Multi-monitoring-point anomaly detection method for chemical production process

Abstract

This invention belongs to the field of anomaly detection in multi-monitoring monitoring of interdisciplinary computer science and chemical engineering. It provides a method for anomaly detection at multiple monitoring points in chemical production processes, with the following steps: 1) capturing trend changes by weighting trend error and mean square error; 2) amplifying changes in abnormal and normal data using range statistics; 3) expressing the internal correlation between time series using the similarity of RP diagrams. The method of this invention has the following advantages: (1) innovative method; (2) improved results and efficiency in time series anomaly identification; (3) wider applicability.

Description

Technical Field

[0001] This invention belongs to the field of anomaly detection in multi-monitoring monitoring of the interdisciplinary field of computer science and chemical engineering. It relates to a method for anomaly detection at multiple monitoring points in chemical production processes, and the result is the detection of anomalies at multiple monitoring points in the chemical production process. Background Technology

[0002] Multi-point monitoring and analysis of chemical safety production processes refers to determining whether anomalies exist based on the deviation between measured values ​​and reference values. Historically, this task was performed by system monitoring experts who established normal behavior thresholds for each measured event / output. If a measured value exceeded the relevant expert-defined threshold, the system was considered not operating as expected. With the mature application of interconnected devices and sensors in cyber-physical systems (CPS) in chemical systems, expert-set thresholds no longer meet practical needs. More importantly, intelligent judgment is required to proactively analyze and predict unreasonable events, infer potential anomalies, thereby significantly reducing the possibility of accidents and promoting the digital transformation of society and enterprises.

[0003] Multivariate time series anomaly detection is a crucial part of multi-monitoring in chemical production processes, significantly impacting normal production in chemical enterprises and remaining a challenging task. Due to the inherent lack of labels in anomaly data and the complexity of anomaly types, anomaly detection is typically defined as unsupervised learning. Classic algorithms are broadly categorized into distance-based, statistical description-based, and density-based algorithms. Density-based algorithms are conceptually similar to distance-based algorithms, both measuring the degree of anomaly based on the Euclidean distance between monitoring node values ​​at a given moment and those from previous time periods. The difference lies in their focus: density-based algorithms prioritize the density of monitoring node values ​​at a given moment, while distance-based algorithms emphasize proximity. Statistical description-based methods determine anomalies by comparing the differences in statistical indicators (mean, variance, range, etc.) between the overall time series and local time points of the monitoring nodes.

[0004] Traditional algorithms suffer from the limitations of manually setting thresholds and failing to clearly measure the contextual dependencies within multivariate time series, neglecting the connections between local components. This weakens the algorithm's detection capabilities to some extent, preventing its effective deployment on production lines and hindering the development of intelligent anomaly detection.

[0005] Deep learning has achieved significant results in multivariate time series anomaly detection tasks, with methods broadly categorized into prediction-based and reconstruction-based approaches. Hundman et al. proposed a method called "Detecting spacecraft anomalies using LSTMS and nonparametric dynamic thresholding," attempting to predict normal values ​​of indicators based on historical data and detect anomalies based on prediction errors. Li et al. proposed a method called "Multivariate time series anomaly detection and interpretation using hierarchical inter-metric and temporal embedding," learning low-dimensional representations, reconstructing normal patterns in the data, and detecting anomalies based on reconstruction errors. However, the multivariates in chemical production processes exhibit strong spatiotemporal correlations. Such methods fail to capture potential privacy relationships between time series, disrupting the uniformity of multi-dimensional time series anomaly detection tasks, leading to numerous false alarms and large-scale invalid alarms, thus failing to meet the application needs of real-world scenarios.

[0006] To address the limitations of neural networks in multivariate time-series anomaly detection, graph neural networks (GNNs), as a successful method for modeling complex patterns in graph-structured data, have been introduced into this task. Graph convolutional networks are algorithms specifically designed for processing graph data. While generating useful feature representations of nodes in the graph network, they can also extract dependencies within the time series, with the correlation between time series primarily derived from the adjacency matrix of the GNN's input. However, chemical production is a complex, dynamic, and chaotic process. Currently, the adjacency matrix input to GNNs focuses on the correlation between parameters and the differences between reconstructed time series values, failing to reliably capture the delays, dynamics, and complex relationships between multivariate time-series data samples to learn effective embeddings and trend changes.

[0007] Based on an investigation and analysis of existing multivariate time series anomaly detection technologies, this invention reveals that graph neural networks directly using loss functions such as MSE and MAE to capture dependencies within time series data suffer from significant interference with the convergence and accuracy of models dealing with noisy and non-stationary variables in real-world production processes. Motivated by this, we designed a trend loss function and a similarity function combining RP and SSIM to address these issues. Experiments demonstrate that this invention can effectively identify time series anomalies, reduce false alarm rates, and improve accuracy, enabling deployment in multivariate time series anomaly monitoring systems for chemical production processes. Summary of the Invention

[0008] The purpose of this invention is to detect anomalies in multivariate time-series monitoring points of complex chemical production processes. It innovatively uses trend loss, RP plot similarity, and range statistics to achieve high efficiency and accuracy in anomaly detection. The main technical points are: 1) weighted summation of trend error and mean squared error to capture trend changes; 2) range statistics to amplify changes in abnormal and normal data; and 3) expressing the internal correlation between time series using RP plot similarity. The method described in this invention is applicable to complex systems similar to those in chemical processes, such as water plant anomaly attacks, power plant malfunctions, server attacks, and abnormal machinery operation.

[0009] The technical solution of the present invention:

[0010] A method for detecting anomalies at multiple monitoring points in a chemical production process, comprising the following steps:

[0011] Step 1: Construct a new time series similarity matrix

[0012] First, recursive plots (RP) are used to capture the delayed and nonlinear behaviors among M sensors (assuming the chemical production process involves M sensors), mapping the behavioral information of the M sensors onto images. Then, the similarity between images is calculated using the graph similarity function SSIM, forming a similarity matrix for the time series. This approach has the ability to accurately describe the delayed and nonlinear correlations between time series, which is of great significance for subsequent accurate anomaly prediction.

[0013] Step 2: Construct a Graph Neural Network (GCN) with trend loss.

[0014] First, the data from M sensors in the chemical production process and the similarity matrix are used as inputs to the GCN network; then, the GCN neural network is trained using our defined trend loss function, and the trained model parameters are saved.

[0015] Definition of trend loss function:

[0016]

[0017]

[0018]

[0019] Minimize the objective function trend loss

[0020] Where, y′ i+1 ,y′ i ,y′ i-1 Let y represent the predicted values ​​before time i, time i, and time i, respectively.i+1 ,y i ,y i-1 Let represent the true values ​​before time i, time i, and time i, respectively, and α be the weighting factor (hyperparameter).

[0021] Step 3: Anomaly Identification Process

[0022] After using the model trained in step 2 to predict unknown multivariate time series, anomalies are inferred. The specific process is as follows: First, using the model obtained in step 2, the values ​​collected by M sensors at each time point are predicted. Then, the average error μ between the predicted value and the true value is calculated. t and range R t The corresponding range value is obtained. (Effectively highlighting anomalies); Finally, the range M is estimated using the SOPT (Streaming POT) method. t The abnormal threshold is determined, and M is automatically judged at each time step based on the defined threshold. t Is the value abnormal?

[0023] Beneficial results of the present invention:

[0024] (1) Methodological innovation

[0025] The innovations of this invention are as follows: 1) We propose a novel graph neural network training method for multivariate time series in chemical production processes. To our knowledge, MADTC is the first graph convolutional neural network that can effectively capture the trend changes of multivariate time series and the implicit time dependencies and delays between multiple variables, learning robust representations of sensor data features and effectively alleviating the problem of overfitting in neural networks. 2) Based on the fluctuation patterns within the time series, we define a new trend change loss function to train the graph neural network, demonstrating significant improvements in both convergence speed and performance. 3) We propose for the first time a range statistic of reconstruction error that amplifies the difference between anomalies and normal errors, highlighting anomalies while maintaining low computational complexity. 4) We transform the graph time series into an RP graph and use structural similarity to calculate the correlation between time series. Through ablation experiments, we demonstrate that this similarity calculation method can more accurately describe the delays and correlations between time series than classic methods such as Euclidean similarity and mutual information entropy.

[0026] (2) Results and efficiency of time series anomaly identification

[0027] This invention, as the first of its kind, sets up a new trend loss function and RP image similarity in multivariate time-series anomaly detection, solving the problem of poor stability and accuracy of current algorithms due to the delays and dynamic correlations between multiple process variables. It effectively alleviates the difficulty of deploying intelligent detection in chemical industry scenarios. The method of this invention has strong robustness, effectively detecting time-series anomalies even with significant interference in the training data. Our experiments on real-world chemical industry data show significant results on three commonly used metrics (R, P, F1). Furthermore, compared to most algorithms that manually set thresholds for anomaly judgment, the automatic anomaly identification of this invention has significant value for deployment on actual production lines.

[0028] (3) Wider applicability

[0029] This invention is applicable to various types of multivariate time-series anomaly detection tasks. Its performance advantages are significant on the following types of data: a) Data with significant fluctuations, such as in chemical production processes and water supply systems, where sudden braking or cold starts cause sharp drops or rises in data, making anomaly identification a challenging task; b) Time series with high data dimensionality; c) Scenarios where complex delays and nonlinear correlations exist between multivariate time series processes; d) Situations where anomalous and normal data are difficult to distinguish. This, to a certain extent, fully demonstrates the generalizability and practicality of this invention under different data models. Attached Figure Description

[0030] Figure 1 This is a diagram of the algorithm structure of the present invention.

[0031] Figure 2 This is a demonstration of abnormal results in an embodiment of the present invention.

[0032] Figure 3 This is a graph showing the effect of the MAE loss function in fitting sensor data in a chemical process according to an embodiment of the present invention.

[0033] Figure 4 This is a graph showing the effect of the MSE loss function in fitting sensor data in a chemical process according to an embodiment of the present invention.

[0034] Figure 5 This is a graph showing the effect of the trend loss function in fitting sensor data in a chemical process according to an embodiment of the present invention. Specific implementation methods

[0035] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0036] The pseudocode for the algorithm is shown in the table below:

[0037]

[0038] Figure 1 This is a demonstration of the algorithm structure of the present invention. The figure above illustrates the complete process of the MADTC algorithm, where the first layer is the similarity calculation between time series (the similarity serves as the input to the adjacency matrix of the graph neural network), the second layer is the training and prediction of the graph neural network using the loss function we set, and the third layer is the process and result of automatic anomaly identification.

[0039] Figure 2 This is a demonstration of the automatic anomaly identification results of an embodiment of the present invention.

[0040] It is worth noting that in the second stage, compared to existing loss functions such as MAE and MSE, our defined loss function can more effectively fit the trend information of the time series. The effects of MAE, MSE, and trend loss on fitting sensor data in chemical processes are shown in the figures below. Figure 3 , Figure 4 and Figure 5 .

Claims

1. A multi-monitoring-point anomaly detection method for a chemical production process, characterized by, The steps are as follows: Step 1, Construct a new time series similarity matrix First, the RP is used to capture the delay, non-linear behavior between the sensors of a chemical production process and map the behavior information of the sensors to images; then, the SSIM is used to calculate the similarity between the images and form a time series similarity matrix. and map the behavior information of the sensors to images; then, the SSIM is used to calculate the similarity between the images and form a time series similarity matrix. Step 2, Construct a graph neural network with trend loss First, the chemical production process The sensor data and the similarity matrix are taken as the input of the GCN network. Then, the GCN neural network is trained using the trend loss function defined by us, and the trained model parameters are saved. Definition of trend loss function: ； ； ； Minimizing the objective function ; wherein, respectively represent the predicted value before the time point, the time point, and the predicted value after the time point, respectively represent the true value before the time point, the time point, and the true value after the time point, is a weighting factor; Step 3, Process of anomaly identification After using the model trained in step 2 to predict unknown multivariate time series, anomalies are inferred. The specific process is as follows: First, using the model obtained in step 2, the model is used to predict each time series. The system collects values ​​from several sensors, and then calculates the average error between the predicted and actual values. and range The corresponding range value is obtained. Finally, the range was estimated using the SOPT method. The abnormal threshold is determined, and the abnormal threshold is automatically used to judge the situation at each time point. Is the value abnormal?

Images