Process anomaly detection method based on time sequence shielding auto-encoder

Abstract

The invention discloses a process anomaly detection method based on a time sequence shielding auto-encoder, and belongs to the technical field of anomaly detection in the process industry. The method comprises the following steps of: (1) standardizing data and extracting a data continuous subset as a new sample by using a sliding time window mode; (2) random sampling is adopted to input variables, and mask operation is carried out on data; (3) extracting and reconstructing hidden features of process data by adopting an automatic encoder which takes LSTM (Long Short Term Memory) into which a learnable vector is introduced as a core; (4) model training; (5) positioning abnormal features and positions; and (6) performing a process anomaly detection process by using TSMAE. According to the method, most input variables can be randomly shielded, information redundancy can be reduced, extraction of hidden features and reconstruction of original data can be realized by virtue of time correlation of LSTM modeling process data, equivalent learnable vectors are filled in the hidden features, and the global information capability of model modeling is enhanced by taking overall reconstruction loss as an optimization target.

Description

Technical Field

[0001] The present invention belongs to the field of anomaly detection in process industry, and in particular relates to a process anomaly detection method based on a temporal masking autoencoder. Background Art

[0002] As an important part of process industry automatic control, anomaly detection is responsible for monitoring and identifying abnormal conditions of system operation so that operators can make timely corrections to ensure system safety. With the introduction of efficient and low-cost data transmission and storage solutions, process industries have accumulated a large amount of time series data during operation. The data contains a large amount of process status information, and it is of great significance to detect anomalies and reduce the losses caused by system failures.

[0003] In recent years, real-time anomaly detection in industrial processes is generally divided into two categories: model-driven methods and data-driven methods. The former requires modeling industrial processes based on prior knowledge and setting fixed thresholds to detect anomalies. The latter only extracts valid information from historical data to detect anomalies. Therefore, data-driven anomaly detection has become an active research field.

[0004] In the data collected by the process industry, multiple process variables evolve with each other to maintain the same behavior. Within a certain period of time, these variables interact with each other within the equipment and between equipment, with obvious variable correlation and time correlation. In order to extract correlation information, more and more deep learning methods are cited in the field of anomaly detection.

[0005] However, the direct application of deep learning methods to anomaly detection in process industries still faces a major challenge, namely the inherent information redundancy of process data. To ensure the stability of process industries, multiple sensors with similar functions are used to collect a set of process data, thereby mitigating the potential risks caused by sensor failures. These collected variables show strong correlations, resulting in information redundancy, which has an adverse effect on the extraction of potential features in the reconstruction model: the reconstruction model ignores the mapping relationship and global information of weakly correlated variables, reducing the anomaly detection performance. In addition, information redundancy will bring additional computational overhead and reduce real-time performance. Summary of the invention

[0006] Aiming at the negative impact of inherent information redundancy in industrial process data acquisition on the performance of anomaly detection algorithms, the present invention proposes a process anomaly detection method based on temporal masking autoencoder, which can effectively reduce the impact of process information redundancy on detection.

[0007] The specific technical solutions are as follows:

[0008] A process anomaly detection method based on a time series masked autoencoder, the method comprising the following steps:

[0009] (1) Data preprocessing

[0010] To ensure the consistency of the collected process data, the data needs to be standardized and converted to a mean of 0 and a standard deviation of 1 so that the variables can be arranged on a common scale. The standardized data is used to extract continuous subsets of the data as new samples using a sliding time window.

[0011] (2) Timing masking and data reconstruction

[0012] Before the raw data is input into the encoder, the input variables are randomly sampled and masked. The raw data is masked by setting a higher masking ratio α to ensure that the influence of information redundancy is weakened or eliminated. This strategy helps the model focus more on the intrinsic features of the modeled data rather than redundant patterns, thereby improving the accuracy of anomaly detection. In addition, by reducing the input features of the encoder, the feature extraction speed can be effectively improved, which provides a guarantee for the real-time requirements of anomaly detection tasks.

[0013] In order to model the temporal correlation in process data and improve the accuracy of process anomaly detection, an autoencoder with LSTM as the core is used to extract the hidden features of process data and restore and reconstruct them. However, using LSTM alone cannot model global information and variable correlation, so a learnable vector is introduced in the autoencoder.

[0014] (3) Model training and anomaly location

[0015] The sigmoid activation function is added to the encoder's LSTM network. Before starting model training, specific hyperparameters must be predefined, including the number of iterations, batch size, and learning rate. The training process is performed in batches, and the Adaptive Moment Estimation (Adam) optimizer is used to find the optimal network parameters. A complete iteration involves training all samples once. Through multiple rounds of iterative optimization, the model achieves reconstruction of the input data with minimal error.

[0016] The Lublin Sugar Factory Dataset (Damadics Benchmark Data Set, DMDS) calculates the difference between the reconstructed output and the original input, uses the 3sigma method to filter outliers and displays them in a visual way. The 3sigma method uses statistical concepts related to standard deviation to filter outliers or outliers.

[0017] Furthermore, the process of step (1) is:

[0018] Step 1.1: Data standardization. Treat the collected process data as a two-dimensional array Where F represents the number of features and T represents the number of data collected. For example, the specific value of F in the evaporation station of the Lublin Sugar Factory is 33, and 86,400 data can be collected in one day. The specific process of standardization is as follows:

[0019]

[0020] Among them, x i and x i Respectively represent the process characteristics before and after standardization, μ i and σ i They represent the mean and variance of the i-th feature respectively. The standardized data is still represented by X.

[0021] Step 1.2: Since TSMAE uses LSTM as the computing core, the temporal continuity of the input model needs to be considered, so a sliding time window extraction is required for the original data X. The new data set extracted using the sliding time window will be regarded as a three-dimensional array Where N is the number of samples in the new data set, W is the time window of fixed size that is systematically moved in the data set, and S is the moving distance of the time window of fixed size. The relationship between N and T, W and S is as follows:

[0022]

[0023] Furthermore, the process of step (2) is:

[0024] Step 2.1: Specifically, define X as the original data and the masking rate α.

[0025] Step 2.2: Randomly sample F×α variables and remove them from the input to generate the encoder input The following quantitative relationship is met:

[0026] L=(1-α)×F (3)

[0027] Among them, L is the number of variables remaining in the end.

[0028] Step 2.3: Extract two-dimensional arrays from the original data to form a data set C = in Define the hidden feature dimension extracted by the LSTM encoder as D, and the visible input C n The hidden features of This is equivalent to mapping each process variable into a D / L dimensional latent matrix. Such a design helps to better capture the different time patterns of each variable, thereby improving the model's ability to model the temporal correlation of process data.

[0029] Step 2.4: The added learnable vector is represented as Its scale is determined by the relationship between the masking rate α, the hidden feature dimension D, and the overall number of variables F. The added learnable vectors help TSMAE model the variable correlation and global information of the process data and ensure that the number of features can be kept consistent during the reconstruction phase.

[0030] Step 2.5: The AutoEncoder (AE) structure is used in this reconstruction. This method uses a two-layer neural network structure, where both the encoder and the decoder are composed of an LSTM. The difference is that the encoder is responsible for extracting the hidden feature H from the unmasked data C. A learnable vector is added to the hidden feature corresponding to the masked feature position, and the decoder reconstructs the original input through the hidden feature and the learnable vector. The visible variable in the time window is defined as Where M represents the number of visible variables. During the training phase, the overall reconstruction deviation is used as the optimization target, and the formula is as follows:

[0031]

[0032] Among them, Y represents the reconstructed output of TSMAE, E(.) and D(.) are the encoder and decoder respectively, and concat(.) represents the concatenation function, where H and G need to be concatenated in the variable dimension.

[0033] Furthermore, the process of step (3) is:

[0034] Step 3.1: Extract training and test samples from the original data using the sliding time window method.

[0035] Step 3.2: In the training phase, the normal time series data obtained after standardization and preprocessing is used as the normal sample training model, and the model parameters are updated according to the reconstruction error of the normal data until the preset training conditions are met.

[0036] Step 3.3: Use the trained TSMAE to evaluate the reconstruction loss of the test sample, and use the 3sigma method to identify abnormal features and corresponding time, and compare the performance indicators of TSMAE and the baseline model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a schematic diagram of the structure of the TSMAE of the present invention;

[0038] Figure 2 It is TSMAE’s process industry anomaly detection full process flow chart;

[0039] Figure 3 It is the heat map of Pearson correlation coefficient of training set variables;

[0040] Figure 4 is the reconstruction loss corresponding to different D on the training set;

[0041] Figure 5 is the visualization result of each method at the first flowmeter anomaly;

[0042] Figure 6 It is the visualization result of each method in the second flowmeter anomaly;

[0043] Figure 7 is the visualization result of each method in the third flowmeter anomaly;

[0044] Figure 8 This is the reconstruction of the 16th variable, where the light grey area indicates the area where the method misjudged;

[0045] Fig. 9 This is the reconstruction of the 28th variable.

[0046] The beneficial effects are as follows:

[0047] The present invention can randomly shield most input variables and reduce information redundancy. It can extract hidden features and reconstruct original data by using the time correlation of LSTM modeling process data, fill equal amounts of learnable vectors in hidden features, and take overall reconstruction loss as optimization target to enhance the global information capability of model modeling. Compared with multiple baseline models, it is further verified that the method improves the anomaly detection inference speed and anomaly detection accuracy, and has universality and versatility. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings.

[0049] Reference Figure 1-Figure 9 , a process anomaly detection method based on a time series masked autoencoder, the method comprising the following steps:

[0050] (1) Data preprocessing

[0051] This example uses real data collected from the evaporation station of the Lublin Sugar Factory to verify its controller as the data set DMDS. The following table shows the basic information of the data set.

[0052] Table 1 Introduction to DMDS dataset

[0053]

[0054] This data set records three groups of controllers and controlled devices. The collected data consists of the control variables of the controllers and the basic information of the controlled devices. Due to the physical relationship between the variables in each set or being in the same device, there is a certain degree of correlation between them. Figure 3The Pearson correlation coefficients between the variables in the training set of the DMDS dataset are shown. Obviously, there is a strong correlation between some variables, indicating that there is information redundancy in the dataset.

[0055] Step 1.1: Data standardization.

[0056] DMDS uses artificial interference equipment to accurately record the equipment, time, and related variables that impose abnormalities. The abnormal data is used for verification and testing, and divided in a 1:1 manner, and the data from the day before the interference is used for training. The collected process data is regarded as a two-dimensional array Where F represents the number of features and T represents the amount of data collected. Due to the different sizes of the collected variables, standardization is used to normalize the original data. The standardization calculation formula used is as follows

[0057]

[0058] in, represents the fth variable vector in the original process data, μ f and σ f ∈ represents x f The mean and variance of represents the standardized f-th variable vector and constitutes the process data P.

[0059] Step 1.2: Perform sliding time window extraction on the original data X.

[0060] From the process data P, a sliding window is used to extract the model input data. Specifically, the length of the time window L and the sliding step S are fixed, and two-dimensional arrays are extracted from P in turn to form a data set. in c n,l represents all variable data collected at time l in the nth time window in the dataset, N represents the number of time window samples extracted from P, and T represents the number of collection moments. The relationship among T, S, N and L is as follows:

[0061]

[0062] (2) Timing masking and data reconstruction

[0063] Step 2.1: Perform feature masking on the data.

[0064] Specifically, define X n For the original data, the masking rate is α, F×α variables are randomly sampled and removed from the input, and finally M variables remain to generate the encoder input C. Among them, M meets the following quantitative relationship:

[0065] M=(1-α)F (7)

[0066] Step 2.2: Define the LSTM based encoder.

[0067] The time window variable is defined as Where L represents the length of the fixed time window and M represents the number of visible variables. The encoder is responsible for extracting hidden features from the unmasked data S D represents the dimension of hidden features extracted by LSTM. LSTM filters and retains input, state, and memory at each moment, thus retaining contextual information for a long time.

[0068] H e =LSTM E (S) (8)

[0069] Among them, LSTM E (.) represents the LSTM function corresponding to the encoder. The calculation process of the LSTM is as follows:

[0070]

[0071] in, They are the weight and bias parameters of the forget gate used to calculate the retained memory and are the weight and bias parameters of the input gate, respectively, which determine which information in the memory cell is updated at the current time, denoted as and are the weights and bias parameters of the candidate gates, respectively, used to calculate the candidate memory and They are the weight parameters and bias parameters of the output gate, which are used to calculate the output vector at the current moment. Represents the memory vector output at the current moment. Sigmoid(.) and Tanh(.) are both activation functions. In addition, fill the learnable vector in the hidden state Where Z represents the number of learnable vectors, and the specific value is calculated by the following formula:

[0072]

[0073] Step 2.3: Replace the masked variables with learnable vectors to reduce information redundancy while enabling AE to model the mapping relationship of weakly correlated variables and ensure the consistency of the number of variables in the reconstruction phase. Note that the learning vector is filled in the position of the masked variable.

[0074] The encoder composed of LSTM is derived from H eThe original data is reconstructed from the set of K and K, and the overall reconstruction error loss is used as the optimization target to constrain the mapping of the learnable vector to the masked variable. The specific process is as follows:

[0075]

[0076] in, Represents the reconstruction output of TSMAE, LSTM D (.) represents the LSTM function corresponding to the decoder. The specific calculation is consistent with Formula 9. concat(.) represents the concatenation function. Here, H e and K are concatenated in the variable dimension.

[0077] Step 2.4: TSMAE parameter update rules.

[0078] In TSMAE, the hidden feature H e The learnable vector K is concatenated in order according to the variable dimension to form the input of the LSTM decoder. LSTM adopts the strategy of sharing weights and biases in the parameter update process, and finally updates the parameters at each moment. Take the parameter update gradient at moment l as an example to show:

[0079]

[0080] in, and They represent the state vector at time l and the output vector at time l+1 in the decoder LSTM respectively. represents the trainable parameter matrix of the hidden layer in LSTM, and They represent the weight and bias of the LSTM output gate in the decoder respectively. Similarly, the weight update gradients of other function gates of the decoder can be obtained.

[0081] The learnable vector K also needs to be updated in time with the back propagation of the gradient. It should be noted that in the process of calculating the updated gradient corresponding to K, H needs to be calculated. e And the overall gradient update vector of K:

[0082]

[0083] in, and They represent the forget vector at time l, the state vector at time l, the update vector at time l, the weights of the forget gate, the memory gate, and the input gate of the LSTM in the decoder, respectively. e,l Represents the state parameter vector of the previous moment. The weight update gradient of the encoder is as follows:

[0084]

[0085] in, Indicates o d,l is only related to visible inputs l The relevant part vector, represents the trainable parameter matrix of the hidden layer in LSTM, and They represent the weight and bias of the LSTM output gate in the encoder respectively.

[0086] (3) Model training and anomaly location

[0087] Step 3.1: Perform model training.

[0088] The hyperparameters used in model training are as follows: The length of the sliding time window L and the sliding step S are set to 50 and 50 respectively. The mask rate α is set to 0.5. The LSTM of the encoder and decoder simply use the same hyperparameters, and the dimension D of the LSTM to extract hidden features is selected based on the fitting performance on the training set. Since D must be an integer and is related to α and F, the selection range is limited to D∈{16,32,64,128,256}. The performance of TSMAE with different D on the training set is compared as shown in the figure. Figure 4 As shown, D = 128 is finally determined. The proposed method is compared with advanced reconstruction anomaly detection models, including Transformer-based Anomaly Detection (TransAD), LSTM-based Variational Autoencoder (LSTM-VAE), Long Short-Term Memory Autoencoder (LSTMAE), Variational Autoencoder (VAE) and Deep Autoencoder (DAE).

[0089] Step 3.2: Locate abnormal features.

[0090] For the DMDS data set, the difference between the reconstructed output and the original input is calculated, and the 3sigma method is used to filter out the abnormal points and display them in a visual way. The 3sigma method uses statistical concepts related to standard deviation to filter out abnormal or outliers.

[0091] According to the 3sigma rule, 68% of the overall data will be within one standard deviation of the mean, 95% will be within two standard deviations of the mean, and nearly 99.7% will be within three standard deviations of the mean. It is usually used to evaluate the probability of observations within a specified range and determine whether data outside the range is abnormal or outlier. Therefore, a hyperparameter σ (usually 3) is set to count the 3sigma upper and lower bounds of each feature in the time dimension, and mark the excess as abnormal.

[0092] Step 3.3: Anomaly detection performance evaluation.

[0093] The adopted method is compared with advanced reconstruction anomaly detection models, including TransAD based on Self-Attention Mechanism (SAM), LSTMAE without masking mechanism, LSTM-VAE, VAE and DAE.

[0094] In order to verify the anomaly detection effect of different methods, accuracy, recall rate and F1 score are used as evaluation indicators, and the reasoning time of different methods is recorded. The anomaly detection ability of different methods is analyzed from the perspective of performance and efficiency. The specific performance is shown in Table 2 below.

[0095] Table 2. Comparison of TSMAE and baseline model performance

[0096]

[0097] The performance index calculation formula in the table is as follows:

[0098]

[0099] Among them, TP is the number of true positive moments, which is the moment when anomalies actually occur and are detected as anomalies; FP is the number of false positive moments, which is the moment when no anomalies actually occur but are detected as anomalies; FN represents the moment when anomalies actually occur but are detected as normal. Precision is the precision rate, which indicates the proportion of actual abnormal samples among the samples determined by the detector to be abnormal. Recall is the recall rate, which indicates the proportion of samples that actually occur abnormally but are determined as abnormal by the detector. F1 is the harmonic mean of Precision and Recall, which is usually used to evaluate the performance of anomaly detection models. F1∈[0,1], and the closer to 1, the better the effect.

[0100] In the DMDS data set, researchers manually interfered with the bypass valve on November 17, 2001, causing an unexpected pressure drop in the evaporator, resulting in abnormal fluid flow and related variables. After a period of interference, the equipment was manually restored, and all variable values ​​returned to normal. The above interference occurred three times, lasting 35s, 38s, and 42s respectively. The abnormality identification of each method is shown in the figure below. Figure 5-Figure 7 As shown in the figure. The dark gray box area indicates the time and variable of the actual abnormality, and the light gray area is the abnormal area detected by the model. It can be clearly seen from the figure that the four models can detect the area where the real abnormality is located, but the baseline model will identify more normal areas as abnormal areas, which directly leads to low recall rate. TSMAE combines the mask mechanism with LSTM to model the global information of normal data and obtain the best effect. Although TransAD uses SAM to model the time and variable correlation of data, the redundancy of process data leads to overall misjudgment of other variables, reducing the F1 score. TransAD's F1 score is lower than TSMAE, which shows that the modeling ability needs to match the data set. From the VAE detection graph, it can be seen that misjudged abnormal points will appear intermittently in the same variable, which shows the importance of modeling time correlation. LSTM-VAE extracts hidden feature distributions containing time information to make reconstruction more accurate, and its F1 is higher than LSTMAE and VAE. However, DAE does not make targeted designs for variables and time correlations in process data, resulting in a large number of misjudgments in DAE's anomaly detection, indicating that DAE does not have the ability to detect process anomalies. Among the above methods, DAE has the fastest anomaly detection inference speed. However, the anomaly detection effect of DAE is much weaker than other methods. TSMAE masks a large number of variables before the process data is input into the encoder, reducing the encoder weight parameters and achieving faster inference speed than LSTMAE.

[0101] Figure 8 and Fig. 9 The original data and reconstructed data of the 16th and 28th variables during the first human interference are shown respectively. The two variables are marked as normal in the data set, but the model determines them to be abnormal. The 16th variable refers to the juice pressure at the evaporator inlet, which is easily affected by other variables, such as the temperature and flow rate in the evaporator. Figure 4 It can be seen that there is a sudden drop in the original data, and the specific reason is that human interference causes the evaporator outlet flow to increase. The misjudgment of this variable cannot indicate that there is a problem with the model performance. The 28th variable refers to the boiler water level and has nothing to do with human interference. The original data fluctuates reasonably within a certain range. TSMAE, LSTMAE and LSTM-VAE can roughly fit the fluctuation trend and do not misjudge it. However, TransAD and VAE basically lose the fit to the data distribution and cannot extract the correct variance, resulting in the variable being judged as abnormal. This shows that the two models lose the capture of time trends during the reconstruction process. The results show that the TSMAE method proposed in this embodiment has faster inference speed and better F1 than the baseline model, with an improvement of 6%-89%.

[0102] This embodiment proposes a process anomaly detection model based on TSMAE. This mechanism randomly masks most input variables, reduces information redundancy, and uses LSTM to model the time correlation of process data to extract hidden features and reconstruct original data, and fills equal amounts of learnable vectors in hidden features. It also uses the overall reconstruction loss as the optimization target to enhance the model's ability to model global information. Compared with multiple baseline models, this method further fully verifies that it improves the anomaly detection reasoning speed and anomaly detection accuracy, and is universal and general.

[0103] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the embodiment concepts. The protection scope of the embodiments should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the embodiments also extends to equivalent technical means that can be conceived by those skilled in the art based on the concepts of the embodiments.

Claims

1. A process anomaly detection method based on a time-series masked autoencoder, characterized in that: The following steps are involved: (1) Data preprocessing Collect the original process data. To ensure the consistency of the collected process data, standardize the data and convert the data to a mean of 0 and a standard deviation of 1 so that the variables can be arranged on a common scale. Use a sliding time window to extract continuous subsets of the standardized data as new samples. (2) Feature masking and data reconstruction 2.1) Before the original data is input into the encoder, the input variables are randomly sampled and masked. By reducing the input features of the encoder, the feature extraction speed is improved to ensure the real-time requirements of the anomaly detection task; 2.2) In order to model the temporal correlation in process data and improve the accuracy of process anomaly detection, an autoencoder with LSTM as the core is used to extract and restore the hidden features of process data, and a learnable vector is introduced into the autoencoder; (3) Model training and anomaly location 3.1) Add the Sigmoid activation function to the encoder's LSTM network; specific hyperparameters must be predefined before starting model training, including the number of iterations, batch size, and learning rate; 3.2) The training process is carried out in batches, and the Adam optimizer is used to find the optimal network parameters 3.3) Training all samples once is regarded as a complete iteration. Through multiple rounds of iterative optimization, the model can achieve the reconstruction of input data with the minimum error. 3.4) Based on the data set, use the corresponding anomaly location method to filter out anomalies or outliers.

2. The process anomaly detection method based on the time series masked autoencoder according to claim 1, characterized in that: The process of (1) is as follows: Step 1.1: Data standardization, treating the collected process data as a two-dimensional array Where F represents the number of features and T represents the amount of data collected; the specific process of standardization is as follows: Among them, x i and Respectively represent the process characteristics before and after standardization, μ i and σ i Respectively represent the mean and variance of the i-th feature. The standardized data is still represented by X; Step 1.2: Perform sliding time window extraction on the original data X. The new data set extracted using the sliding time window will be regarded as a three-dimensional array. Where N is the number of samples in the new data set, W is the time window of fixed size that is systematically moved in the data set, and S is the moving distance of the time window of fixed size. The relationship between N and T, W and S is as follows:

3. The process anomaly detection method based on time-series masked autoencoder according to claim 1, characterized in that: The process of (2) is as follows: Step 2.1: Define X as the original data, with a masking rate of α; Step 2.2: Randomly sample F×α variables and remove them from the input to generate the encoder input Where L is the number of variables remaining and conforms to the following quantitative relationship: L=(1-α)×F; (3) Step 2.3: Extract two-dimensional arrays from the original data to form a data set C = in Define the hidden feature dimension extracted by the LSTM encoder as D, and the visible input C n The hidden features of This is equivalent to mapping each process variable into a D / L-dimensional hidden matrix; Step 2.4: The added learnable vector is represented as The size of the learnable vector is determined by the relationship between the mask rate α, the hidden feature dimension D, and the total number of variables F. The added learnable vector helps TSMAE model the variable correlation and global information of process data, and ensures that the number of features can be kept consistent during the reconstruction phase.

4. The process anomaly detection method based on the time series masked autoencoder according to claim 3, characterized in that: The AE structure is used in the reconstruction work. The AE structure includes a two-layer neural network structure. The encoder and decoder are both composed of an LSTM. The encoder is responsible for extracting hidden features H from the unmasked data C and adding learnable vectors to the hidden features corresponding to the masked feature positions. The decoder reconstructs the original input through the hidden features and learnable vectors. The visible variables in the time window are defined as Where M represents the number of visible variables; in the training phase, the overall reconstruction deviation is used as the optimization target, and the formula is as follows: Among them, Y represents the reconstructed output of TSMAE, E(.) and D(.) are the encoder and decoder respectively, and concat(.) represents the concatenation function.

5. The process anomaly detection method based on time-series masked autoencoder according to claim 1, characterized in that: The process of (3) is as follows: Step 3.1: Extract training and test samples from the original data using the sliding time window method; Step 3.2: In the training phase, the normal time series data obtained through standardization and preprocessing is used as the normal sample training model, and the model parameters are updated according to the reconstruction error of the normal data until the preset training conditions are met; Step 3.3: Use the trained TSMAE to evaluate the reconstruction loss of the test sample, and use the 3sigma method to identify abnormal features and corresponding time, and compare the performance indicators of TSMAE and the baseline model.

Images