Equipment anomaly detection method based on semi-supervised multi-scale convolutional auto-encoder and generative adversarial network

Abstract

The invention belongs to but is not limited to the technical field of industrial equipment anomaly detection, and discloses an equipment anomaly detection method based on a semi-supervised multi-scale convolutional auto-encoder and a generative adversarial network, which comprises the following steps of: performing model training by using part of marked data and a large amount of unmarked data, performing deep feature extraction by using a multi-scale convolutional neural network, and performing feature extraction by using the multi-scale convolutional neural network; meanwhile, as an encoder, an auto-encoder and a generative adversarial network are fused, the ability of the encoder to learn data features is indirectly improved through adversarial training of a decoder and a discriminator, and then fine adjustment is performed on a newly added classification layer of the encoder by using limited label information, so that the ability of the encoder to extract input data features is improved; and the model can better understand the classification boundary of the data, so that more accurate anomaly detection is realized. The whole model training adopts a two-stage training strategy of unsupervised pre-training and supervised fine tuning in semi-supervised learning, and finally an encoder is used as a classifier for equipment anomaly detection.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the technical field of industrial equipment anomaly detection, and in particular relates to an equipment anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network. Background Art

[0002] Anomaly detection of industrial equipment is an important means to ensure system safety and prevent industrial accidents. It can reduce the risk of failure, improve equipment reliability, and provide a stable and safe operating environment for industrial production. At present, anomaly detection methods can be divided into three basic methods: supervised, unsupervised, and semi-supervised anomaly detection, depending on whether the data set is labeled. The label refers to the mark of whether the data is normal or not. However, equipment usually cannot "run with problems", resulting in very limited abnormal data that can be obtained, and most of the data is unlabeled, which brings huge challenges to the anomaly detection problem in practice.

[0003] Supervised anomaly detection methods require that the training data set be labeled with abnormal and normal labels in advance. Since most of the current anomaly detection work is online monitoring and the data labels are unknown, this type of anomaly detection method is rarely used directly. This method is often used for algorithm verification and evaluation of the performance of anomaly algorithms. In view of the high cost of obtaining a large amount of labeled data in practical applications, most of the existing anomaly detection models mainly turn their attention to unsupervised methods. Unsupervised anomaly detection methods effectively overcome the problem of difficulty in obtaining device abnormal data. However, in many practical situations, some labeled data and unlabeled data can be used at the same time, and a particularly small number of abnormal samples in the labeled data can be used. Unsupervised methods ignore this valuable information. Unlike unsupervised learning, the goal of semi-supervised learning is to use both labeled and unlabeled data and maximize the value of all available data. At present, deep generative models parameterized by neural networks have achieved state-of-the-art performance in the practice of many semi-supervised learning tasks.

[0004] CN117150402A discloses a method and model for detecting anomalies in power data based on a generative adversarial network. The invention uses an anomaly detection model obtained by training a generator and a discriminator based on a generative adversarial network to process power time series data, and continuously optimizes the data generation ability and anomaly identification ability of the generator and the discriminator through adversarial training, thereby improving the anomaly detection ability of the model. However, the method adopts an unsupervised anomaly detection method, ignoring some available labeled data information. In addition, the method realizes anomaly detection of time series data by calculating the anomaly score of each sample and comparing it with a set threshold. However, in the threshold-based detection method, there is uncertainty in the threshold selection. If the threshold is not selected properly, errors may occur in the detection results.

[0005] In view of the above analysis, the technical problems that urgently need to be solved in the prior art are: there is uncertainty in the selection of threshold values ​​and errors in the detection results. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention provides a device anomaly detection method and system based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network, which makes full use of a small amount of labeled data and combines a large amount of unlabeled data to maximize the value of all available data and avoid the uncertainty of threshold selection in unsupervised anomaly detection. At the same time, the feature extraction capability of the deep generative model parameterized by the neural network is enhanced, thereby improving the model anomaly detection performance.

[0007] The present invention is implemented as follows: a device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network comprises the following steps:

[0008] Step 1: Data collection: monitor the equipment operation status through a variety of sensors, collect the changes of various physical quantities (such as temperature, pressure, vibration, etc.) during the operation of the equipment, including normal data and abnormal data;

[0009] Step 2: Data preprocessing: normalizing the data, extracting preliminary features, and balancing and partitioning the data;

[0010] Step 3, model construction, which overall combines the framework of autoencoder and generative adversarial network, adopts multi-scale convolutional neural network, and constructs the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network;

[0011] Step 4: Model training. The input sample data X of the entire model consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X. The anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network is trained in two stages. The first stage is unsupervised pre-training and the second stage is supervised fine-tuning.

[0012] Step 5, model testing: After the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification to achieve abnormality detection and discrimination.

[0013] Furthermore, in step 3, SMAE-GAN consists of three main parts: encoder, decoder and discriminator, and the network structure adopts multi-scale convolution.

[0014] Furthermore, the encoder and discriminator have the same multi-scale convolutional structure and the same hyperparameter settings, and the structures between the encoder and decoder are also corresponding, except that the convolution operation in the encoder is replaced by the transposed convolution operation in the decoder.

[0015] Furthermore, the multi-scale convolutional neural network constructs multiple parallel convolution paths, each path uses convolution kernels of different sizes, and uses convolution kernels of different sizes to construct a multi-scale fusion model. By changing the range of the receptive field and combining different depths of the convolution layer, the neurons on the feature map have different receptive field ranges, and multiple convolution kernels and multi-layer convolutions can realize feature extraction of different scales and forms; each path has multiple groups of identical convolution layers and pooling layers to obtain features of the same scale, and finally the torch.cat() function is used to realize multi-scale feature fusion.

[0016] Furthermore, the encoder uses the extracted multi-scale features to reconstruct the original samples and identify anomalies through the decoder; the input of the discriminator includes real input samples and reconstructed samples; using the extracted multi-scale features, the discriminator uses the fully connected layer to distinguish the true from the false input data.

[0017] Furthermore, the first stage of model training is the unsupervised pre-training stage, with a large amount of unlabeled data un_X as input data; first, the parameters of the autoencoder AE and the discriminator D are initialized. The autoencoder AE is fixed, the discriminator is trained first, the discriminator loss is calculated, the gradient is calculated through back propagation and the discriminator parameters are updated; then the autoencoder AE is trained again, the autoencoder loss is calculated, the gradient is calculated through back propagation and the encoder and decoder parameters of the autoencoder are updated; the above steps are repeated until the discriminator loss is small enough and the autoencoder parameters are saved.

[0018] Furthermore, in the second stage of model training, a classification layer is added after the encoder part of the trained autoencoder AE, and then the autoencoder AE trained in the first stage is fine-tuned in a supervised manner. A small amount of labeled data su_X is used as input data, the autoencoder loss is calculated, and the encoder and decoder parameters of the autoencoder are updated by calculating the gradient through back propagation until the loss of the autoencoder is small enough and the parameters of the autoencoder are saved.

[0019] Another object of the present invention is to provide a device anomaly detection system based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network of the method, comprising:

[0020] The data acquisition module monitors the operating status of the equipment through a variety of sensors and collects changes in various physical quantities (such as temperature, pressure, vibration, etc.) during the operation of the equipment, including normal data and abnormal data;

[0021] Data preprocessing module, which normalizes the data, extracts preliminary features, and balances and divides the data;

[0022] The model building module combines the framework of autoencoder and generative adversarial network as a whole, adopts multi-scale convolutional neural network, and builds the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network;

[0023] Model training module, the entire model input sample data X consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X; the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network is trained in two stages, the first stage is unsupervised pre-training, and the second stage is supervised fine-tuning;

[0024] Model testing module, after the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification to achieve abnormal detection and discrimination.

[0025] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network.

[0026] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network.

[0027] Another object of the present invention is to provide an information data processing terminal, which includes the device anomaly detection system based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network.

[0028] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0029] First, in view of the technical problems existing in the above-mentioned prior art, some creative technical effects are brought about after solving the problems. The specific description is as follows:

[0030] (1) This method uses a semi-supervised anomaly detection method, combining supervised and unsupervised learning, avoiding the uncertainty of threshold selection in unsupervised anomaly detection, and making full use of a small amount of labeled data combined with a large amount of unlabeled data, giving full play to the value of all available data and improving the performance of the anomaly detection model;

[0031] (2) By constructing a multi-scale convolutional neural network module for deep feature extraction and fusing features of different scales, the model's ability to express sample features is enhanced.

[0032] (3) The model framework integrates the autoencoder and the generative adversarial network. Through adversarial training between the decoder of the autoencoder and the discriminator of the generative adversarial network, the decoder is pushed to generate data that is as realistic as possible, thereby indirectly improving the encoder's ability to learn data features and improving the model's anomaly detection performance.

[0033] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following two important aspects:

[0034] (1) The present invention has strong commercial prospects. By improving the accuracy of anomaly detection and data utilization, it can be widely used in financial fraud detection, industrial equipment fault detection, network security monitoring and other fields. Compared with traditional methods, the present invention significantly improves the stability and robustness of detection while reducing the need for manual data labeling, which can reduce the cost of enterprises in data labeling and abnormal event processing, and further promote the intelligent development of data-intensive industries. In addition, by strengthening the model's early warning capabilities for abnormal situations, it helps companies reduce possible economic losses and improve overall operating efficiency.

[0035] (2) The device anomaly detection method based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network proposed in the present invention has brought significant technical progress. Anomaly detection has always faced two major technical difficulties: "insufficient label data" and "unsupervised learning threshold setting selection". Traditional unsupervised methods rely too much on threshold selection. If the threshold is not selected properly, it may lead to errors in the detection results. The present invention introduces a small amount of label data to perform detection in a semi-supervised manner, maximizes the value of all available data, and avoids the uncertainty of threshold selection in unsupervised anomaly detection. At the same time, the fusion of autoencoders, multi-scale convolutional networks and generative adversarial networks has made the model achieve breakthroughs in obtaining sample diversity and accurately extracting sample features, effectively solving the long-standing contradiction between detection accuracy and data dependence. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of a device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network provided by an embodiment of the present invention;

[0037] Figure 2 It is a structural diagram of a semi-supervised multi-scale convolutional autoencoder and anomaly detection model of a generative adversarial network (SMAE-GAN) provided in an embodiment of the present invention;

[0038] Figure 3It is a structural diagram of a device anomaly detection system based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] like Figure 1 As shown, a device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network includes the following steps:

[0041] Step 1, data collection: monitor the operating status of the equipment through multiple sensors, and collect changes in various physical quantities (such as temperature, pressure, vibration, etc.) during the operation of the equipment, including normal data and abnormal data;

[0042] Step 2, data preprocessing: In order to improve the quality and availability of the collected data and the effectiveness of subsequent modeling, the data is normalized, preliminary features are extracted, and the data is balanced and divided;

[0043] Step 3: Build an anomaly detection model based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network: Figure 2 As shown in Figure 1, the model consists of three main parts: encoder, decoder and discriminator. The encoder and discriminator have the same multi-scale convolution structure and the same hyperparameter settings, and the structure between the encoder and decoder is also corresponding, except that the convolution operation in the encoder is replaced by the transposed convolution operation in the decoder.

[0044] Among them, the multi-scale convolutional neural network constructs three parallel convolution paths, each path uses convolution kernels of different sizes, and uses convolution kernels of different sizes to build a multi-scale fusion model. By changing the range of the receptive field and combining different depths of the convolution layer, the neurons on the feature map have different receptive field ranges, and multiple convolution kernels and multi-layer convolutions realize feature extraction of different scales and forms. The sizes of the convolution kernels are selected as 3×1, 5×1, and 7×1 respectively. Each path has three sets of identical convolution layers and pooling layers to obtain features of the same scale. Finally, the torch.cat() function is used to realize multi-scale feature fusion. The calculation formula is:

[0045] F c = torch.cat(f 3 (x),f 5 (x),f 7 (x)

[0046] Among them, f3 (x), f 5 (x), f 7 (x) represents the feature output extracted by convolution kernels of different sizes, F c Represents the extracted multi-scale features. Finally, the encoder uses the extracted multi-scale features F c The decoder is used to reconstruct the original samples and identify anomalies.

[0047] The input data of the discriminator includes real input samples and reconstructed samples. Using the extracted multi-scale features, the discriminator distinguishes true from false input data through a fully connected layer.

[0048] Step 4, model training: The entire model input sample data X consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X.

[0049] The model training process is divided into two stages:

[0050] The first stage is the unsupervised pre-training stage. A large amount of unlabeled data un_X is used as input data to enter the autoencoder, and the data un_X' with the same structure as the input data is reconstructed. The model learns the mapping from the input data to itself by minimizing the reconstruction error of the data, so that the encoder can obtain the implicit feature distribution of the input data. The optimization goal is as follows:

[0051]

[0052] Among them, un_X' is the reconstructed sample of un_X.

[0053] un_X and un_X' are used as the input of the discriminator. The output of the discriminator is the probability of the original data and the generated data (essentially a binary classifier). Its main function is to conduct adversarial training with the decoder to improve the quality of the decoder-generated data and enhance the feature extraction capability of the encoder. The optimization goal is:

[0054]

[0055]

[0056] Among them, D(un_X) and D(un_X') represent the results of the discriminator's judgment on the real sample and the fake sample, that is, the probability that the sample is "true"; D f (·) represents the intermediate feature layer of the discriminator, D f (un_X) and D f(un_X') represents the output of the intermediate feature layer of the discriminator on the real sample and the reconstructed sample respectively. Since the model focuses on learning a good feature representation of the real sample, rather than distinguishing the reconstructed sample from the real sample, the model borrows the idea of ​​feature matching and minimizes the difference between the output of the intermediate feature layer of the discriminator on the real sample and the reconstructed sample, thereby enhancing the encoder's potential representation ability for the real sample.

[0057] Taking the above objective functions into consideration, the discriminator and the autoencoder are trained independently. The discriminator is trained first. The decoder hopes that the discriminator will judge the reconstructed data as true, and the discriminator hopes to judge the reconstructed data as false. The entire network finally reaches Nash equilibrium in the confrontation, indicating that the network has been trained. The final objective loss function of the first stage of SMAE-GAN model training is divided into two parts:

[0058] Loss1 AE =L1+L3

[0059] Loss D =L2

[0060] In the second stage, a classification layer is added after the encoder part of the trained autoencoder, and then the autoencoder trained in the first stage is fine-tuned in a supervised manner. A small amount of labeled data su_X is used as the input data of the autoencoder, and its classification loss and reconstruction loss are calculated respectively. The optimization goal is:

[0061]

[0062] Where p(y|su_X,y<2) represents the probability value of the label 0 and 1 obtained by the encoder for su_X, and su_X' is the reconstructed sample of su_X.

[0063] Taking the above objective functions into consideration, the final objective loss function of the second stage of SMAE-GAN model training is:

[0064] Loss2 AE =αL4+βL5

[0065] Among them, α+β=1.

[0066] Step 5, model testing: After the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification.

[0067] like Figure 3 As shown, an embodiment of the present invention provides a device anomaly detection system based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network, including:

[0068] The data acquisition module monitors the operating status of the equipment through a variety of sensors and collects changes in various physical quantities (such as temperature, pressure, vibration, etc.) during the operation of the equipment, including normal data and abnormal data;

[0069] Data preprocessing module, which normalizes the data, extracts preliminary features, and balances and divides the data;

[0070] The model building module combines the framework of autoencoder and generative adversarial network as a whole, adopts multi-scale convolutional neural network, and builds the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network;

[0071] Model training module, the entire model input sample data X consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X; the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network is trained in two stages, the first stage is unsupervised pre-training, and the second stage is supervised fine-tuning;

[0072] Model testing module, after the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification to achieve abnormal detection and discrimination.

[0073] An embodiment of the present invention provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network.

[0074] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0075] The present invention provides a specific application of a device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network. In two application embodiments of the present invention, two typical objects are selected according to the differences in the equipment operation data collection methods: wind turbines (collecting low-frequency signals through the SCADA system) and large rotor systems (collecting high-frequency vibration signals through the CMS system) to verify the application of the proposed method, as follows:

[0076] Example 1: Wind turbine blade icing prediction

[0077] Step 1, data collection: Monitor the operating status of the equipment through multiple sensors, and collect changes in various physical quantities (such as temperature, pressure, vibration, etc.) during the operation of the equipment, including normal data and abnormal data; the details are as follows:

[0078] This data set is collected from the SCADA system of a domestic wind turbine. It records 28 features such as timestamp, wind speed, generator speed, grid-side active power, yaw position, etc. The sampling frequency is 7s / time. The data contains the status information of wind turbine blades under various operating conditions, which is used to predict blade icing.

[0079] Step 2: Data preprocessing: In order to improve the quality and availability of the collected data and the effectiveness of subsequent modeling, the data is normalized, preliminary features are extracted, and the data is balanced and divided as follows:

[0080] The data collected by this dataset contains 28 features, of which the date and group number features have no training significance, so these two features are deleted and the remaining 26 features are used as input features. In order to eliminate the dimensional differences between data of different dimensions, the data is normalized to the maximum and minimum:

[0081]

[0082] Where x=(x 1 ,x 2 ,…,x T ) is a timestamp sequence of length T formed by the sampled multidimensional time variable, x t represents the multidimensional data obtained at timestamp t, x t ∈R m , min(x) represents the minimum value of the original data x, and max(x) represents the maximum value of the original data x.

[0083] The total number of samples in the data set used in this embodiment is 374,147, of which the number of unlabeled samples in the training set is 150,000, the number of labeled samples is 10,000, and the remaining samples are used as the test set. Since this data set is an unbalanced data set, the ratio of normal samples to abnormal samples in the labeled samples in the training set is set to 3:1, and the ratio of normal samples to abnormal samples in the unlabeled samples is about 7:1.

[0084] Step 3, model construction, which overall combines the framework of autoencoder and generative adversarial network, adopts multi-scale convolutional neural network, and constructs the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network;

[0085] Step 4: Model training. The input sample data X of the entire model consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X. The anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network is trained in two stages. The first stage is unsupervised pre-training and the second stage is supervised fine-tuning.

[0086] Step 5, model testing: After the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification to achieve abnormality detection and discrimination.

[0087] Example 2: Prediction of rotor component fall-off failure

[0088] Step 1, data collection: Monitor the operating status of the equipment through multiple sensors, and collect changes in various physical quantities (such as temperature, pressure, vibration, etc.) during the operation of the equipment, including normal data and abnormal data; the details are as follows:

[0089] The data set is collected by 6 eddy current displacement sensors and contains radial vibration and axial vibration information of the rotor system. The data is collected from six months before the rotor failure to the time of the rotor failure. The data is collected at equal angles, with 32 points sampled per rotation and 32 rotations in total, so each sample contains 1024 sampling points. The data is used to detect and predict the shedding failure of rotor components.

[0090] Step 2: Data preprocessing: In order to improve the quality and availability of the collected data and the effectiveness of subsequent modeling, the data is normalized, preliminary features are extracted, and the data is balanced and divided as follows:

[0091] The data set is collected by 6 eddy current displacement sensors and contains radial vibration and axial vibration information of the rotor system. The data is collected from six months before the rotor failure to the time of the rotor failure. The data is collected at equal angles, with 32 points sampled per rotation and 32 rotations in total, so each sample contains 1024 sampling points. The data is used to detect and predict the shedding failure of rotor components.

[0092] The faulty unit collected data in 7 stages, which can be divided into two stages, namely the normal operation stage and the abnormal stage. The speed fluctuation of the original data is large, and different speeds are bound to change the signal characteristics. Therefore, the data is preprocessed and the data at a stable speed is intercepted for analysis. According to the time domain waveform analysis of the data, it is found that the two-stage data collected by the two shaft displacement sensors A and B have no distinction, so this data only uses the data signals collected by the sensors installed near the joint end and the non-joint end, a total of 4 signals. This data set belongs to vibration signals, and the sampling frequency of vibration signals is usually very high. Each sample in the two stages contains 1024 points. Vibration signals usually have rich time domain information, and time domain features such as kurtosis, root mean square value, peak-to-peak value, etc. can usually intuitively reflect the vibration characteristics and physical state of the signal. These features have clear physical meanings in the engineering field, and using these features in machine learning models can improve the efficiency and performance of the model. Therefore, this experiment uses a sliding window size of 1024 to extract the time domain features of each sample as the final input data. The time domain features extracted this time are: kurtosis, maximum value, root mean square value, peak-to-peak value, peak index, and margin index. Six features are extracted from each dimensional signal, and the final input sample dimension is 24 dimensions.

[0093] The total number of samples in the data set used in this embodiment is 1865, of which 572 are normal operation samples and 1293 are abnormal stage samples. Since abnormal samples are difficult to obtain in practical applications, the ratio of normal samples to abnormal samples in the labeled samples in the training set is set to 10:3, and the ratio of normal samples to abnormal samples in the unlabeled samples is set to 4:1. The number of labeled samples is 13, the number of unlabeled samples is 500, and the remaining samples are used as test sets.

[0094] Step 3, model construction, which overall combines the framework of autoencoder and generative adversarial network, adopts multi-scale convolutional neural network, and constructs the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network;

[0095] Step 4: Model training. The input sample data X of the entire model consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X. The anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network is trained in two stages. The first stage is unsupervised pre-training and the second stage is supervised fine-tuning.

[0096] Step 5, model testing: After the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification to achieve abnormality detection and discrimination.

[0097] In order to illustrate the superiority of the proposed method in the absence of labeled samples, the experiment selected commonly used supervised learning algorithms: Convolutional Neural Network (CNN), Support Vector Machine (SVM), Random Forest (RF) and semi-supervised learning algorithms: Semi-Supervised Generative Adversarial Network (SGAN), Autoencoder (AE) for comparative experiments, using recall rate R, precision rate P, F 1 The value is used as the model evaluation index, and the final results are shown in Table 1.

[0098] Table 1 Evaluation indicators of different models on test samples of embodiments of the present invention

[0099]

[0100] The anomaly detection model based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network (SMAE-GAN) constructed by the present invention can make full use of the information provided by a large amount of unlabeled data to improve the accuracy of model anomaly detection compared with pure supervised learning models CNN, SVM, and RF. Compared with the common semi-supervised learning models SGAN and AE, the method proposed by the present invention integrates AE and GAN, in which the decoder and the discriminator are trained adversarially, and the autoencoder can be trained by the reconstruction error of the unlabeled data, which not only improves the utilization efficiency of the data, but also enables the model to better learn the data distribution through this joint training method, thereby improving the ability of the encoder to learn data features. The verification is carried out on industrial low-frequency data sets and high-frequency data sets respectively, and an F1 score of more than 90% can be obtained even with a small number of labeled samples, which is more in line with the data scene requirements of actual industrial sites.

[0101] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network, characterized in that: The steps include: Step 1, data collection, monitoring the equipment operation status through a variety of sensors, collecting changes in various physical quantities during the operation of the equipment, the physical quantities including temperature, pressure, vibration, including normal data and abnormal data; Step 2: Data preprocessing: normalizing the data, extracting preliminary features, and balancing and partitioning the data; Step 3, model construction, which overall combines the framework of autoencoder and generative adversarial network, adopts multi-scale convolutional neural network, and constructs the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network; Step 4: Model training. The input sample data X of the entire model consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X. The anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network is trained in two stages. The first stage is unsupervised pre-training and the second stage is supervised fine-tuning. Step 5, model testing: After the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification to achieve abnormality detection and discrimination.

2. The device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network as claimed in claim 1, characterized in that: In step 3, SMAE-GAN consists of three main parts: encoder, decoder and discriminator, and the network structure adopts multi-scale convolution; the multi-scale convolution structure of the encoder and the discriminator is the same, the hyperparameter settings are consistent, and the structure between the encoder and the decoder is also corresponding, except that the convolution operation in the encoder is replaced by the transposed convolution operation in the decoder.

3. The device anomaly detection method based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network as claimed in claim 2, characterized in that: The multi-scale convolutional neural network constructs multiple parallel convolution paths, each path uses convolution kernels of different sizes, and uses convolution kernels of different sizes to build a multi-scale fusion model. By changing the range of the receptive field and combining different depths of the convolution layer, the neurons on the feature map have different receptive field ranges. Multiple convolution kernels and multi-layer convolutions can realize feature extraction of different scales and forms; each path has multiple sets of identical convolution layers and pooling layers to obtain features of the same scale, and finally the torch.cat() function is used to realize multi-scale feature fusion.

4. The device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network as claimed in claim 3, characterized in that: The encoder uses the extracted multi-scale features to reconstruct the original samples and identify anomalies through the decoder; the input of the discriminator includes real input samples and reconstructed samples; using the extracted multi-scale features, the discriminator uses the fully connected layer to distinguish the true from the false input data.

5. The device anomaly detection method based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network as claimed in claim 1, characterized in that: The first stage of model training is the unsupervised pre-training stage, with a large amount of unlabeled data un_X as input data; first initialize the parameters of the autoencoder AE and the discriminator D; fix the autoencoder AE, train the discriminator first, calculate the discriminator loss, calculate the gradient through back propagation and update the discriminator parameters; then train the autoencoder AE again, calculate the autoencoder loss, calculate the gradient through back propagation and update the encoder and decoder parameters of the autoencoder; repeat the above steps until the loss of the discriminator is small enough and save the parameters of the autoencoder.

6. The device anomaly detection method based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network as claimed in claim 1, characterized in that: In the second stage of model training, a classification layer is added after the encoder part of the trained autoencoder AE, and then the autoencoder AE trained in the first stage is fine-tuned in a supervised manner. A small amount of labeled data su_X is used as input data, the autoencoder loss is calculated, and the encoder and decoder parameters of the autoencoder are updated by calculating the gradient through back propagation until the loss of the autoencoder is small enough and the parameters of the autoencoder are saved.

7. A device anomaly detection system based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network according to any one of claims 1 to 6, characterized in that: include: The data acquisition module monitors the operation status of the equipment through a variety of sensors and collects changes in various physical quantities during the operation of the equipment, including temperature, pressure, and vibration, including normal data and abnormal data; Data preprocessing module, which normalizes the data, extracts preliminary features, and balances and divides the data; The model building module combines the framework of autoencoder and generative adversarial network as a whole, adopts multi-scale convolutional neural network, and builds the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network; Model training module, the entire model input sample data X consists of a large amount of unlabeled data un_X and a small amount of labeled data su_X; the anomaly detection model SMAE-GAN based on semi-supervised multi-scale convolutional autoencoder and generative adversarial network is trained in two stages, the first stage is unsupervised pre-training, and the second stage is supervised fine-tuning; Model testing module, after the SMAE-GAN model training is completed, the network parameters of the encoder are fixed, and the test set data is input into the encoder network for normal and abnormal binary classification to achieve abnormal detection and discrimination.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the device anomaly detection method based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal includes the device anomaly detection system based on a semi-supervised multi-scale convolutional autoencoder and a generative adversarial network as described in claim 7.

Images