Industrial time series anomaly detection method based on chain prior attention constraint

Abstract

The application discloses an industrial time series anomaly detection method based on chain prior attention constraint. First, a training data set is constructed according to historical industrial time series under normal working conditions. Then, a data reconstruction model based on chain prior attention constraint is constructed, and the model is trained by using the training data set to obtain a trained data reconstruction model and a final training error. The detection threshold is determined according to the final training error. Finally, the real-time collected industrial time series is input into the trained data reconstruction model, the model outputs an anomaly score, and the anomaly detection result of the current industrial time series is generated in combination with the detection threshold. The application solves the problems of difficulty in time series feature extraction of the model and lack of prior knowledge, and further constructs an accurate time series feature extraction and reconstruction model, and is a reliable industrial time series anomaly detection method.

Description

Technical Field

[0001] This invention belongs to the field of industrial anomaly detection, and specifically relates to an industrial time series anomaly detection method based on chain-based prior attention constraints. Background Technology

[0002] Currently, with the continuous development of industrial technology, a massive amount of industrial time series data has been generated. Therefore, big data-driven anomaly detection of industrial time series based on neural networks is an important technological direction.

[0003] Attention mechanism is a widely used computational mechanism in neural networks. By assigning different weights to different parts of the input information, it enables the model to focus on the key information most relevant to the current task, thereby improving the expressive power of time series models and being widely used in industrial time series anomaly detection.

[0004] However, the relevant technologies still face challenges in modeling long-term dependencies and lack prior knowledge, which can lead to insufficient model representation of long-term temporal features. This affects the stability and reliability of the model in anomaly detection in complex industrial scenarios, and urgently needs to be addressed. Summary of the Invention

[0005] This invention provides an industrial time series anomaly detection method based on chain-based prior attention constraints to solve the problems of difficulty in extracting time series features from data and lack of prior knowledge in related technologies. It then constructs an accurate time series feature extraction and reconstruction model, forming a reliable industrial time series anomaly detection technology.

[0006] The technical solution of the present invention is as follows: A first aspect of this invention provides a method for detecting anomalies in industrial time series based on chained prior attention constraints, comprising the following steps: S1: Collect historical data from industrial equipment, devices, or systems, and construct a training dataset based on historical industrial time series under normal operating conditions; specifically, this includes the selection of monitoring variables reflecting the operating status of the device; the selection of historical data under normal operating conditions; data cleaning and preprocessing.

[0007] Through the above technical means, the embodiments of the present invention can realize the construction of high-quality historical datasets of normal working conditions, so as to meet the high-quality construction of subsequent data reconstruction models based on chain-based prior attention constraints, and thus meet the high-precision identification of industrial anomalies.

[0008] The embodiments of the present invention can realize the construction of high-quality historical datasets of normal working conditions, so as to meet the high-quality construction of subsequent data reconstruction models based on chain-like prior attention constraints, and thus meet the high-precision identification of industrial anomalies.

[0009] S2: Construct a data reconstruction model based on chained prior attention constraints and train the model using the training dataset to obtain the trained data reconstruction model and the final training error; S3: Determine the detection threshold based on the final training error; S4: Input the real-time collected industrial time series into the trained data reconstruction model. The model outputs anomaly scores, which are then combined with detection thresholds to generate anomaly detection results for the current industrial time series.

[0010] Optionally, in one embodiment of the present invention, the data reconstruction model based on chained prior attention constraints includes: The input of the data reconstruction model is used as the input of the embedding layer. The output of the embedding layer and the result of fusing with the corresponding position encoding are used as the input of the sequence decomposition layer. The sequence decomposition layer is used to decompose its input (i.e., time series embedding) into trend component features and seasonal component features. After the two data reconstruction units process the trend component features and seasonal component features respectively, they output the corresponding reconstructed data. Finally, the two reconstructed data are fused to obtain the final reconstructed data, which is used as the output of the data reconstruction model.

[0011] Optionally, in one embodiment of the present invention, the data reconstruction unit includes a hidden mapping layer and a decoding reconstruction layer connected by a chain-based prior attention constraint. The hidden mapping layer is used to map the input of the data reconstruction unit to the hidden space to obtain data features. The decoding reconstruction layer is used to transform the data features back to the original space using affine transformation to realize the reconstruction of the input of the data reconstruction unit.

[0012] Optionally, in one embodiment of the present invention, the hidden mapping layer based on chained prior attention constraints includes N sequentially connected chained prior attention modules.

[0013] Optionally, in one embodiment of the present invention, the chained prior attention module satisfies the following formula:

[0014]

[0015]

[0016]

[0017] Among them, Z nZ represents the features extracted by the nth layer of the chained prior attention module; n-1 This represents the feature extracted by the (n-1)th layer of the chained prior attention module; Represents the chained prior attention. Layer computation functions; Represents the chained attention structure matrix; Indicates the first The prior attention matrix of the layer; Indicates time Learnable weights; Indicates time neighborhood Learnable weights; Indicates time No. Characteristics of the layer; Indicates time No. Characteristics of the layer; Represents the first element in the chain attention structure matrix. Line number The value of the column; Represents the first element in the prior attention matrix. Line number The value of the column; Indicates time The set of neighboring nodes; It is a nonlinear transformation; Indicates time The scale parameter of the prior Cauchy Kernel for information propagation at other times, specifically, the scale parameter in the distribution constraint. It is calculated using a set of learnable weights; Indicates time and The absolute value of the difference.

[0018] The embodiments of the present invention can utilize forward attention constraints to fix the learning direction of the attention mechanism, thereby conforming to the inherent information propagation pattern of time series, so as to achieve better extraction of time series information, thereby ensuring effective learning and capture of normal patterns, and thus maintaining sensitivity to anomalies.

[0019] This is used to apply a Cauchy Kernel-based prior distribution constraint to attention, which is constructed based on the forgetting mechanism of memory over time and the long-tailed distribution characteristics of attention.

[0020] The embodiments of the present invention can utilize prior attention distribution constraints to guide the learning of attention weights with prior knowledge, thereby achieving accurate learning of dependencies between sequence points, thus realizing effective extraction of long-term sequence dependency features, thereby improving the effective capture of normal patterns, and enhancing the model's ability to identify abnormal sequences.

[0021] Optionally, in one embodiment of the present invention, the decoding and reconstruction layer includes a plurality of fully connected layers connected in sequence.

[0022] This invention can guide the learning direction of the attention mechanism by introducing chain-like attention structure constraints and prior attention distribution constraints, thereby solving the problems of difficulty in extracting time-series features of data and lack of prior knowledge. This improves the model's sensitivity to features of normal operating conditions, enhances the ability to identify anomalies, and enables the construction of a data reconstruction model based on chain-like prior attention constraints. This allows for reliable industrial time-series anomaly detection, meeting the high precision and reliability requirements for anomaly detection in modern industrial scenarios.

[0023] The embodiments of the present invention can utilize a chain-like prior attention module to achieve layer-by-layer extraction of temporal features, thereby enabling accurate extraction of temporal features under normal operating conditions. This allows the abnormal pattern features obtained by the model to have significant differences from normal patterns, thereby enhancing the score difference between normal and abnormal patterns and helping to accurately detect anomalies.

[0024] The embodiments of the present invention can utilize a data reconstruction model based on chain-based prior attention constraints to score the degree of anomaly of a sequence through the reconstruction error of the sequence, and rely on the advanced nature of chain-based prior attention constraints to ensure the use of information and accurate detection of anomalies.

[0025] The detection threshold is determined based on the training error during the model building process. The optimal threshold for detecting anomalies is determined by using sufficient training samples and combining the model training process.

[0026] The embodiments of the present invention can utilize the model construction process to reasonably determine the anomaly detection threshold, and rely on sufficient normal operating condition data to balance the false alarm rate and the detection rate, thereby ensuring the anomaly detection performance in actual industry.

[0027] This invention can construct a high-quality historical dataset of normal operating conditions and combine it with a neural network with chain-like prior attention constraints to learn a model that can effectively represent the features of normal operating conditions. This effectively improves the ability to extract temporal features under normal operating conditions, thereby effectively improving the ability to identify anomalies, and thus enabling the model to have a good detection effect on anomalies.

[0028] A second aspect of the present invention provides an industrial time series anomaly detection device based on chain-based prior attention constraints, comprising: The data acquisition module is used to acquire industrial time series data. The training dataset construction module is used to build training datasets based on historical industrial time series under normal operating conditions. The model building module is used to reconstruct a data model based on chained prior attention constraints, and to train the model using the training dataset to obtain the trained data reconstruction model and the final training error. A threshold generation unit is used to generate a detection threshold based on the final training error; The anomaly detection module is used to input the acquired real-time industrial time series into the trained data reconstruction model, and generate the anomaly detection results of the current industrial time series based on the anomaly score output by the model and the detection threshold.

[0029] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the industrial time series anomaly detection method based on chained prior attention constraints as described in the above embodiments.

[0030] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described industrial time series anomaly detection method based on chained prior attention constraints.

[0031] A fifth aspect of the present invention provides a computer program product, including a computer program that, when executed, is used to implement the above-described industrial time series anomaly detection method based on chained prior attention constraints.

[0032] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0033] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an industrial time series anomaly detection method based on chained prior attention constraints according to an embodiment of the present invention; Figure 2 This is a computational model diagram of a conventional attention mechanism according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a chain-like attention structure constraint according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the prior attention distribution constraint according to an embodiment of the present invention, wherein (a) is a numerical example of long-tail distribution under different scale parameters, and (b) is an attention distribution matrix of a sequence containing 96 time points obtained based on chain-like prior attention calculation; Figure 5 This is a schematic diagram of a chain-based prior attention module according to an embodiment of the present invention; Figure 6 This is a network structure diagram of a data reconstruction model based on chained prior attention constraints according to an embodiment of the present invention; Figure 7 This is a process flow diagram of a polypropylene production apparatus according to an embodiment of the present invention; Figure 8 This is a block diagram of an industrial time series anomaly detection device based on chain-based prior attention constraints according to an embodiment of the present invention. Figure 9 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0034] In the figure: 10-Industrial time series anomaly detection device based on chain-based prior attention constraint; 100-Data acquisition module; 200-Model building module; 300-Anomaly detection module; 901-Memory; 902-Processor; 903-Communication interface. Detailed Implementation

[0035] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0036] The following describes an industrial time series anomaly detection method based on chained prior attention constraints according to embodiments of the present invention with reference to the accompanying drawings.

[0037] To address the challenges of modeling long-term dependencies and lack of prior knowledge in the aforementioned related technologies, which leads to insufficient model representation of the dynamic characteristics and time-varying patterns of time series and thus affects the anomaly detection performance in complex industrial scenarios, this invention provides an industrial time series anomaly detection method based on chained prior attention constraints. In this technology, a chained prior attention calculation module is constructed through the combined effect of chained attention structure constraints and prior attention distribution constraints. This module serves as the core for building a data reconstruction model based on chained prior attention constraints, providing the model with inherent time-series dependencies as prior knowledge. This enables the model to accurately capture the dynamic characteristics of time series in complex scenarios, precisely grasp the essential characteristics of normal operating conditions, and enhance the model's sensitivity to anomalies, thereby improving actual anomaly detection performance. Simultaneously, it effectively improves the model's anomaly detection accuracy and robustness under complex operating conditions, achieving highly reliable anomaly monitoring in complex, nonlinear, and multivariate coupled industrial environments, meeting the high-precision requirements of actual industrial production for anomaly detection technology. This solves the practical problems of difficulty in modeling long-term dependencies and lack of prior knowledge in current related technologies, and further solves the problem of insufficient model representation of dynamic characteristics and time-varying laws of time series, which affects the anomaly detection performance of the model in complex industrial scenarios.

[0038] Specifically, this description first describes the process of an industrial time series anomaly detection method based on chain-based prior attention constraints provided by an embodiment of the present invention, and then provides technical details.

[0039] Figure 1 This is a flowchart illustrating an industrial time series anomaly detection method based on chained prior attention constraints provided in an embodiment of the present invention. Based on the flowchart, the industrial time series anomaly detection method based on chained prior attention constraints includes the following steps: In step S1, historical data of industrial equipment, devices or systems are collected, and data under normal operating conditions are selected to construct a dataset for training the model.

[0040] It is understood that the industrial time series anomaly detection method based on chain-based prior attention constraints in this embodiment of the invention is an unsupervised anomaly detection technique. It constructs an anomaly detection model through sufficient historical normal operating condition data. Therefore, the construction of a high-quality historical normal operating condition dataset for training the model is a key step.

[0041] In actual execution, the embodiments of the present invention can obtain historical data and historical operating status records of the monitored object, as well as relevant variable information that can accurately reflect the operating status of the monitored object, thereby realizing the construction of the training dataset.

[0042] For example, in embodiments of the present invention, a polypropylene production process dataset can be selected, and variables such as temperature, pressure, and flow rate that can reflect its actual operating status can be selected. Historical data and historical operating status records can be collected, and then data from a suitable time period can be selected to construct high-quality data under normal operating conditions for training the detection model, and a dataset can be constructed.

[0043] Furthermore, embodiments of the present invention can complete the construction of anomaly detection models and the determination of thresholds through the constructed dataset, laying the foundation for accurate model construction and anomaly detection.

[0044] In step S2, a data reconstruction model based on chained prior attention constraints is constructed and trained using the training dataset to obtain the trained data reconstruction model and the final training error. It is understood that the data reconstruction model construction process based on chain-based prior attention constraints in this embodiment of the invention is based on the normal operating condition historical dataset constructed in step S1. Its essence lies in reflecting the normal operating status of industrial equipment, devices or systems by learning the data features under normal operating conditions. Based on this, the constructed data reconstruction model based on chain-based prior attention constraints can effectively complete the reconstruction of data under normal operating conditions, and thus remain sensitive to data under unknown abnormal operating conditions, i.e., unable to complete effective reconstruction, and thereby detect anomalies.

[0045] Optionally, in one embodiment of the present invention, the data reconstruction model based on chained prior attention constraints includes: The input of the data reconstruction model is used as the input of the embedding layer. The output of the embedding layer and the result of fusing with the corresponding position encoding are used as the input of the sequence decomposition layer. The sequence decomposition layer is used to decompose its input (i.e., time series embedding) into trend component features and seasonal component features. After the two data reconstruction units process the trend component features and seasonal component features respectively, they output the corresponding reconstructed data. Finally, the two reconstructed data are fused to obtain the final reconstructed data, which is used as the output of the data reconstruction model.

[0046] Optionally, in one embodiment of the present invention, the data reconstruction unit includes a hidden mapping layer and a decoding reconstruction layer connected by a chain-based prior attention constraint. The hidden mapping layer is used to map the input of the data reconstruction unit to the hidden space to obtain data features; the decoding reconstruction layer is used to transform the data features back to the original space using affine transformation to realize the reconstruction of the input of the data reconstruction unit.

[0047] The core of the data reconstruction model based on chain-prior attention constraints includes two key processes: hidden layer mapping and decoding reconstruction. Hidden layer mapping is used to extract high-quality temporal features from the data, while decoding reconstruction is used to map these high-quality temporal features back to the original space. This process is performed separately on the sequence decomposition results based on global average pooling to maintain a decoupled understanding of seasonal and trend patterns, thereby achieving a more accurate grasp of temporal features. In addition, the data reconstruction model based on chain-prior attention constraints is built based on reconstruction loss constraints and also completes the anomaly scoring process based on reconstruction loss.

[0048] In actual implementation, the embodiments of the present invention can complete the feature extraction and decoding reconstruction process on actual industrial data, complete the model training based on reconstruction loss constraints, and complete the calculation on real-time data to obtain anomaly scores based on reconstruction loss.

[0049] For example, in a polypropylene production process, the embodiments of the present invention can use a data reconstruction model based on chain-prior attention constraints to train on a constructed polypropylene production process dataset, so that it can effectively extract features under normal operating conditions of the polypropylene production process; furthermore, since the trained data reconstruction model based on chain-prior attention constraints lacks awareness of abnormal operating conditions, it can maintain sensitivity to anomalies and thus identify anomalies.

[0050] In step S3, the detection threshold is determined based on the training error during the model building process.

[0051] It is understood that the embodiments of the present invention employ a data reconstruction model construction process based on chained prior attention constraints to determine the anomaly detection threshold. By introducing prior dependency constraints during the model training phase, attention weights are guided to propagate in a chain along the time series, thereby enhancing the model's ability to characterize the evolution of abnormal patterns. Based on this, the model can more accurately learn the statistical feature distribution of normal behavior and adaptively determine the anomaly detection threshold accordingly, effectively reducing the interference of noise fluctuations on threshold setting and improving the stability and reliability of anomaly detection results.

[0052] In actual implementation, this embodiment of the invention uses a data reconstruction model based on chain-based prior attention constraints to model time series features, and adaptively generates anomaly detection thresholds based on the anomaly score distribution obtained in the model reconstruction process, thereby achieving accurate determination of abnormal states.

[0053] For example, in a polypropylene production process, this embodiment of the invention uses the multi-source process variables in the production process as time series inputs, models the inherent temporal dependencies of the data through a data reconstruction model based on chain-based prior attention constraints, and then obtains an anomaly score reflecting changes in the production state. Based on the anomaly score, an anomaly detection threshold is adaptively determined, thereby achieving accurate identification of abnormal states in the polypropylene production process.

[0054] In step S4, the real-time collected industrial data is input into the constructed data reconstruction model based on chain-based prior attention constraints, and real-time anomaly detection is performed based on the set threshold.

[0055] It is understood that the embodiments of the present invention introduce chain-like prior attention constraints into the neural network to guide the model to focus on key features with continuous evolutionary relationships in industrial time series, thereby enhancing the model's ability to represent abnormal patterns and their evolution processes, and improving the accuracy and robustness of anomaly detection results.

[0056] In actual implementation, this embodiment of the invention uses a data reconstruction model based on chain-based prior attention constraints to perform online reasoning on the data continuously generated by the monitored industrial equipment, devices or systems. The anomaly score output by the model is compared with the anomaly detection threshold determined in step S3. When the anomaly score exceeds the threshold, the current industrial process is determined to be in an abnormal state.

[0057] For example, in a polypropylene production process, this embodiment of the invention uses process parameters such as reaction temperature, reaction pressure, material flow rate, and component concentration as time series inputs. Through a data reconstruction model based on chain-like prior attention constraints, it achieves real-time monitoring of abnormal operating conditions during the production process, thereby ensuring the safety and stability of the production process.

[0058] Figure 2 This is a computational model diagram of a traditional attention mechanism. This embodiment of the invention uses this computational model diagram to illustrate the weighted calculation principle of the traditional attention mechanism based on query vectors, key vectors, and value vectors, as well as its limitations in effectively characterizing long-term temporal dependencies and insufficient utilization of prior structural information when processing time-series data.

[0059] like Figure 2 As shown, the core idea of ​​the attention mechanism is to assign different weights to different features by calculating the correlation between input features, thereby highlighting information that is more important to the current task. In specific implementation, input features are typically mapped to query vectors (Query, Q), key vectors (Key, K), and value vectors (Value, V). Subsequently, the similarity between the query vector and the key vector is calculated, and the results are normalized to obtain the attention weights. The calculation formula is as follows:

[0060] in, This represents the attention mechanism; This represents the softmax function; This represents the dimension of the key vector, used to scale the inner product result to stabilize the training process. This represents the transpose operation. Finally, the output of the attention mechanism is obtained by performing a weighted summation on the value vector.

[0061] Understandably, the attention mechanism automatically determines which information is more important to the current decision by measuring the correlation between the features at the current moment and the features at previous moments, and assigns them higher weights, while giving lower weights to information with lower correlation. This allows the model to focus on key features from a large amount of time-series information, thereby improving the modeling effect.

[0062] In actual execution, attention mechanisms are usually embedded in neural network models to perform time-by-time calculations on the input time-series data.

[0063] For example, in the polypropylene production process, the attention mechanism can automatically assign higher weights to historical moments that have a greater impact on the current production state and lower weights to time information with lower correlation by calculating the correlation between features at different time steps. This highlights the time-series features that have a key impact on the polypropylene production process and provides a basis for subsequent abnormal state identification.

[0064] However, in the aforementioned polypropylene production process, traditional attention mechanisms typically perform independent calculations based on the correlation between the characteristics of the current moment and historical moments. This lacks prior constraints on the inherent temporal evolution of the process and easily overlooks the continuous propagation characteristics of abnormal states over time. Specifically, when the production process experiences slowly evolving or chain-like abnormal conditions, traditional attention mechanisms struggle to effectively characterize the transmission relationship of abnormal influences between adjacent time steps. This results in unstable attention weight allocation, and abnormal features are easily masked by fluctuations in normal operating conditions, thus affecting the accuracy and reliability of anomaly detection results.

[0065] To address the aforementioned issues, this invention proposes an industrial time series anomaly detection method based on chained prior attention constraints. By introducing chained prior constraints of the time series into the attention weight calculation process, the model is guided to continuously model key features along the time dimension, thereby enhancing its ability to characterize the anomaly evolution process. Based on this, the model can more stably extract time-series features related to the anomalous state and generate more reliable anomaly scores and detection results accordingly.

[0066] Through the above technical solutions, the embodiments of the present invention can effectively improve the anomaly detection performance in complex industrial scenarios such as polypropylene production processes, reduce the impact of noise interference on detection results, and provide strong support for the safe operation and stable control of industrial processes.

[0067] Optionally, in one embodiment of the present invention, the hidden mapping layer based on chained prior attention constraints includes N sequentially connected chained prior attention modules.

[0068] For each chained prior attention module, the output of the previous chained prior attention module is residually concatenated with the output of the current chained prior attention module, and then normalized to obtain the final output of the current chained prior attention module. The input of the hidden layer mapping layer is used as the input of the first chained prior attention module, and the output of the last chained prior attention module is used as the output of the hidden layer mapping layer.

[0069] Optionally, in one embodiment of the present invention, the chained prior attention module includes n fully connected layers, and the chained prior attention module satisfies the following formula:

[0070]

[0071]

[0072]

[0073] Among them, Z n Z represents the features extracted by the nth layer of the chained prior attention module; n-1 This represents the feature extracted by the (n-1)th layer of the chained prior attention module; Represents the chained prior attention. Layer computation functions; Represents the chained attention structure matrix; This represents the prior attention matrix of the nth layer; Indicates time Learnable weights; express neighborhood Learnable weights; Indicates time No. Characteristics of the layer; Indicates time No. Characteristics of the layer; Represents the first element in the chain attention structure matrix. Line number The value of the column; Represents the first element in the prior attention matrix. Line number The value of the column; Indicates time , neighboring nodes; It is a nonlinear transformation; Indicates time The scale parameter of the prior Cauchy Kernel for information propagation at other times, specifically, the scale parameter in the distribution constraint. It is calculated using a set of learnable weights; Indicates time and The absolute value of the difference.

[0074] In the above formula, the attention matrix A is used to impose forward attention constraints on attention, which are constructed based on the forward propagation characteristics of time-series data. T represents the number of samples. For example, for a time series of length 4, the chained attention structure matrix A satisfies the following formula:

[0075] The embodiments of the present invention can utilize forward attention constraints to fix the learning direction of the attention mechanism, thereby conforming to the inherent information propagation pattern of time series, so as to achieve better extraction of time series information, thereby ensuring effective learning and capture of normal patterns, and thus maintaining sensitivity to anomalies.

[0076] Figure 3 This is a schematic diagram of a chain attention structure constraint according to an embodiment of the present invention.

[0077] It is understood that, through the chain-like attention structure constraint, the neural network no longer considers the permutation invariant when calculating attention weights, but only considers the correlation between the features of the current time and the features of each historical time, and combines the prior continuous evolution information in the time series to ensure that the influence of the normal pattern between adjacent time steps can be effectively captured, thereby improving the model's ability to express normal working conditions and its sensitivity and stability to anomalies.

[0078] In actual implementation, this embodiment of the invention inputs real-time collected industrial time-series data into a neural network model containing chained attention structure constraints. Attention weights are calculated based on the chained structure constraints, and features at each time step are weighted and aggregated. Subsequently, by combining the anomaly score output by the model with a preset threshold, the system can determine in real time whether the industrial process is in an abnormal state.

[0079] For example, in the polypropylene production process, this embodiment of the invention inputs the collected process parameters such as reaction temperature, reaction pressure, material flow rate, and component concentration as time series data into a chain-like attention-constrained neural network model. By chaining attention weights along the time series, the model can highlight time segments that have a continuous impact on changes in key operating conditions, thereby achieving real-time identification of abnormal states and effectively ensuring the safety and stability of the production process.

[0080] This is used to apply a Cauchy Kernel-based prior distribution constraint to attention, which is constructed based on the forgetting mechanism of memory over time and the long-tailed distribution characteristics of attention. The aim is to guide the model to pay attention to features at important time steps while appropriately reducing the weight of more distant historical information.

[0081] Figure 4 This is a schematic diagram of the prior attention distribution constraint according to an embodiment of the present invention. Figure 4 (a) represents the long-tail distribution under different scale parameters. Figure 4 (b) is an attention distribution matrix obtained based on chain-based prior attention calculation.

[0082] It is understood that, through prior attention distribution constraints, the embodiments of the present invention enable the model to not only rely on the correlation of the input data itself when calculating attention weights, but also combine the memory decay characteristics and long-tail distribution features of time series, thereby achieving priority attention to normal features and reasonable decay of long-term dependent information, and improving the accuracy and robustness of anomaly detection.

[0083] In actual implementation, this embodiment of the invention inputs real-time collected industrial time-series data into a neural network model containing prior attention distribution constraints. The model calculates the attention weight for each time step according to the constraint rules, while reducing the weight of unimportant or historically distant features. Subsequently, by combining the anomaly score output by the model with a preset or adaptively generated threshold, real-time identification and alarm of industrial process anomalies are achieved.

[0084] For example, in the polypropylene production process, this embodiment of the invention inputs the collected time-series parameters such as reaction temperature, reaction pressure, and material flow rate into a neural network model that includes prior attention distribution constraints. By combining the memory-forgetting mechanism and the long-tail distribution characteristics of attention, the model can effectively suppress interference from irrelevant historical information far removed from the current moment while continuously focusing on key operating condition features, thereby achieving accurate identification of anomalies in the production process.

[0085] The embodiments of the present invention can utilize prior attention distribution constraints to guide the learning of attention weights with prior knowledge, thereby achieving accurate learning of dependencies between sequence points, thus realizing effective extraction of long-term sequence dependency features, thereby improving the effective capture of normal patterns, and enhancing the model's ability to identify abnormal sequences.

[0086] Nonlinear transformation The specific form depends on the actual application. Specifically, multiple nonlinear functions can be tried using a small batch of data, and the nonlinear function with the best performance can be selected for application. In this embodiment, the nonlinear transformation... It is a linear rectification function.

[0087] Optionally, in one embodiment of the present invention, the decoding and reconstruction layer includes a plurality of fully connected layers connected in sequence.

[0088] Figure 5 This is a schematic diagram of a chain-based prior attention module according to an embodiment of the present invention. Figure 5 As shown in the embodiment of the present invention, the chain-based prior attention module comprehensively utilizes chain-based attention structure constraints and prior attention distribution constraints, including layer-by-layer computation and residual connections between layers. Through this module, the model can perform weighted aggregation of time-series features in each layer, while maintaining the original input information through residual connections. This ensures that key features are not lost during multi-layer transmission and enhances the continuous modeling capability for normal pattern evolution.

[0089] It is understood that the chained prior attention module in this embodiment of the invention combines chained attention structure constraints with prior attention distribution constraints, enabling the neural network to capture the chained dependencies of key features in the time series while reasonably attenuating the weights on historical distant information when calculating attention weights. The layer-by-layer computation and residual connection design ensure that the model can retain the original feature information at a deep level, while gradually enhancing its ability to represent normal patterns, thereby improving the accuracy and robustness of anomaly detection.

[0090] In actual implementation, this embodiment of the invention inputs the collected industrial time series data into a chained prior attention module. The model calculates the attention weight for each time step layer by layer and accumulates feature representations through residual connections between layers. Based on the chained structure and prior constraints, the model can highlight continuous attention to key features while suppressing interference from irrelevant or distant historical information, thereby generating stable features.

[0091] For example, in the polypropylene production process, the embodiments of the present invention input the collected sequences of reaction temperature, pressure, material flow rate, etc., into a neural network containing a chain-like prior attention module. By combining chain-like structural constraints and prior attention distribution constraints, the continuous evolution of key operating conditions can be captured at multiple levels, thereby identifying abnormal states and realizing real-time monitoring of the production process, thus ensuring production safety and process stability.

[0092] Figure 6 This is a schematic diagram of the network structure of a data reconstruction model based on chain-based prior attention constraints according to an embodiment of the present invention. The core of this data reconstruction model based on chain-based prior attention constraints includes two key processes: hidden layer mapping and decoding reconstruction. Hidden layer mapping is used to extract high-quality temporal features from the data, while decoding reconstruction is used to map these high-quality temporal features back to the original space. This process is performed separately on the sequence decomposition results based on global average pooling to maintain a decoupled understanding of seasonal and trend patterns, thereby achieving a more accurate grasp of temporal features. Furthermore, the data reconstruction model based on chain-based prior attention constraints is constructed based on reconstruction loss constraints and also completes the anomaly scoring process based on the reconstruction loss.

[0093] Optionally, in a data reconstruction model based on chained prior attention constraints according to an embodiment of the present invention, a sequence decomposition method based on global average pooling is used to decompose the input time series embedding into trend components and seasonal components. By splitting the sequence features into different components, the model can model long-term changing trends and periodic fluctuation features respectively, providing a more refined feature representation for the subsequent chained prior attention module, thereby improving the sensitivity and accuracy of anomaly detection. The calculation formula is as follows:

[0094]

[0095] in, This indicates a moving average with a fill operation. This indicates a fill operation. Represents the embedding of time series, These represent trend component features and seasonal component features, respectively. In particular, in trend and seasonal component modeling, Corresponding to .

[0096] It is understood that, by performing trend and seasonal decomposition on the time series, the data reconstruction model based on chain-based prior attention constraints can focus on long-term trend changes and periodic fluctuations respectively when calculating attention weights. The trend component provides a global perception of overall changes, while the seasonal component emphasizes the details of periodic or recurring patterns, thereby enhancing the model's ability to capture normal operating patterns and improving the accuracy of anomaly identification in complex industrial processes.

[0097] Optionally, in one embodiment of the present invention, the data reconstruction model based on chained prior attention constraints includes: mapping the original input to a latent space by stacking multiple chained prior attention modules to extract key data features; subsequently, using affine transformation to transform the extracted data features back to the original space, thereby reconstructing the original time series. This structure can capture the evolutionary patterns of normal modes in the time series at a deep level, while preserving the original information, providing a stable basis for reconstruction error for anomaly detection.

[0098] It is understood that the embodiments of the present invention employ a data reconstruction model based on chain-based prior attention constraints. By stacking chain-based prior attention modules, the key features of the time series are encoded in the latent space, enabling the model to model time dependencies at a deep level. The latent features are then reconstructed back to the original space through affine transformation, allowing the model to generate reconstruction results that are highly consistent with the input sequence. By comparing the reconstruction error with a set threshold, the model can accurately determine abnormal states.

[0099] In actual implementation, this embodiment of the invention inputs real-time collected industrial time series data into a data reconstruction model based on chain-based prior attention constraints. The system extracts time series features layer by layer through stacked chain-based prior attention modules and performs representation learning in the latent space. Subsequently, affine transformation is used to map the latent features back to the original sequence space, the reconstruction error is calculated, and a threshold is combined for real-time anomaly detection, thereby enabling immediate identification and alarm of abnormal states during industrial process operation.

[0100] For example, in the polypropylene production process, this embodiment of the invention extracts latent features from the collected reaction temperature, pressure, and material flow rate data using a stacked chain-like prior attention module, and then reconstructs the original sequence through affine transformation. The model can generate significant reconstruction errors for abnormal operating conditions and generate stable anomaly scores accordingly, achieving accurate detection and real-time alarm for abnormal states in the production process, thereby ensuring the safety and stability of the production process.

[0101] Figure 7This is a process flow diagram of a polypropylene production apparatus according to an embodiment of the present invention. Polypropylene is a non-toxic, odorless, and tasteless milky-white crystalline polymer polymerized from propylene monomers. It is one of the most consumed plastics and an important raw material for the production of medical masks, and is also widely used in important sectors of the national economy such as textiles and medical supplies. Therefore, ensuring the safe, stable, and efficient operation of polypropylene production equipment is not only of paramount economic importance but also of crucial safety significance. This embodiment of the invention uses the actual polypropylene production process of a chemical plant as a verification case.

[0102] like Figure 7 The polypropylene production process flow shown is illustrated. The polymerization reaction is the core production stage of the Unipol propylene polymerization process, mainly consisting of a reactor, a circulating gas cooler, and a circulating gas compressor. Most of the gaseous stream exiting the top of the reactor is pressurized by the circulating gas compressor and then circulated back to the reactor via a heat exchanger to maintain normal fluidization within the reactor. A small portion goes to the product degassing chamber to remove inert components accumulated in the reactor gas phase, thus maintaining a stable concentration of gaseous components in the polymerization system. Liquid propylene from the propylene feed pump, along with triethylaluminum catalyst from the triethylaluminum feed pump and electron donor from the electron donor feed pump, continuously enter the reactor through the circulating gas loop at the outlet of the circulating gas cooler. Hydrogen continuously enters the reactor through the circulating gas loop at the inlet of the circulating gas compressor, and catalyst from the slurry catalyst feed pump is directly injected into the reactor.

[0103] Optionally, the polypropylene production process monitoring case described in this embodiment of the invention uses 17 variables reflecting its production operation status as shown in Table 1 for monitoring.

[0104] Table 1. Variables used in monitoring the polypropylene production process.

[0105] Furthermore, the polypropylene production process described in this embodiment of the invention can be monitored through steps S1 to S4.

[0106] The following examples illustrate the effectiveness of the industrial time series anomaly detection method based on chain-based prior attention constraints according to the present invention.

[0107] This invention uses the polypropylene production process in industrial production as the main application scenario, and combines two standard datasets to verify the effectiveness of the Chain Prior Constraint-based Time Series Anomaly Detection Method (CPCAD). Table 2 provides specific information on the three actual case datasets used: (1) soil telemetry detection satellite; (2) server dataset; (3) polypropylene production process dataset.

[0108] Table 2. Detailed information on the three real-world case datasets.

[0109] The industrial time series anomaly detection model based on chain-based prior attention constraints used in this embodiment of the invention is specifically configured as follows: the model's computational dimension is set to 8, two chain-based prior attention modules are used to encode and extract features from the input, and a two-layer fully connected neural network is used to decode and reconstruct the latent representation. The total training epochs are 10, with a learning rate of 0.001. Furthermore, this embodiment of the invention can be compared with models such as Isolation Forest (IF), Empirical Cumulative Distribution Anomaly Detector (ECOD), Transformer-based Anomaly Detector (TranAD), Convolutional Neural Network-based Autoencoder (CAE-M), and Recurrent Neural Network-based Anomaly Detector (OmniAnomaly). Precision, recall, F1 score, area under the ROC curve (AUCROC), and area under the PR curve (AUCPR) are used as evaluation metrics to verify the polarity of the industrial time series anomaly detection method based on chain-based prior attention constraints in this invention. Table 3 compares the anomaly detection results of the soil telemetry satellite dataset, Table 4 compares the anomaly detection results of the server dataset, and Table 5 compares the anomaly detection results of the polypropylene production process dataset.

[0110] Table 3 Comparison of anomaly detection results in soil remote sensing satellite datasets (%)

[0111] Table 4 Comparison of anomaly detection results for the server dataset (%)

[0112] Table 5 Comparison of anomaly detection results (%) in the polypropylene production process dataset

[0113] As can be seen, the industrial time series anomaly detection model based on chained prior attention constraints in this embodiment of the invention significantly outperforms the comparative model in all metrics across three different datasets, fully demonstrating the model's advantages in anomaly pattern capture and identification. Specifically, on the server dataset, the model of this embodiment achieves an F1 score of 95.02%, an AUCPR of 67.29%, and an AUCPR of 46.83%; on the polypropylene production process dataset, the F1 score reaches 96.04%, the AUCPR reaches 99.14%, and the AUCPR is 98.13%. These results demonstrate the high accuracy and reliability of the industrial time series anomaly detection model based on chained prior attention constraints in this embodiment of the invention for detecting anomalies in complex industrial processes across various fields.

[0114] In particular, compared with the best-performing comparative model, the embodiments of the present invention achieved an average F1 improvement of 3.48%, an AUCPR improvement of 10.78%, and an AUCPR improvement of 11.53% on three datasets, indicating that the chained prior attention constraint has a significant effect on capturing key temporal features in time series, suppressing noise interference, and improving model stability.

[0115] It is understood that, by combining chain structure constraints and prior attention distribution constraints, the embodiments of the present invention enable the model to fully consider the continuous evolution and long-term dependency characteristics of temporal patterns in time series, thereby maintaining high-performance anomaly detection capabilities under different industrial scenarios and data types.

[0116] In actual implementation, embodiments of the present invention can input real-time collected industrial time series data into the model for online monitoring and anomaly scoring calculation. By comparing with a set threshold, real-time identification and alarm of abnormal states can be achieved, ensuring the safe and stable operation of industrial processes.

[0117] For example, in the polypropylene production process, the model described in this embodiment of the invention can continuously model and aggregate features of time series data such as reaction temperature, pressure, and material flow rate. When anomalies occur, it generates significant reconstruction errors and anomaly scores, thereby achieving accurate early warning and real-time monitoring of abnormal states, providing reliable support for production safety and process optimization.

[0118] Next, with reference to the accompanying drawings, an industrial time series anomaly detection device based on chain-based prior attention constraints according to an embodiment of the present invention is described.

[0119] Figure 8 This is a block diagram of an industrial time series anomaly detection device based on chain-based prior attention constraints according to an embodiment of the present invention.

[0120] like Figure 8As shown, the industrial time series anomaly detection device 10 based on chain-based prior attention constraints includes: a data acquisition module 100, a model building module 200, and an anomaly detection module 300.

[0121] Among them, the data acquisition module 100 is used to collect and construct industrial normal operating condition datasets; The model building module 200 is used to build a data reconstruction model based on chained prior attention constraints and to determine the detection threshold through the model building process. The anomaly detection module 300 is used to calculate anomaly scores based on the constructed data reconstruction model based on chain-based prior attention constraints, and to detect anomalies in combination with a predetermined threshold.

[0122] Optionally, in one embodiment of the present invention, the data acquisition module 100 includes a data acquisition module and a training dataset construction module.

[0123] The data acquisition module is used to acquire industrial time series data. The training dataset building module is used to build training datasets based on historical industrial time series under normal operating conditions.

[0124] Optionally, in one embodiment of the present invention, the construction module 200 includes a model construction module and a threshold generation unit.

[0125] The model building module is based on a data reconstruction model with chained prior attention constraints, and uses the training dataset to train the model to obtain the trained data reconstruction model and the final training error. A threshold generation unit is used to generate a detection threshold based on the final training error; The construction unit is used to build a data reconstruction model based on chained prior attention constraints based on the first and second computing units, and to determine the detection threshold.

[0126] It should be noted that the foregoing explanation of the embodiment of the industrial time series anomaly detection method based on chain-based prior attention constraints also applies to the industrial time series anomaly detection device based on chain-based prior attention constraints in this embodiment, and will not be repeated here.

[0127] The industrial time series anomaly detection device based on chain-based prior attention constraints proposed in this invention utilizes a chain-based prior attention module that simultaneously leverages chain structure constraints and prior attention distribution constraints to perform layer-by-layer modeling and weighted aggregation of time series features. Through chain structure constraints, the model can capture the continuous evolution characteristics of time-series patterns in the time dimension, effectively characterizing normal patterns and thus achieving accurate anomaly identification. Through prior attention distribution constraints, the model combines memory decay characteristics and long-tail distribution features to reasonably reduce the weight of distant historical information, emphasizing the focus on key time step features, thereby suppressing noise interference and improving the stability of anomaly detection. Combining layer-by-layer computation and residual connections, the chain-based prior attention module can retain original feature information at a deeper level while strengthening the expressive power of normal patterns, thereby enhancing sensitivity to anomaly patterns. This enables the industrial time series anomaly detection device based on chain-based prior attention constraints proposed in this invention to maintain high-precision and robust anomaly identification capabilities in different industrial scenarios.

[0128] Figure 9 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device may include: The memory 901, the processor 902, and the computer program stored on the memory 901 and capable of running on the processor 902.

[0129] When the processor 902 executes the program, it implements the industrial time series anomaly detection method based on chained prior attention constraints provided in the above embodiments.

[0130] Furthermore, electronic devices also include: Communication interface 903 is used for communication between memory 901 and processor 902.

[0131] The memory 901 is used to store computer programs that can run on the processor 902.

[0132] The memory 901 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0133] If the memory 901, processor 902, and communication interface 903 are implemented independently, then the communication interface 903, memory 901, and processor 902 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0134] Optionally, in a specific implementation, if the memory 901, processor 902, and communication interface 903 are integrated on a single chip, then the memory 901, processor 902, and communication interface 903 can communicate with each other through an internal interface.

[0135] The processor 902 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0136] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described industrial time series anomaly detection method based on chained prior attention constraints.

[0137] This invention also provides a computer program product, including a computer program that can run computer instructions. When these computer instructions are executed by a processor, they implement the industrial time series anomaly detection method based on chained prior attention constraints provided in this invention.

[0138] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0139] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0140] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0141] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0142] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0143] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0144] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0145] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for detecting anomalies in industrial time series based on chained prior attention constraints, characterized in that, Includes the following steps: S1: Construct a training dataset based on historical industrial time series under normal operating conditions; S2: Construct a data reconstruction model based on chained prior attention constraints and train the model using the training dataset to obtain the trained data reconstruction model and the final training error; S3: Determine the detection threshold based on the final training error; S4: Input the real-time collected industrial time series into the trained data reconstruction model. The model outputs anomaly scores, which are then combined with detection thresholds to generate anomaly detection results for the current industrial time series.

2. The industrial time series anomaly detection method based on chain-based prior attention constraints according to claim 1, characterized in that, In S2, the data reconstruction model based on chained prior attention constraints includes: The input of the data reconstruction model is used as the input of the embedding layer. The output of the embedding layer and the result of fusing with the corresponding positional encoding are used as the input of the sequence decomposition layer. The sequence decomposition layer is used to decompose its input into trend component features and seasonal component features. The two data reconstruction units process the trend component features and seasonal component features respectively and output the corresponding reconstructed data. Finally, the two reconstructed data are fused to obtain the final reconstructed data, which is used as the output of the data reconstruction model.

3. The industrial time series anomaly detection method based on chained prior attention constraints according to claim 2, characterized in that, The data reconstruction unit includes a connected hidden mapping layer based on chained prior attention constraints and a decoding reconstruction layer. The hidden mapping layer is used to map the input of the data reconstruction unit to the hidden space to obtain data features. The decoding and reconstruction layer is used to transform data features back to the original space using affine transformation, thereby reconstructing the input of the data reconstruction unit.

4. The industrial time series anomaly detection method based on chained prior attention constraints according to claim 3, characterized in that, The hidden mapping layer based on chained prior attention constraints includes N sequentially connected chained prior attention modules.

5. The industrial time series anomaly detection method based on chained prior attention constraints according to claim 4, characterized in that, The chained prior attention module satisfies the following formula: Among them, Z n Z represents the features extracted by the nth layer of the chained prior attention module; n-1 This represents the feature extracted by the (n-1)th layer of the chained prior attention module; Represents the chained prior attention. Layer computation functions; Represents the chained attention structure matrix; This represents the prior attention matrix of the nth layer; Indicates time Learnable weights; Indicates time neighborhood Learnable weights; Indicates time No. Characteristics of the layer; Indicates time No. Characteristics of the layer; Represents the first element in the chain attention structure matrix. Line number The value of the column; Represents the first element in the prior attention matrix. Line number The value of the column; Indicates time The set of neighboring nodes; It is a nonlinear transformation; Indicates time The scale parameter of the prior Cauchy Kernel for information propagation at other times; Indicates time and The absolute value of the difference.

6. The industrial time series anomaly detection method based on chained prior attention constraints according to claim 3, characterized in that, The decoding and reconstruction layer comprises several fully connected layers connected in sequence.

7. An industrial time series anomaly detection device based on chain-based prior attention constraints, characterized in that, include: The data acquisition module is used to acquire industrial time series data. The training dataset construction module is used to build training datasets based on historical industrial time series under normal operating conditions. The model building module is used to reconstruct a data model based on chained prior attention constraints, and to train the model using the training dataset to obtain the trained data reconstruction model and the final training error. A threshold generation unit is used to generate a detection threshold based on the final training error; The anomaly detection module is used to input the acquired real-time industrial time series into the trained data reconstruction model, and generate the anomaly detection results of the current industrial time series based on the anomaly score output by the model and the detection threshold.

8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the industrial time series anomaly detection method based on chained prior attention constraints as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement an industrial time series anomaly detection method based on chained prior attention constraints as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement an industrial time series anomaly detection method based on chained prior attention constraints as described in any one of claims 1-6.

Images