An Abnormal Detection Method for Heterogeneous Low-Quality Industrial Time-Series Data

Abstract

The present invention discloses a method for detecting anomalies in heterogeneous low-quality industrial time-series data. The method includes: collecting heterogeneous time-series data, performing continuous process variable augmentation, and then constructing a training set; in the pre-training stage of heterogeneous feature integration, constructing a heterogeneous representation learning network and using the pre-training sample set in the training set for pre-training; in the fine-tuning stage of fusion detection adaptation, constructing a heterogeneous fusion detection network and using the fine-tuning sample set in the training set for training; using the finally trained heterogeneous fusion detection network to obtain the reconstructed values of continuous process variables, and completing anomaly detection according to the anomaly detection control limit. The present invention can accurately model the temporal dependence relationship between heterogeneous variables in the low-quality scenario of variable heterogeneity and measurement point loss, effectively improving the accuracy of anomaly detection, and providing a strong support for realizing the intelligent operation and maintenance and optimization decision-making of industrial production systems.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormal detection in the process industry, and particularly to an abnormal detection method for heterogeneous and low-quality industrial time-series data. Background Art

[0002] Production quality control is the core of the manufacturing industry, and the monitoring and control of the process state are the key to achieving this goal. In the modern industrial production and operation and maintenance processes, the continuous and stable operation of equipment and systems is the basis for ensuring product quality, reducing energy consumption, improving economic efficiency, and maintaining production safety. With the development of industrial systems towards large-scale, complex, and intelligent directions, the number of devices has increased sharply, the process flow has a high degree of coupling, the production environment has become increasingly dynamic and variable, and various potential abnormalities have shown a trend of diversification and concealment. The traditional abnormal recognition method relying on manual experience has been difficult to meet the requirements of real-time, accurate, and large-scale monitoring. As a key technology for identifying the deviation of the system from the normal operating state, abnormal detection runs through all aspects of industrial intelligent operation and maintenance. By promptly discovering potential problems such as equipment aging, process deviation, and sensor failures, abnormal detection can not only effectively reduce the system downtime rate and extend the service life of equipment, but also provide decision-making support for production process optimization and energy efficiency improvement, and avoid safety accidents and major economic losses caused by the accumulation of failures. Especially in the context of the rapid development of big data and artificial intelligence technologies, data-driven abnormal detection methods can fully explore the deep feature information in the industrial process, realize the intelligent perception and accurate identification of the operating state of complex systems, and greatly improve the automation and intelligent level of operation and maintenance. Therefore, the research and development of efficient abnormal detection technologies facing the industrial big data environment and capable of coping with dynamic change characteristics have become an important technical support for promoting industrial intelligent transformation and building a safe and efficient production system, with great engineering value and broad application prospects.

[0003] With the development of the Internet of Things, 5G networks, and intelligent manufacturing, the data collected in industrial processes is usually mixed heterogeneous data containing discrete variables (DVs) and continuous variables (CVs). Continuous variables (such as temperature, pressure, flow rate, etc.) mainly record the dynamic changes of process parameters, while discrete variables (such as equipment status, valve opening and closing, etc.) reflect the key state information of operating conditions. In addition, due to objective reasons such as measurement, storage, and transmission limitations, some continuous variables are discretized into discrete variables during the recording process. This data heterogeneity poses significant challenges to the establishment of anomaly detection. Despite the continuous progress of detection technologies, existing research still pays insufficient attention to discrete variables. Common processing methods include directly deleting discrete variables during preprocessing or simply modeling discrete variables together with continuous variables. The former ignores the rich process knowledge contained in discrete variables, often leading to a decline in model performance; the latter ignores the information asymmetry problem between the two types of variables. The information granularity of discrete variables is relatively low, lacking the time-varying characteristics of continuous variables, such as autocorrelation, weekly patterns, and local fluctuations. Discrete variables usually show sudden changes or steady states due to their limited value ranges. Simple encoding (such as 0, 1, 2) may lose their important information and at the same time mislead the model into thinking that these values have a linear relationship, although they may actually represent different states. In addition, there are significant differences in their distribution characteristics. Continuous variables usually follow a Gaussian distribution, showing smooth numerical changes, while discrete variables tend to follow a Bernoulli distribution or other finite-category distributions. The difference in distribution types makes it more difficult to capture the complex non-linear relationship between the two, and simple encoding methods often cannot effectively extract the process information in discrete variables, thus affecting the anomaly detection effect. The key to solving the above heterogeneous problems lies in bridging the information gap between discrete variables and continuous variables. Discrete variables often imply unobserved continuous-value factors, but due to external limitations such as measurement, storage, and transmission limitations, this continuous information may be lost. For example, the frequent changes in process alarm signals may reflect the continuous change trend of system temperature. Therefore, mining the implicit continuity of discrete variables, restoring this continuous value, and obtaining its corresponding latent continuous variables (LCVs) can not only enrich the information content of discrete variables but also effectively bridge the information gap, providing a feasible path for solving the anomaly detection problem in mixed heterogeneous data.

[0004] In practical applications, anomaly detection of time series data often faces the data limitation of missing sampling of input variables. First, due to factors such as sensor measurement cycles, equipment failures, and communication interruptions, the problem of missing sampling in industrial process data collection is very common. These missing data usually contain key process state information, and the common data deletion or mean filling methods in existing models are difficult to accurately reflect the true dynamic changes of the process, thus reducing the model's ability to model the dynamic behavior of industrial processes. In addition, missing sampling destroys the continuity of time series data and affects the effects of feature extraction and sequence modeling. Especially in industrial environments with high-dimensional multi-variables, missing values will weaken the expression ability of the dependence relationship between variables, further affecting the performance and stability of the model.

[0005] In recent years, self-supervised learning, as an effective method, has shown great potential in reducing the dependence on large-scale labeled data. Under the self-supervised learning paradigm, the model autonomously mines potential structures and semantic information from unlabeled data through auxiliary tasks or generative tasks. Self-supervised learning can obtain high-quality data representations without manual labeling. In the field of industrial anomaly detection, abnormal samples are usually extremely scarce, and the abnormal patterns show diversity and complexity, which puts higher requirements on the data representation ability and generalization ability of the anomaly detection model. Self-supervised learning can make full use of a large amount of normal process data, establish a robust feature representation system by learning the dynamic characteristics under normal patterns, and thus still be able to support efficient and accurate anomaly detection in the case of extremely few abnormal samples. At present, this method has achieved remarkable results in fields such as computer vision and natural language processing. Self-supervised methods centered on contrastive learning usually use auxiliary tasks to improve the instance discrimination ability. Subsequently, the proposed masked autoencoder (MAE) method learns visual or semantic representations by reconstructing part of the observed data, further promoting the development of self-supervised learning and demonstrating excellent performance. However, when applied to handle heterogeneous missing data in industrial scenarios, existing MAE methods have obvious limitations. Industrial data not only contains significant time series continuity and dynamic changes, but also contains complex interactions between discrete variables and continuous variables. Traditional MAE methods mainly focus on the reconstruction of continuous variables and have insufficient ability to model the implicit continuity of discrete variables, making it difficult to comprehensively represent the characteristics of heterogeneous data and capture the dynamic change patterns behind discrete variables. This deficiency is particularly prominent in industrial data processing tasks, resulting in the model being unable to effectively recover in the case of missing discrete information, thus weakening the overall anomaly detection performance.

[0006] Aiming at the problem of low-quality data caused by variable heterogeneity and sampling missing in industrial processes, it is urgent to design a new method for anomaly detection of heterogeneous low-quality industrial time series data to improve the performance and robustness of the model in complex industrial environments. Summary of the Invention

[0007] The object of the present invention is to provide an abnormal detection method for heterogeneous low-quality industrial time series data in view of the deficiencies of the prior art. The present invention can significantly improve the accuracy and robustness of the key parameter inference and abnormal detection model when dealing with heterogeneous missing industrial data.

[0008] The object of the present invention is achieved by the following technical solutions: An abnormal detection method for heterogeneous low-quality industrial time series data, comprising the following steps:

[0009] (1) Collect normal heterogeneous time series data in the process industry process, and scan it using a sliding time window method to obtain a serialized data set;

[0010] (2) Based on the serialized data set, design a continuous process variable augmentation method, and generate masked continuous process variables and pseudo-discrete process variables through mask matrix generation and standard deviation discretization to construct a training set;

[0011] (3) In the heterogeneous feature integration pre-training stage, construct a heterogeneous representation learning network, and use the training set for pre-training to obtain a pre-trained heterogeneous representation learning network;

[0012] (4) In the fusion detection adaptation fine-tuning stage, construct a heterogeneous fusion detection network based on the pre-trained heterogeneous representation learning network, and use the training set for training to obtain a finally trained heterogeneous fusion detection network; and estimate the abnormal detection control limit according to the reconstruction error;

[0013] (5) During application, input the serialized continuous process variables with missing measurement points and the serialized discrete process variables in the online operation data into the finally trained heterogeneous fusion detection network to obtain the reconstructed values of the corresponding continuous process variables; compare the reconstructed values of the continuous process variables with their corresponding true values and the abnormal detection control limit to complete the abnormal detection.

[0014] Further, the step (1) specifically includes:

[0015] First, collect normal heterogeneous time series data in the process industry process to construct a set of heterogeneous time series data, denoted as , where represents a mixed process variable data set containing all data points, represents the total number of samples in the set , represents the total number of mixed process variables; represents a mixed process variable data set with missing measurement points in the continuous process variables, represents the total number of samples in the set ;

[0016] Then perform z-score normalization on all continuous process variables;

[0017] Then use the sliding time window method to scan all the above standardized heterogeneous time series data to obtain a serialized data set;

[0018] For the set , add the binary mask matrix to it and use it as the fine-tuning sample set.

[0019] Further, the step (2) specifically includes:

[0020] For each sample in the serialized data set corresponding to the set , where represents the continuous process variable of the m-th sample in the serialized data set corresponding to the set , and represents the discrete process variable of the m-th sample in the serialized data set corresponding to the set ;

[0021] For each continuous process variable , perform a masking operation through a randomly determined binary mask matrix to obtain a masked continuous process variable;

[0022] Next, calculate the mean and standard deviation of each continuous process variable , and perform data discretization according to the principle to obtain the corresponding pseudo-discrete process variable;

[0023] Finally, construct a pre-training sample set based on the continuous process variable, its corresponding masked continuous process variable, pseudo-discrete process variable, and the original discrete process variable, and construct a training set in combination with the fine-tuning sample set.

[0024] Further, the heterogeneous representation learning network includes a missing data encoder, a discrete data encoder, a first feature fusion unit, a second feature fusion unit, a reconstruction decoder, and a recovery decoder;

[0025] Input the masked continuous process variable into the missing data encoder to extract the masked data feature; input the pseudo-discrete process variable into the discrete data encoder to extract the pseudo-discrete data feature; input the original discrete process variable into the discrete data encoder to extract the discrete data feature; input the masked data feature and the discrete data feature into the second feature fuser to obtain the fused masked data feature; input the pseudo-discrete data feature and the discrete data feature into the first feature fuser to obtain the fused pseudo-discrete data feature; input the fused masked data feature into the reconstruction decoder to obtain the reconstructed value of the original continuous process variable; input the fused pseudo-discrete data feature into the recovery decoder to obtain the recovered value of the original continuous process variable.

[0026] Further, in step (3), use the training set for pre-training to obtain a pre-trained heterogeneous representation learning network, specifically including:

[0027] Input the masked continuous process variable, pseudo-discrete process variable, and original discrete process variable in the pre-training sample set of the training set into the heterogeneous representation learning network to obtain the reconstructed value and recovered value of the original continuous process variable;

[0028] Calculate the reconstruction loss function according to the reconstructed value of the original continuous process variable, the true value of the original continuous process variable in the pre-training sample set, and its corresponding binary mask matrix; calculate the mean square error as the recovery loss function according to the recovered value of the original continuous process variable and the true value of the original continuous process variable corresponding to it in the pre-training sample set; obtain the total loss function of the heterogeneous representation learning network by weighted summation of the reconstruction loss function and the recovery loss function;

[0029] Taking the minimization of the total loss function of the heterogeneous representation learning network as the optimization objective, optimize and adjust the parameters of the missing data encoder, discrete data encoder, first feature fuser, second feature fuser, reconstruction decoder, and recovery decoder until the preset number of training rounds is reached or the total loss function of the heterogeneous representation learning network converges, to obtain a pre-trained heterogeneous representation learning network, and obtain a pre-trained missing data encoder, discrete data encoder, second fuser, and reconstruction decoder.

[0030] Further, the heterogeneous fusion detection network includes a pre-trained missing data encoder, discrete data encoder, second feature fuser, and reconstruction decoder;

[0031] Input the serialized continuous process variables with missing measurement points into the pre-trained missing data encoder to extract missing data features; input the serialized discrete process variables into the pre-trained discrete data encoder to extract discrete data features; input the missing data features and discrete data features into the pre-trained second feature fusion device to obtain fused missing data features; input the fused missing data features into the pre-trained reconstruction decoder to obtain the reconstructed values of the continuous process variables.

[0032] Further, in step (4), use the training set for training to obtain the finally trained heterogeneous fusion detection network, which specifically includes:

[0033] Input the serialized continuous process variables with missing measurement points and the serialized discrete process variables in the fine-tuning sample set in the training set into the heterogeneous fusion detection network to obtain the reconstructed values of the continuous process variables;

[0034] Calculate the total loss function of the heterogeneous fusion detection network according to the mean square error between the reconstructed values of the continuous process variables and the corresponding continuous process variables in the fine-tuning sample set;

[0035] Taking the minimization of the total loss function of the heterogeneous fusion detection network as the optimization objective, fine-tune the parameters of the missing data encoder, discrete data encoder, second feature fusion device and reconstruction decoder until the preset number of training rounds is reached or the total loss function of the heterogeneous fusion detection network converges, and obtain the finally trained heterogeneous fusion detection network.

[0036] Further, the estimating the anomaly detection control limit according to the reconstruction error specifically includes:

[0037] First, based on the samples in the fine-tuning sample set, calculate the reconstructed root mean square error corresponding to each sample according to the reconstructed values of the continuous process variables and the corresponding true values;

[0038] Then, use the radial basis kernel function to calculate the probability density function of the reconstructed root mean square error;

[0039] Finally, use the kernel density estimation method to calculate the anomaly detection control limit based on the probability density function of the reconstructed root mean square error.

[0040] The beneficial effects of the present invention are as follows: By analyzing the data or signals generated in the industrial production process, the present invention can identify significant deviations of the process state from the normal working state, and thus discover potential faults or abnormal behaviors; In view of the non-linear correlation between continuous variables and discrete variables, the present invention mines the potential associations between variables by learning and fusing heterogeneous time-series representations, and thus realizes anomaly detection in a unified continuous space; In view of the problem of missing data, the present invention realizes continuous missing reconstruction through a missing data encoder, a second feature fusion device, and a reconstruction decoder, uses the discrete variable information to assist in completing the reconstruction task of partially missing data, endows the model with the ability to efficiently fill in the missing data of continuous variables, and weakens the impact of data missing on the model performance; In view of the potential continuity hidden in discrete variables, the present invention introduces the idea of hidden continuity restoration, and completes the restoration of pseudo-discrete variables with the assistance of discrete variable information, eliminates the interference of the variable information asymmetry problem on data analysis, and improves the model's ability to capture and restore the underlying dynamic patterns of discrete variables; The present invention can make full use of unlabeled normal process data, improves the robustness of representation learning through a pre-training strategy, considers the problem of data distribution shift, and the model can use recently collected normal process data for parameter fine-tuning to complete the adaptation of the model to the new data distribution; Different from the simple way of dealing with heterogeneous missing data in traditional anomaly detection methods, the present invention analyzes heterogeneous missing data in a unified continuous value space through self-supervised mask reconstruction and hidden continuity restoration design, making the model have stronger anomaly detection ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flowchart of the method for detecting anomalies in heterogeneous low-quality industrial time-series data of the present invention;

[0042] Figure 2 is a schematic diagram of the task of the method for detecting anomalies in heterogeneous low-quality industrial time-series data of the present invention;

[0043] Figure 3 is an architecture flowchart of the heterogeneous representation learning network in the pre-training stage and the heterogeneous fusion detection network in the fine-tuning stage of the present invention;

[0044] Figure 4 is a visualization result diagram of the restoration of the potential continuity of pseudo-discrete variables in Embodiment 1 of the present invention; among them, Figure 4 in (a) is a schematic diagram of the original pseudo-discrete variable, Figure 4 in (b) is a diagram of the true continuous variable and the result diagram of the restored hidden continuous variable;

[0045] Figure 5 is a visualization result diagram of the restoration of the potential continuity of pseudo-discrete variables in Embodiment 2 of the present invention; among them, Figure 5 in (a) is a schematic diagram of the original pseudo-discrete variable, Figure 5In (b), the real continuous variable and the result graph of the recovered hidden continuous variable are shown. Detailed implementation manners

[0046] Here, exemplary embodiments will be described in detail, and examples thereof are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0047] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "said" used in the present invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0048] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0049] The present invention will be described in detail below with reference to the drawings. Without conflict, the features in the following embodiments and implementation manners may be combined with each other.

[0050] See Figure 1 , the heterogeneous low-quality industrial time series data anomaly detection method of the present invention specifically includes the following steps:

[0051] (1) Collect normal heterogeneous time series data in the process industry process, and scan it using the sliding time window method to obtain a serialized data set.

[0052] Specifically, first collect normal heterogeneous time series data in the process industry process to construct a set of heterogeneous time series data, denoted as , where represents a mixed process variable data set containing all data points, as shown in Figure 2 , represents the set the total number of samples in, represents the total number of mixing process variables; represents the set the sampling value of the mixing process variables including all data points for the p-th sample point in the set represents the set the sampling value of the continuous process variables for the p-th sample point in the set, where the values are continuous represents the total number of continuous process variables represents the set the sampling value of the discrete process variables for the p-th sample point in the set, with values in discrete states such as 0 or 1 represents the total number of discrete process variables, having ; represents the set of mixing process variable data with missing measurement points and labels for continuous process variables represents the set the total number of samples in ; represents the set the sampling value of the mixing process variables with missing measurement points for the continuous process variables for the q-th sample point in the set represents the set the sampling value of the continuous process variables with missing measurement points for the q-th sample point in the set represents the set the sampling value of the discrete process variables including all data points for the q-th sample point in the set; define as a binary mask matrix, used to indicate whether the continuous process variables are missing, 0 represents present, 1 represents missing. When collecting continuous process variables, when a certain measurement point is missing, the value at the corresponding position in the binary mask matrix is set to 1 is a mask vector used to indicate the missing measurement points of the continuous process variables for the q-th sample point.

[0053] Then perform z-score standardization on all continuous process variables to obtain the standardized continuous process variables. Z-score standardization can eliminate the influence of variable dimensions. Among them, all continuous process variables include the set and the set the continuous process variables in. The calculation formula for z-score standardization is:

[0054]

[0055] In the formula, and are the mean and standard deviation of the corresponding continuous process variables respectively, represents the continuous process variable to be standardized, represents the standardized continuous process variable.

[0056] Then, the sliding time window method is adopted to scan all the above standardized heterogeneous time series data to obtain a serialized data set: For the set , each sample in its serialized data set is represented as , where is the time window length; For the set , each sample in its serialized data set is represented as ; For the binary mask matrix , each sample in its serialized data set is represented as ; For the set , the binary mask matrix is added to it, then each sample in its corresponding serialized data set is represented as , , and it is used as the fine-tuning sample set .

[0057] It should be noted that since it is impossible to input all 8000 data points into the model, which is too large, generally, window serialization has to be done first, and then the serialized data is input into the model for the reconstruction of continuous process variables. Serialized data samples are obtained through the sliding time window method. At this time, a sample refers to a certain sliding window matrix data. Before serialization, a sample refers to a certain time point vector data. For in the data, m is used to represent the m-th sample, that is, the m-th sliding window . For the data { }, n is used to represent the n-th sample, that is, the n-th sliding window { }.

[0058] (2) Based on the serialized data set, design a continuous process variable augmentation method to generate masked continuous process variables and pseudo-discrete process variables through mask matrix generation and standard deviation discretization to construct a training set. The training set constructed in the current step (2) and the test set constructed in the above step (1) can support the self-supervised pre-training task in the subsequent steps.

[0059] Specifically, for each sample in the serialized data set corresponding to the set , it is represented as , where represents the continuous process variable of the m-th sample in the serialized data set corresponding to the set , and represents the discrete process variable of the m-th sample in the serialized data set corresponding to the set . For each continuous process variable , a binary mask matrix is randomly determined for each. Exemplarily, in this embodiment, for each element in the binary mask matrix, the initial random assignment follows a uniform distribution over the interval to represent its missing probability, and the missing rate is set to . When the value of the binary mask matrix element is less than , it is determined that the element is missing and assigned a value of 0; otherwise, it is determined that it is not missing and assigned a value of 1. Then, the masking operation is performed using the binary mask matrix to simulate data missingness, obtaining the masked continuous process variable, which is expressed as:

[0060]

[0061] where, represents the dot product of vectors, represents the masked continuous process variable, and the corresponding masked continuous process variable data set is represented as . Next, the mean and standard deviation of each continuous process variable are calculated. According to the principle, data discretization is performed. Specifically, according to the position of the variable value of each continuous process variable relative to the set upper and lower bounds, it is transformed into discrete states to obtain the corresponding pseudo-discrete process variable , and then the pseudo-discrete process variable data set is obtained. Exemplarily, in this embodiment, the upper bound of each continuous process variable is set to the corresponding mean plus one standard deviation, and the lower bound is set to the corresponding mean minus one standard deviation; according to the position of the variable value relative to the upper and lower bounds, it is transformed into three discrete states: for , if its variable value is less than the lower bound, it is converted to 0; if its variable value is between the upper and lower bounds, it is converted to 1; if its variable value is greater than the upper bound, it is converted to 2; after processing in this way, the pseudo-discrete process variable can be obtained. Finally, through the above steps, data augmentation of the continuous process variable containing all data points in the set is performed to obtain its corresponding masked continuous process variable and pseudo-discrete process variable and as shown in Figure 3 , realizing data augmentation of the continuous process variable. According to the continuous process variable, its corresponding masked continuous process variable, pseudo-discrete process variable, and the original discrete process variable, a pre-training sample set is constructed, and combined with the fine-tuning sample set to construct the training set .

[0062] It should be understood that the pseudo-discrete process variable is a "false" discrete variable, which can be understood as simulating a discrete scenario. It is a "pseudo" discrete process variable created by discretizing a continuous process variable and has no relation to the original discrete process variable.

[0063] (3) In the heterogeneous feature integration pre-training stage, construct a heterogeneous representation learning network. Specifically, the heterogeneous representation learning network includes a missing data encoder, a discrete data encoder, a first feature fusion device, a second feature fusion device, a reconstruction decoder, and a recovery decoder, as Figure 3 shown; and use the pre-training sample set in the training set to pre-train the heterogeneous representation learning network to obtain a pre-trained heterogeneous representation learning network, and obtain a pre-trained missing data encoder, discrete data encoder, second feature fusion device, and reconstruction decoder.

[0064] In this embodiment, the missing data encoder, the discrete data encoder, the reconstruction decoder, and the recovery decoder can all be implemented using a multi-layer fully connected network, and the encoding and decoding of data in the feature space are realized through linear transformation and non-linear activation functions. The first feature fusion device and the second feature fusion device can be implemented using a cross-attention mechanism, and the extraction of the correlation of variables in the feature space is realized through feature fusion. Specifically, consider the query sequence and the key-value sequence , where M is the length of the query sequence, N is the length of the key-value sequence, and is the feature dimension. Suppose there are h heads, and each head independently uses different learnable weight matrices to map the input data to the query matrix , the key matrix , and the value matrix , which is expressed as:

[0065]

[0066] In the formula, ,, , , are the learnable weight matrices of the query, key, and value respectively. Each head independently calculates the attention output, which is expressed as:

[0067]

[0068] In the formula, represents the output of the i-th head, represents the cross-attention mechanism, represents the softmax activation function, is the hidden dimension of the projection subspace. Concatenate the outputs of all heads and use a learnable weight matrix to map to the output, and obtain:

[0069]

[0070] In the formula, represents the vector connection operation. At the same time, in order to stabilize the training and improve the gradient propagation, the cross-attention mechanism usually includes residual connections and layer normalization, and the final output is:

[0071]

[0072]

[0073] In the formula, represents the layer normalization operation, represents the fully connected network, represents the result of the final feature fusion.

[0074] Specifically, as Figure 3 shown, the masked continuous process variable is input into the missing data encoder to extract the masked data feature , denoted as:

[0075]

[0076] In the formula, represents the missing data encoder, , ; it should be noted that the masked continuous process variable input into the missing data encoder is data in the form of a time window and is serialized. The pseudo-discrete process variable is input into the discrete data encoder to extract the pseudo-discrete data feature , denoted as:

[0077]

[0078] In the formula, represents the discrete data encoder, , . The original discrete process variable is input into the discrete data encoder to extract the discrete data feature , denoted as:

[0079]

[0080] In the formula, . The masked data feature and the discrete data feature are input into the second feature fuser to obtain the fused masked data feature , denoted as:

[0081]

[0082] In the formula, represents the second feature fuser, , ; through the second feature fuser, the correlation between the masked continuous process variable and the original discrete process variable can be extracted, and the masked data feature and the discrete data feature are fused to obtain the fused masked data feature. The pseudo-discrete data feature and the discrete data feature are input into the first feature fuser to obtain the fused pseudo-discrete data feature , denoted as:

[0083]

[0084] In the formula, represents the first feature fuser, , ; through the first feature fuser, the correlation between the pseudo-discrete process variable and the original discrete process variable can be extracted, and the pseudo-discrete data feature and the discrete data feature are fused to obtain the fused pseudo-discrete data feature. The fused masked data feature is input into the reconstruction decoder to obtain the reconstruction value of the original continuous process variable , denoted as:

[0085]

[0086] In the formula, represents the reconstruction decoder, . The fused pseudo-discrete data feature is input into the recovery decoder to obtain the recovery value of the original continuous process variable , denoted as:

[0087]

[0088] In the formula, represents the recovery decoder, .

[0089] It should be noted that the missing data encoder, discrete data encoder, reconstruction decoder, and recovery decoder composed of multi-layer fully connected networks are built, and their network parameters are randomly initialized; the first feature fuser and the second feature fuser implemented by the cross-attention mechanism are built, and their network parameters are randomly initialized.

[0090] In this embodiment, the heterogeneous representation learning network is pre-trained using the pre-training sample set in the training set to obtain a pre-trained missing data encoder, discrete data encoder, second feature fusion device, and reconstruction decoder, specifically including: the masked continuous process variables in the pre-training sample set in the training set , pseudo-discrete process variables and the original discrete process variables are input into the heterogeneous representation learning network, and through the process shown in Figure 3 , the reconstruction value and recovery value of the original continuous process variable can finally be obtained. According to the reconstruction value of the original continuous process variable and the true value of the original continuous process variable in the pre-training sample set and its corresponding binary mask matrix, the reconstruction loss function is calculated; according to the recovery value of the original continuous process variable and the true value of the original continuous process variable corresponding to it in the pre-training sample set, the mean square error is calculated as the recovery loss function; by weighted summing the reconstruction loss function and the recovery loss function, the total loss function of the heterogeneous representation learning network is obtained. Taking the minimization of the total loss function of the heterogeneous representation learning network as the optimization goal, the parameters of the missing data encoder, discrete data encoder, first feature fusion device, second feature fusion device, reconstruction decoder, and recovery decoder are optimized and adjusted until the preset number of training rounds is reached or the total loss function of the heterogeneous representation learning network converges, so as to obtain a pre-trained heterogeneous representation learning network, and obtain a pre-trained missing data encoder, discrete data encoder, second feature fusion device, and reconstruction decoder.

[0091] Further, the calculation formula of the reconstruction loss function is:

[0092]

[0093] In the formula, represents the reconstruction loss function, represents the time window length, i represents the i-th time point of the time window, represents the total number of continuous process variables, represents the true value of the continuous process variable of the m-th sample in the pre-training sample set, is the reconstruction value of the corresponding original continuous process variable, represents the vector two-norm.

[0094] Further, the calculation formula of the recovery loss function is:

[0095]

[0096] In the formula, represents the recovery loss function, represents the recovery value of the original continuous process variable corresponding to the m-th sample in the pre-training sample set.

[0097] Furthermore, the calculation formula for the total loss function of the heterogeneous representation learning network is:

[0098]

[0099] In the formula, represents the total loss function of the heterogeneous representation learning network, represents the weight coefficient.

[0100] (4) In the fusion detection adaptation fine-tuning stage, a heterogeneous fusion detection network is constructed based on the pre-trained heterogeneous representation learning network. This heterogeneous fusion detection network includes the pre-trained missing data encoder, discrete data encoder, second feature fusion device, and reconstruction decoder in step (3); the heterogeneous fusion detection network is trained using the fine-tuning sample set in the training set to obtain the finally trained heterogeneous fusion detection network; and the anomaly detection control limit is estimated according to the reconstruction error.

[0101] In this embodiment, as Figure 3 shown, the serialized continuous process variable with missing measurement points is input into the pre-trained missing data encoder to extract the missing data feature , denoted as:

[0102]

[0103] In the formula, , . The serialized discrete process variable is input into the pre-trained discrete data encoder to extract the discrete data feature , denoted as:

[0104]

[0105] In the formula, . The missing data feature and the discrete data feature are input into the pre-trained second feature fusion device to obtain the fused missing data feature , denoted as:

[0106]

[0107] In the formula, ; The second feature fusion unit can extract the correlation between the continuous process variables and the discrete process variables with missing measurement points, and fuse the missing data features and the discrete data features to obtain the fused missing data features. The fused missing data features are input into the pre-trained reconstruction decoder to obtain the reconstructed values of the continuous process variables , that is, the reconstructed values of the continuous process , denoted as:

[0108]

[0109] In the formula, represents the vectorization operation.

[0110] In this embodiment, the heterogeneous fusion detection network is trained using the fine-tuning sample set in the training set to obtain the finally trained heterogeneous fusion detection network, which specifically includes: serializing the continuous process variables with missing measurement points in the fine-tuning sample set in the training set and the serialized discrete process variables are input into the heterogeneous fusion detection network. Through the process as Figure 3 shown, the reconstructed values of the continuous process variables can finally be obtained. According to the reconstructed values of the continuous process variables and the mean square error between the corresponding continuous process variables in the fine-tuning sample set, the total loss function of the heterogeneous fusion detection network is calculated. Taking minimizing the total loss function of the heterogeneous fusion detection network as the optimization objective, the parameters of the missing data encoder, the discrete data encoder, the second feature fusion unit, and the reconstruction decoder are fine-tuned until the preset number of training rounds is reached or the total loss function of the heterogeneous fusion detection network converges, and the finally trained heterogeneous fusion detection network is obtained for online anomaly detection.

[0111] Furthermore, the calculation formula of the total loss function of the heterogeneous fusion detection network is:

[0112]

[0113] In the formula, represents the total loss function of the heterogeneous fusion detection network, represents the time window length, i represents the i-th time point of the time window, represents the total number of continuous process variables, represents the binary mask vector of the continuous process variables of the n-th sample in the fine-tuning sample set, represents the true value of the continuous process variables of the n-th sample in the fine-tuning sample set, is the reconstructed value of the corresponding original continuous process variable, represents the vector two-norm.

[0114] Furthermore, the anomaly detection control limit is estimated based on the reconstruction error, specifically including: First, based on the samples in the fine-tuning sample set, the root mean square error (RMSE) of each sample is calculated according to the reconstructed value and the corresponding true value of the continuous process variable, denoted as:

[0115]

[0116] In the formula, represents the root mean square error of the nth sample in the fine-tuning sample set, . Then, the probability density function of the root mean square error of the reconstruction is calculated using the radial basis kernel function , denoted as:

[0117]

[0118] In the formula, is the bandwidth of the radial basis function. The smaller the bandwidth, the smoother the probability density function curve obtained. Finally, based on the probability density function of the root mean square error of the reconstruction, the kernel density estimation method is used to calculate the anomaly detection control limit , denoted as:

[0119]

[0120] In the formula, represents the preset confidence level. When the integral value reaches the confidence level , the corresponding is the anomaly detection control limit value at this confidence level. When the root mean square error of a certain sample is greater than the anomaly detection control limit , it is considered that an anomaly is detected.

[0121] Furthermore, as time goes by, considering that the data distribution in the actual industrial production process may shift due to reasons such as equipment aging, working condition fluctuations, and environmental condition changes, when new normal process variable data is collected, it can still be used as a new fine-tuning sample set to fine-tune the parameters of the heterogeneous fusion detection network, so that the anomaly detection model can dynamically adapt to the new normal data distribution, effectively alleviate the performance degradation caused by distribution shift, and ensure the stable operation of the model throughout the production life cycle.

[0122] (5)During application, the serialized continuous process variables with missing measurement points and the serialized discrete process variables in the online operation data are input into the finally trained heterogeneous fusion detection network to obtain the reconstructed values of the corresponding continuous process variables; the reconstructed root mean square error is calculated based on the reconstructed values of the continuous process variables and their corresponding true values, and it is compared with the anomaly detection control limit to complete anomaly detection: if the calculated reconstructed root mean square error is greater than the anomaly detection control limit, it is considered that an anomaly is detected; if the calculated reconstructed root mean square error is less than or equal to the anomaly detection control limit, it is considered normal.

[0123] In summary, the present invention deeply mines the correlation information between discrete variables and continuous variables by learning and fusing heterogeneous time series features. Continuous missing reconstruction is realized through a missing data encoder, a second feature fusion device, and a reconstruction decoder, endowing the model with the ability to complete missing data. The discrete data encoder, the first feature fusion device, and the recovery decoder are used to realize discrete hidden continuous recovery, endowing the model with the ability to recover the hidden continuity of discrete data. The present invention makes full use of unlabeled normal process data through pre-training, and fine-tunes recent data with data distribution shift to enhance the adaptability of the model, significantly improving the accuracy and robustness of the anomaly detection model when dealing with heterogeneous missing industrial data.

[0124] The embodiments of the present invention list two typical examples, namely: anomaly detection of the operation process data of a gas turbine in a thermal power plant and anomaly detection of the operation process data of a server cluster. The heterogeneous low-quality industrial time series data anomaly detection method described in the present invention can cover any anomaly detection tasks with heterogeneous low-quality problems in many fields including but not limited to energy, information technology, etc.

[0125] Example 1: Anomaly Detection of Heterogeneous Low-Quality Time Series Data in the Operation Process of Gas Turbine Equipment

[0126] In the energy field, the stable operation of typical gas turbine equipment, as one of the key elements to ensure the reliability of power supply and system efficiency, has a crucial impact on the overall operation efficiency of a thermal power plant. When a gas turbine operates, the dynamic change of the unit load will cause the coal feeding amount to be continuously adjusted according to the actual demand, and then the operating state of the coal mill will continue to change, showing typical non-stationary transient characteristics. In this case, with the help of an effective anomaly detection method, potential faults can be identified in a timely manner, which is of great significance for optimizing the power generation process, improving economic benefits, reducing the risk of equipment shutdown, and extending the service life of equipment.

[0127] A gas turbine mainly consists of three core components: a compressor, a combustion chamber, and a gas turbine. In this embodiment, heterogeneous samples are collected from various time nodes during equipment operation to construct a heterogeneous time series dataset. This dataset covers 30 process variables, mainly including the following types of features: 10 discrete temperature variables of the gas turbine measurement points of the unit, indicating whether the current measurement point exceeds the preset threshold in a state of 0 or 1; 8 continuous variables of the compressor of the unit, covering various aspects of information such as vibration, pressure, and valve opening; 12 other variables from the turbine, combustion chamber, and overall measurement points of the unit.

[0128] To evaluate the performance of the method described in the present invention in anomaly detection, this embodiment focuses on the online application stage where abnormal conditions may occur. In this stage, the time series data of the process variables in the online operation data are input into the fine-tuned heterogeneous fusion detection network to obtain the corresponding reconstructed values of the continuous process variables. In addition, classic reconstruction methods such as partial least squares regression (PLSR), extreme learning machine (ELM), and advanced deep sequence learning models such as long and short term memory (LSTM), temporal convolution network (TCN), and Transformer are selected as comparison models to verify the effectiveness of the method described in the present invention. To ensure the smooth deployment and application of the comparison methods, the sampling missing sequences are processed by the method of neighboring mean filling, and simple categorical encoding is performed on the discrete variables. In addition, commonly used evaluation indicators in the field of anomaly detection, including Precision, Recall, and F1-score, are selected to comprehensively evaluate the detection accuracy. The calculation formulas for these indicators are as follows:

[0129]

[0130]

[0131]

[0132] Among them, represents Precision, represents Recall, represents F1-score, is the number of samples that are actually abnormal and detected as abnormal, is the number of samples that are actually abnormal but detected as normal, is the number of samples that are actually normal but detected as abnormal. The precision metric measures the proportion of samples detected as "abnormal" by the model that are actually "abnormal", indicating the accuracy of the model in identifying abnormalities. The recall metric measures the proportion of samples that are actually "abnormal" and correctly identified by the model, indicating the model's ability to capture abnormalities. The F1-score metric can be used to comprehensively evaluate the performance of the model, weighing precision and recall to avoid bias between the two. The comparison results of the final detection performance of the method described in the present invention and five other reconstruction models are shown in Table 1.

[0133] Table 1: Comparison results of anomaly detection performance of different methods (%)

[0134]

[0135] Use the kernel density estimation (KDE) method to estimate the probability density function of the RMSE of the fine-tuning samples. Based on the given confidence level , in this embodiment, take , then take the quantile of the above probability density function as the control limit of the reconstruction error, that is, the alarm threshold. When the RMSE index corresponding to the new sample exceeds the alarm threshold, it is considered that a small-probability event occurs, indicating that a process fault is detected and the process is abnormal. According to the data in Table 1, it can be seen that the method described in the present invention performs excellently in the three evaluation metrics of Precision, Recall, and F1-score, where Precision reaches 90.34%, Recall is 86.34%, and F1-score is as high as 88.29%, significantly superior to other reconstruction models. Compared with the second-best TCN method, the F1-score of the method described in the present invention is increased by 6.07%, which indicates that the method described in the present invention has obvious advantages in the accuracy and robustness of anomaly detection. The implicit continuous recovery results of the pseudo-discrete variables corresponding to a certain three continuous variables are finally obtained as shown in Figure 4 , where the original pseudo-discrete variables are as shown in (a) of Figure 4 , and the true continuous variables and the implicitly continuous variables obtained by recovery of the method described in the present invention are as shown in (b) of Figure 4 . Figure 4 The visualization results shown indicate that the discrete data encoder, the first feature fusion unit, and the recovery decoder in the method described in the present invention have good implicit continuous recovery capabilities. Whether for periodic signals or mutation signals, each module can more accurately reproduce the underlying implicit continuity of the pseudo-discrete variables generated by their augmentation, proving the effectiveness and reliability of the implicit continuous recovery idea.

[0136] In addition, to verify the generality and effectiveness of the method described in the present invention in causal analysis tasks in other fields, a brief implementation example in the field of information technology is given below.

[0137] Embodiment 2: Anomaly Detection of Heterogeneous and Low-Quality Time-Series Data during the Operation of Server Cluster Devices

[0138] In the field of information technology, the stable operation of servers is a core element for ensuring the efficient development of various online services, playing a fundamental supporting role in enterprise information processing and service supply. During the operation of servers, due to the interaction of many complex factors such as the drastic fluctuations in the concurrent user access volume and the sudden start and stop of background tasks, the internal computing rhythm and data transmission mode are always in a state of dynamic adjustment, which makes key performance indicators such as CPU utilization rate and memory usage show a high degree of instability and complexity. In such a complex environment, by using advanced machine learning algorithms and intelligent data analysis models to deeply mine and accurately analyze server operation data, potential abnormal signs can be quickly and accurately captured. This is of inestimable significance for ensuring the continuous operation of enterprise key services, improving the smoothness of the user experience, reducing economic losses caused by server failures, and enhancing the enterprise's reputation and status in the digital field.

[0139] In this embodiment, the server cluster device (Server Machine Dataset, SMD) is taken as the research object, and the operation data of the server cluster device is collected to verify the anomaly detection performance of the method described in the present invention. This dataset covers the telemetry data of multiple servers in different time periods. Specifically, the dataset contains 38 process variables, involving multiple monitoring indicators, such as CPU usage, memory usage, disk I / O, network traffic, etc. Among them, 19 are discrete variables, indicating whether the corresponding variable exceeds the preset threshold in a state of 0 or 1, and the other 19 are continuous value variables. The final comparison results of the inference performance are shown in Table 2.

[0140] Table 2: Comparison Results of Anomaly Detection Performance of Different Methods (%)

[0141]

[0142] Different from many laboratory simulation data, the SMD dataset is derived from the actual server operating environment and has extremely high practical application value. This data may be affected by various complex factors such as external system interference, data loss, and sensor errors. Therefore, it faces greater challenges to perform high-precision anomaly detection on it. By comparison, the method described in the present invention has the highest F1-score, which fully demonstrates that the method described in the present invention has significant advantages in anomaly detection in the real-world server operation and maintenance scenarios. By analyzing low-quality mixed time series data, the model can discover potential faults in advance and take preventive measures to avoid system crashes or service interruptions. Figure 5 shows the visualization results of the implicit continuity restoration of the pseudo-discrete variables corresponding to three continuous variables, where the original pseudo-discrete variables are as shown in Figure 5 (a) in, the true continuous variables of the method described in the present invention and the restored implicit continuous variables are as shown in Figure 5 (b) in, Figure 5 The visualization results shown indicate that with the support of implicit continuity restoration, the discrete data encoder, the first feature fusion device, and the restoration decoder in the method described in the present invention can focus on the inherent continuity of the implicit continuous variables and their fine-grained time dynamic characteristics hidden in the discrete variables, enabling effective anomaly detection of variables in a unified continuous value space.

[0143] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for detecting anomalies in heterogeneous low-quality industrial time-series data, characterized in that, It includes the following steps: (1) Collect the normal heterogeneous time series data in the process industry process, and scan it using the sliding time window method to obtain a serialized data set; (2) Based on the serialized data set, design a continuous process variable augmentation method, and generate masked continuous process variables and pseudo-discrete process variables through mask matrix generation and standard deviation discretization to construct a training set; The specific steps of step (2) include: For the set X in For each sample in the corresponding serialized dataset Where Represents the set X in The continuous process variable of the m-th sample in the corresponding serialized dataset of the set X, Represents the set X in The discrete process variable of the m-th sample in the corresponding serialized dataset of the set X; For each continuous process variable Perform a masking operation through a randomly determined binary mask matrix to obtain a masked continuous process variable; Next, calculate the mean and standard deviation of each continuous process variable and perform data discretization according to the 3σ principle to obtain the corresponding pseudo-discrete process variable; Finally, construct a pre-training sample set according to the continuous process variables and their corresponding masked continuous process variables, pseudo-discrete process variables, and the original discrete process variables, and construct a training set in combination with the fine-tuning sample set; (3) In the heterogeneous feature integration pre-training stage, construct a heterogeneous representation learning network and use the training set for pre-training to obtain a pre-trained heterogeneous representation learning network; (4) In the fusion detection adaptation fine-tuning stage, construct a heterogeneous fusion detection network based on the pre-trained heterogeneous representation learning network and use the training set for training to obtain a finally trained heterogeneous fusion detection network; and estimate the anomaly detection control limit according to the reconstruction error; (5) During application, input the serialized continuous process variables with missing measurement points and the serialized discrete process variables in the online operation data into the finally trained heterogeneous fusion detection network to obtain the reconstructed values of the corresponding continuous process variables; compare the reconstructed values of the continuous process variables with their corresponding true values and the anomaly detection control limit to complete anomaly detection.

2. The heterogeneous low-quality industrial time-series data anomaly detection method according to claim 1, wherein The specific steps of step (1) include: First, collect the normal heterogeneous time series data in the process industry process to construct a set of heterogeneous time series data, denoted as Ω = {X in , X mi}, where represents the set of mixed process variable data containing all data points, N in represents the total number of samples in the set X in , and d represents the total number of mixed process variables; represents the set of mixed process variable data with missing measurement points in the continuous process variables, N mi represents the total number of samples in the set X mi ; Then perform z-score standardization on all continuous process variables; Then scan all the standardized heterogeneous time series data using the sliding time window method to obtain a serialized data set; For the set X mi , add the binary mask matrix to it and use it as the fine-tuning sample set.

3. The heterogeneous low-quality industrial time-series data anomaly detection method according to claim 1, wherein The heterogeneous representation learning network includes a missing data encoder, a discrete data encoder, a first feature fusion device, a second feature fusion device, a reconstruction decoder, and a recovery decoder; Input the masked continuous process variables into the missing data encoder to extract masked data features; input the pseudo-discrete process variables into the discrete data encoder to extract pseudo-discrete data features; input the original discrete process variables into the discrete data encoder to extract discrete data features; input the masked data features and the discrete data features into the second feature fusion device to obtain fused masked data features; input the pseudo-discrete data features and the discrete data features into the first feature fusion device to obtain fused pseudo-discrete data features; input the fused masked data features into the reconstruction decoder to obtain the reconstructed values of the original continuous process variables; Input the fused pseudo-discrete data features into the recovery decoder to obtain the recovery values of the original continuous process variables.

4. The heterogeneous low-quality industrial time-series data anomaly detection method according to claim 1, characterized in that In step (3), using the training set for pre-training to obtain a pre-trained heterogeneous representation learning network specifically includes: Input the masked continuous process variables, pseudo-discrete process variables, and the original discrete process variables in the pre-training sample set of the training set into the heterogeneous representation learning network to obtain the reconstructed values and recovery values of the original continuous process variables; Calculate the reconstruction loss function based on the reconstructed values of the original continuous process variables, the true values of the original continuous process variables in the pre-training sample set, and their corresponding binary mask matrices; calculate the mean square error between the restored values of the original continuous process variables and the true values of the corresponding original continuous process variables in the pre-training sample set as the restoration loss function; obtain the total loss function of the heterogeneous representation learning network by weighted summation of the reconstruction loss function and the restoration loss function; Taking the minimization of the total loss function of the heterogeneous representation learning network as the optimization objective, optimize and adjust the parameters of the missing data encoder, discrete data encoder, first feature fuser, second feature fuser, reconstruction decoder, and restoration decoder until the preset number of training rounds is reached or the total loss function of the heterogeneous representation learning network converges, and obtain the pre-trained heterogeneous representation learning network, and get the pre-trained missing data encoder, discrete data encoder, second fuser, and reconstruction decoder.

5. The heterogeneous low-quality industrial time-series data anomaly detection method according to claim 1, wherein The heterogeneous fusion detection network includes a pre-trained missing data encoder, discrete data encoder, second feature fuser, and reconstruction decoder; Input the serialized continuous process variables with missing measurement points into the pre-trained missing data encoder to extract missing data features; Input the serialized discrete process variables into the pre-trained discrete data encoder to extract discrete data features; input the missing data features and discrete data features into the pre-trained second feature fuser to obtain fused missing data features; Input the fused missing data features into the pre-trained reconstruction decoder to obtain the reconstructed values of the continuous process variables.

6. The heterogeneous low-quality industrial time-series data anomaly detection method according to claim 1, characterized in that In step (4), use the training set for training to obtain the finally trained heterogeneous fusion detection network, which specifically includes: Input the serialized continuous process variables with missing measurement points and the serialized discrete process variables in the fine-tuning sample set in the training set into the heterogeneous fusion detection network to obtain the reconstructed values of the continuous process variables; Calculate the total loss function of the heterogeneous fusion detection network according to the mean square error between the reconstructed values of the continuous process variables and the corresponding continuous process variables in the fine-tuning sample set; Taking the minimization of the total loss function of the heterogeneous fusion detection network as the optimization objective, fine-tune the parameters of the missing data encoder, discrete data encoder, second feature fuser, and reconstruction decoder until the preset number of training rounds is reached or the total loss function of the heterogeneous fusion detection network converges, and obtain the finally trained heterogeneous fusion detection network.

7. The heterogeneous low-quality industrial time-series data anomaly detection method according to claim 1, wherein The estimation of the anomaly detection control limit based on the reconstruction error specifically includes: First, based on the samples in the fine-tuning sample set, calculate the reconstructed root mean square error corresponding to each sample according to the reconstructed values and the corresponding true values of the continuous process variables; Then, use the radial basis kernel function to calculate the probability density function of the reconstructed root mean square error; Finally, use the kernel density estimation method to calculate the anomaly detection control limit based on the probability density function of the reconstructed root mean square error.

Images