation feature map; and to perform convolution processing on the second-dimensional target temporal correlation feature map through a third convolutional layer to obtain a third-dimensional target temporal correlation feature map. Temporal analysis is performed on the target temporal correlation feature map of the first dimension based on the first long short-term memory subnetwork to obtain first temporal pattern information; temporal analysis is performed on the target temporal correlation feature map of the second dimension based on the second long short-term memory subnetwork to obtain second temporal pattern information; temporal analysis is performed on the target temporal correlation feature map of the third dimension based on the third long short-term memory subnetwork to obtain third temporal pattern information. The first deconvolutional layer performs feature reconstruction processing on the superposition result of the first temporal pattern information and the first-dimensional target temporal correlation feature map to obtain the first reconstructed temporal correlation feature map corresponding to the first-dimensional target temporal correlation feature map. The second deconvolutional layer performs feature reconstruction processing on the superposition result of the first reconstructed temporal correlation feature map, the second temporal pattern information, and the second-dimensional target temporal correlation feature map to obtain the second reconstructed temporal correlation feature map corresponding to the second-dimensional target temporal correlation feature map. The third deconvolutional layer performs feature reconstruction processing on the superposition result of the second reconstructed temporal correlation feature map, the third temporal pattern information, and the third-dimensional target temporal correlation feature map to obtain the target reconstructed temporal correlation feature map.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images