Factory equipment anomaly detection method based on time series prediction large model
Through the time series prediction large model based on Transformer architecture, the problem of insufficient modeling capabilities for complex time series data is solved, and accurate monitoring and abnormal detection of equipment operation status is realized, which improves the accuracy and timeliness of detection, and is suitable for Various industrial equipment.
Patent Information
- Application Number
- CN202411442721.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-10-16
AI Technical Summary
The existing technology lacks the modeling ability of complex time series data, and cannot effectively capture subtle changes in the operating state of the equipment, resulting in insufficient accuracy and timeliness of equipment abnormality detection and poor method versatility.
The time series prediction model based on Transformer architecture is adopted, and the complex mode of the device's operating state is extracted through high-dimensional feature acquisition units and anomaly detection and prediction units, combining the self-attention mechanism, cross-attention mechanism and dynamic normalization variance, and the complex mode of the device's operating state is extracted. Improve the accuracy and real-timeness of abnormal detection through multi-layer detection mechanism.
It greatly improves the accuracy and timeliness of equipment abnormality detection, is suitable for various types of industrial equipment, is not restricted by equipment types and operating environment, and enhances the detection capability under different working conditions.
Smart Images

Figure CN118965242A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of industrial production technology, and in particular to a method for detecting abnormality of plant equipment based on a time series prediction large model. Background Art
[0002] In modern industrial production, the normal operation of equipment is crucial to ensure production efficiency and safety. However, various abnormalities may occur in the equipment during long-term operation, such as failure, wear, aging, etc. If these abnormalities are not discovered and handled in time, they may cause serious production accidents and economic losses. Therefore, equipment anomaly detection has become an important research topic in industrial production.
[0003] Traditional equipment anomaly detection methods mainly rely on manual experience and regular inspections. These methods are not only time-consuming and labor-intensive, but also prone to missed detections and false detections. With the development of sensor technology and data acquisition technology, anomaly detection by monitoring the event sequence data of equipment operating status has become a new trend. However, when processing complex industrial equipment data, the existing event sequence prediction methods have insufficient modeling capabilities for complex time series data and cannot effectively capture subtle changes in the equipment operating status, resulting in insufficient accuracy and timeliness of anomaly detection, and poor method versatility.
[0004] Therefore, there is an urgent need for a factory equipment anomaly detection method based on a large time series prediction model. Summary of the invention
[0005] 1. Technical issues to be resolved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method for detecting anomalies of factory equipment based on a large time series prediction model, which solves the technical problems in the prior art of insufficient modeling capability for complex time series data, inability to effectively capture subtle changes in the operating status of equipment, resulting in insufficient accuracy and timeliness of anomaly detection, and poor method versatility.
[0006] (II) Technical solution In order to achieve the above object, the main technical solutions adopted by the present invention include: The embodiment of the present invention provides a method for detecting abnormality of plant equipment based on a time series prediction large model, comprising: S100, obtaining operation status data of the plant equipment to be detected, and performing data preprocessing on the operation status data of the plant equipment to be detected to obtain preprocessed data; S200, inputting the preprocessed data into a pre-trained time series prediction large model to obtain high-dimensional features, anomaly detection results and data prediction values corresponding to the preprocessed data; The time series prediction model includes: a high-dimensional feature acquisition unit and an anomaly detection and prediction unit.
[0007] Optionally, the method further comprises: S300, when the result of the anomaly detection is no anomaly, determine whether the data prediction value is within the upper and lower limit thresholds, if not, report the anomaly; if yes, proceed to step S400; S400, obtaining a prediction curve according to the data prediction value, and determining whether the prediction curve deviates from the historical curve. If so, proceeding to step S500; The historical curve is a curve corresponding to the preprocessed data; S500, linearly mapping the high-dimensional features corresponding to the preprocessed data to obtain a vector of a specific dimension, matching the vector of the specific dimension with the corresponding feature vector in the feature library constructed in advance one by one, obtaining the cosine similarity of the two vectors, and reporting an abnormality if the cosine similarity is less than a threshold set in advance; The vector of the specific dimension is a vector of dimension [1,1024]; The threshold value set in advance is set according to specific needs.
[0008] Optionally, the upper and lower thresholds are obtained according to parameter restriction boundaries, expert experience boundaries or through a dynamic sliding window statistics method.
[0009] Optionally, the pre-built feature library includes features of dimensions of [1,1024] corresponding to key data of different plant equipment under different working conditions and different time periods; The key data of the equipment in different plant areas under different working conditions and different time periods are determined by clustering method and specific needs of the plant areas; Among them, the key data of the different factory equipment under different working conditions and different time periods are input into the high-dimensional feature acquisition unit to obtain the corresponding high-dimensional features; the corresponding high-dimensional features are linearly mapped to obtain the [1,1024] dimensional features corresponding to the key data under different working conditions and different time periods.
[0010] Optionally, before S100, the step further includes: S000, using a training data set to train the time series prediction big model to obtain a trained time series prediction big model; The training data set includes: a normal operating state data training set and an abnormal operating state data set; The normal operating status data training set includes the normal operating data of each device in each factory; the abnormal operating status data set includes the abnormal operating data of each device in each factory.
[0011] Optionally, the S000 specifically includes: S010, inputting the data of each time series in the normal operating status data training set into the corresponding high-dimensional feature acquisition unit, using the MLM technology, randomly selecting a certain period of time in each iteration for mask processing, and performing model forward prediction; using the Adam optimizer, repeating the training for multiple rounds until the MSE loss is minimized; S020. Input the data of each time series in the normal operating status data training set and the abnormal operating status data training set into the corresponding abnormal data training set, use the mean square error as the prediction loss, the cross entropy as the anomaly detection loss, use Adam as the optimizer, and repeat the training for multiple rounds until the mean square error and the cross entropy are minimized.
[0012] Optionally, the S200 specifically includes: S210, inputting the preprocessed data into a high-dimensional feature acquisition unit to acquire high-dimensional features corresponding to the preprocessed data; S220: Input the high-dimensional features corresponding to the preprocessed data into an anomaly detection and prediction unit to obtain an anomaly detection result and a predicted data value.
[0013] Optionally, the high-dimensional feature acquisition unit uses a temporal Transformer architecture as a basic architecture, and the high-dimensional feature acquisition unit includes: a temporal data processing layer, a temporal data embedding layer, a learnable prompt embedding layer, a time correlation attention mechanism layer, and a spatial correlation attention mechanism layer.
[0014] Optionally, the anomaly detection and prediction unit includes: a high-dimensional time series feature extraction layer and a multi-task head layer.
[0015] Optionally, in S100, preprocessing the operating status data of the to-be-detected plant equipment includes: Invalid data in a non-operating state is deleted from the operating status data of the plant equipment to be detected, and the operating status data of the plant equipment from which the invalid data in the non-operating state is deleted is converted into a normal distribution interval to obtain preprocessed data.
[0016] (III) Beneficial effects The beneficial effects of the present invention are as follows: a method for detecting abnormalities in plant equipment based on a large model of time series prediction of the present invention, by using the self-attention mechanism of the Transformer architecture to extract the internal relationship features of the bit numbers, using the cross-attention mechanism to extract the relationship features between the bit numbers, and using prompt embedding to guide the data feature learning of different equipment and different working conditions, can better learn the relationship between data of different frequencies of different equipment in different factories; at the same time, a dynamic normalized variance and a dynamic attention scaling factor are proposed to dynamically process the data, so as to better weigh the data with different distributions, and greatly improve the accuracy of abnormality detection under different equipment and working conditions; the present invention proposes a time-related attention mechanism, which can give higher weights to closer time points when calculating the attention mechanism; in addition, the multi-layer detection mechanism of the present invention, considering wider coverage, can greatly improve the accuracy of abnormality detection, while enhancing real-time performance, and helping maintenance personnel to respond quickly; further, the method of the present invention is applicable to various types of industrial equipment, and is not limited by the type of equipment and the operating environment. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a flow chart of a method for detecting abnormality of plant equipment based on a time series prediction large model according to Embodiment 1 of the present invention; Figure 2 A flowchart of a method for detecting abnormality of plant equipment based on a time series prediction large model according to Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the time series prediction model of Example 2 of the present invention. DETAILED DESCRIPTION
[0018] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.
[0019] Time series data: Time series data, in the present invention, specifically refers to data with time attributes that is continuously collected by device sensors / manuals.
[0020] Bit number: the name of the data collected by different sensors, also called features in some places. In order to distinguish it from the hidden spatial features of the model, the present invention uses bit numbers to represent the names of data collected by different data sources.
[0021] The embodiment of the present invention proposes a method for detecting abnormalities of factory equipment based on a time series prediction large model. In order to solve the problems of insufficient modeling ability of complex time series data, poor model migration ability and scarce abnormal data in the prior art, the present invention adopts a time series prediction large model based on the Transformer architecture. Compared with other algorithms, it can effectively capture the complex patterns in the equipment operation data through hybrid pre-training and fine-tuning training, break the data barriers between different data distributions in different factories, and greatly improve the accuracy of abnormality detection; the present invention adopts a multi-layer detection mechanism, considers finer coverage and wider coverage, and greatly improves the accuracy of abnormality detection; in addition, by real-time monitoring of the equipment operation status and performing multi-level detection, the detected anomalies can be accurately classified, which is more conducive to business applications.
[0022] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0023] Example 1
[0024] See also Figure 1 , a method for detecting abnormality of plant equipment based on a time series prediction model in this embodiment includes: Step S100: acquiring the operating status data of the plant equipment to be detected, and performing data preprocessing on the operating status data of the plant equipment to be detected to obtain the preprocessed data; In this embodiment, the operating status data includes temperature data, pressure data, vibration data, etc.; in addition to the bit number of the equipment itself, such as temperature data, vibration data, pressure data, liquid level data, current data, valve opening data, etc., it can also be the bit number of the equipment input and output, such as inlet flow, outlet flow, etc.
[0025] Step S200: input the preprocessed data into a pre-trained time series prediction model to obtain high-dimensional features, anomaly detection results, and data prediction values corresponding to the preprocessed data; The time series prediction model includes: high-dimensional feature acquisition unit, anomaly detection and prediction unit.
[0026] The present invention can process a variety of data, has universality, and has a certain ability to automatically find correlations, and does not rely on artificially constructed bit number relationships.
[0027] In this embodiment, the operating status data of the equipment is collected in real time through sensors installed on the equipment or workers. The collected operating status data must eventually be uploaded to the data management platform, and the platform automatically converts the data format into the format of the specified model input.
[0028] For example, the data input into the time series prediction model needs to be in a normally distributed range. At this time, the data management platform will automatically normalize the data after data cleaning, and convert the cleaned data into a normally distributed range.
[0029] In the specific implementation process, data cleaning mainly refers to the collection data when the cleaning equipment is shut down, the unstable data during the sensor startup process, and other invalid data in the non-operating state.
[0030] When the result of anomaly detection is that an anomaly exists, an alarm is issued in a timely manner, and the detected anomaly is roughly classified. At the same time, corresponding processing suggestions are provided to help maintenance personnel respond and solve problems quickly.
[0031] The present embodiment provides a method for detecting abnormalities in factory equipment based on a time series prediction big model. By introducing the time series prediction big model technology based on the Transformer architecture, it solves the problems of insufficient modeling capabilities for complex time series data, poor model migration capabilities, and scarcity of abnormal data in the prior art, thereby achieving precise monitoring of the equipment operating status and abnormality detection, and improving the accuracy and timeliness of equipment abnormality detection.
[0032] Example 2
[0033] See also Figure 2 , a method for detecting abnormality of plant equipment based on a time series prediction model in this embodiment includes: Step S100: Acquire the operating status data of the plant equipment to be detected, and perform data preprocessing on the operating status data of the plant equipment to be detected to obtain the preprocessed data.
[0034] In this embodiment, the operating status data includes temperature data, pressure data, vibration data, etc.; in addition to the bit number of the equipment itself, such as temperature data, vibration data, pressure data, liquid level data, current data, valve opening data, etc., it can also be the bit number of the equipment input and output, such as inlet flow, outlet flow, etc.
[0035] Step S200: input the preprocessed data into a time series prediction model that has been trained in advance to obtain high-dimensional features, anomaly detection results, and data prediction values corresponding to the preprocessed data.
[0036] The time series prediction model includes: high-dimensional feature acquisition unit, anomaly detection and prediction unit.
[0037] Step S300: When the result of the anomaly detection is no anomaly, determine whether the data prediction value is within the upper and lower limit thresholds. If not, report the anomaly; if so, proceed to step S400.
[0038] The upper and lower thresholds are obtained based on parameter limit boundaries, expert experience boundaries, or through dynamic sliding window statistics.
[0039] Step S400, obtain the prediction curve according to the data prediction value, and determine whether the prediction curve deviates from the historical curve. If so, proceed to step S500.
[0040] The historical curve is the curve corresponding to the preprocessed data.
[0041] Step S500: linearly map the high-dimensional features corresponding to the preprocessed data to obtain a vector of a specific dimension, match the vector of the specific dimension with the corresponding feature vector in the feature library constructed in advance one by one, obtain the cosine similarity of the two vectors, and report an abnormality if the cosine similarity is less than a threshold set in advance.
[0042] The vector of a specific dimension is a vector of dimension [1,1024]; the threshold set in advance is set according to specific requirements.
[0043] In the specific implementation process, the logic of judging whether the data is abnormal is as follows: the anomaly detection model outputs the anomaly detection result and the data prediction value. If the anomaly detection result output at this time is that there is an anomaly, the anomaly is directly reported; if the anomaly detection result output is that there is no anomaly, it is necessary to combine the upper and lower thresholds to determine. If the device has a clear parameter limit boundary or expert experience boundary, the boundary is selected as the upper and lower thresholds. If not, the dynamic sliding window statistics method is used to determine the upper and lower thresholds. If the data prediction value exceeds the threshold, an early warning is reported; if it does not exceed the threshold, it is determined whether the prediction curve corresponding to the data prediction value deviates from the historical curve of the past period of time. Here, the data points of the historical curve and the prediction curve are clustered. If they do not belong to the same category, it means that there may be deviation. When the result does not deviate, it returns to normal and does not report the abnormality; if it deviates, it is necessary to compare the features and match them with the normal data of the bit number. If the cosine similarity is less than the threshold set in advance, the abnormality is reported, otherwise it returns to normal and does not report the abnormality.
[0044] The thresholds set in advance are set according to specific needs.
[0045] After reporting the abnormality, the alarm can be divided into real-time alarm and predictive alarm according to business needs. It can be classified into different process alarms such as temperature, flow, pressure, etc., can be classified into equipment alarms such as motor, electrical, mechanical, etc., and can also be classified into various indicator alarms to meet various business requirements.
[0046] In this embodiment, the feature library constructed in advance includes features of dimensions of [1,1024] corresponding to key data of different plant equipment under different working conditions and different time periods; The key data of different plant equipment under different working conditions and different time periods are determined through clustering methods and the specific needs of the plant; Among them, the key data of different factory equipment under different working conditions and different time periods are input into the high-dimensional feature acquisition unit to obtain the corresponding high-dimensional features; the corresponding high-dimensional features are linearly mapped to obtain the [1,1024] dimensional features corresponding to the key data under different working conditions and different time periods.
[0047] In the specific implementation process, the key data of different plant equipment under different working conditions and different time periods are the operating data during normal operation. And the feature library will be updated periodically.
[0048] In a specific implementation process, before step S100, the following steps are also included: Step S000: training the time series prediction model using a training data set to obtain a trained time series prediction model; The training data set includes: normal operation status data training set and abnormal operation status data set; The normal operation status data training set includes the normal operation data of each plant and each equipment; the abnormal operation status data set includes the abnormal operation data of each plant and each equipment.
[0049] In the specific implementation process, the training data set is the historical operating data of each device in each factory, including normal and abnormal data. The data comes from different equipment in different factories, and the data of each device is regarded as a set; the data is divided into normal data and abnormal data.
[0050] Step S000 specifically includes: Step S010: Input the data of each time series in the normal operation status data training set into the corresponding high-dimensional feature acquisition unit, use the MLM (masked language modeling) technology, randomly select a certain period of time in each iteration for masking, and perform model forward prediction; use the Adam optimizer, repeat the training for multiple rounds until the MSE loss is minimized; Among them, MLM is a commonly used pre-training technique in NLP. Given an input sequence, a part is randomly selected for masking, and then the model needs to predict the masked part based on the context of the sequence. In the pre-training of time series tasks, a part of the time series is randomly masked, and the masked part is predicted through the multivariate context sequence, thereby improving the feature extraction ability of the high-dimensional feature acquisition unit for time series.
[0051] In the specific training process, in each training iteration, a certain period of time is randomly selected for the multivariate time series sequence to perform masking to determine the part that needs to be predicted; then the model is forward predicted, and the MSE loss between the model output and the original unmasked true sequence is calculated; then the gradient is calculated through back propagation, and the Adam optimizer is selected to update the model parameters. Repeat the training for multiple rounds until the MSE loss is minimized, and then stop the training.
[0052] Step S020: input the data of each time series in the normal operating status data training set and the abnormal operating status data training set into the corresponding abnormal data training set, use the mean square error as the prediction loss, the cross entropy as the anomaly detection loss, use Adam as the optimizer, and repeat the training for multiple rounds until the mean square error and cross entropy are minimized.
[0053] During the training process of the anomaly detection and prediction unit, labeled data is required for anomaly detection tasks. The abnormal data is labeled, and normal data with similar data volume is selected for labeling. No labeling is required for prediction tasks.
[0054] The specific training process is the same as the training process of the high-dimensional feature acquisition unit. In each training iteration, data is input, the model is inferred, the loss is calculated, the gradient is calculated through backpropagation, the model is updated, and the iteration is repeated. The training is stopped after the goal is reached. The mean square error is used as the prediction loss, the cross entropy is used as the detection loss, and Adam is used as the optimizer.
[0055] In this embodiment, step S200 specifically includes: Step S210: input the preprocessed data into a high-dimensional feature acquisition unit to acquire high-dimensional features corresponding to the preprocessed data; Step S220: input the high-dimensional features corresponding to the preprocessed data into an anomaly detection and prediction unit to obtain anomaly detection results and predicted data values.
[0056] Among them, the high-dimensional feature acquisition unit uses the temporal Transformer architecture as the basic architecture, and the high-dimensional feature acquisition unit includes: a temporal data processing layer, a temporal data embedding layer, a learnable prompt embedding layer, a time correlation attention mechanism layer, and a spatial correlation attention mechanism layer.
[0057] The anomaly detection and prediction unit includes: a high-dimensional time series feature extraction layer and a multi-task head layer.
[0058] In this embodiment, the time embedding layer of the high-dimensional feature acquisition unit is used to add time embedding to each time step data to capture the time signal of the time series data, and the time embedding layer is a learnable linear embedding layer, which maps the timestamp to a fixed-dimensional vector space, whose dimension is consistent with the feature dimension after encoding the time series data, which facilitates feature splicing.
[0059] The time series data processing layer divides the time series data into blocks according to the set patch size, performs feature mapping, and uses sinusoidal position encoding to extract position features. Finally, the time features, data features, and position features are concatenated as the features of this patch data.
[0060] The learnable prompt embedding layer introduces a dataset-specific learnable prompt embedding to capture the correlation between different devices or signals in a unified dataset. All data features from the same dataset and different patches are concatenated with the same prompt embedding to enhance the generalization ability of the model.
[0061] The time-related attention mechanism layer uses the self-attention mechanism of the Transformer architecture to extract the internal relationship features of the bit numbers, and uses the cross-attention mechanism to extract the relationship features between the bit numbers. In order to better learn the relationship between different frequency data of different equipment in different factories, the present invention also proposes a dynamic normalized variance and a dynamic attention scaling factor to better balance data with different distributions. The method is to perform cluster analysis on the data, determine the variance and scaling factor according to the clustering category of the data, and different clustering categories can have different variances and scaling factors.
[0062] In order to better solve the problem of high-frequency and low-frequency signal mixing in industrial equipment, the high-dimensional feature acquisition unit also introduces a spatial correlation attention mechanism layer. When calculating the attention weight, higher weights are given to closer time points to solve the time dependency problem in high-frequency and low-frequency signal processing, thereby enhancing the high-dimensional feature acquisition model's ability to capture the dynamics of time series.
[0063] In this embodiment, see Figure 3 The network architecture diagram of the large model for time series prediction. Based on the direct or indirect correlation between devices, a time dimension encoder and a bit number dimension encoder are designed. Taking into account the use of data from different factories for mixed training, an independent prompt guidance is designed. Considering the shared features of multiple tasks and the independent operation of multiple tasks, a multi-head structure is designed.
[0064] In this embodiment, in step S100, preprocessing the operating status data of the plant equipment to be detected includes: Invalid data in a non-operating state is deleted from the operating status data of the plant equipment to be detected, and the operating status data of the plant equipment from which the invalid data in the non-operating state is deleted is converted into a normal distribution interval to obtain preprocessed data.
[0065] In this embodiment, a high-dimensional feature acquisition unit and anomaly detection and prediction unit based on the Transformer architecture are adopted. Compared with other algorithms, through hybrid pre-training and fine-tuning training, it can effectively capture the complex patterns in the equipment operation data, break the data barriers between different data distributions in different factories, and greatly improve the accuracy of anomaly detection; the present invention adopts a multi-layer detection mechanism, considering finer and wider coverage, and greatly improves the accuracy of anomaly detection; in addition, by real-time monitoring of the equipment operation status and performing multi-level detection, the detected anomalies can be accurately classified, which is more conducive to business applications.
[0066] Example 3
[0067] The present embodiment provides a method for detecting abnormalities in plant equipment based on a large time series prediction model. It is assumed that in a large manufacturing plant, the production line includes multiple key equipment, such as compressors, conveyor belts, motors, cooling systems, etc.
[0068] During the implementation process, the operating status data of the factory equipment is first obtained, such as data collected by temperature sensors installed on compressors, conveyor belts, motors and cooling systems, data collected by pressure sensors, data from vibration sensors, etc., and the above data is uploaded to the data processing platform; at the same time, the operating status of the equipment recorded by workers on a daily basis, such as operating hours, maintenance status, etc., is also uploaded to the data processing platform.
[0069] Then, the data in the shutdown state and the sensor startup process in the data processing platform are filtered out, the valid operation data is retained, and the valid operation data is normalized to convert the data into a value between 0 and 1 to ensure the consistency of the model input.
[0070] The processed data is input into the trained time series prediction model for prediction and monitoring, including algorithm prediction, prediction result exceeding the limit and feature matching failure.
[0071] When an abnormality occurs, real-time alarm and prediction alarm are issued, and the abnormal situation is classified into process alarms such as temperature, flow, pressure, and equipment alarms such as motor, electrical, and mechanical. In addition, the system provides processing suggestions based on the classified abnormal situation, such as checking sensors, adjusting equipment parameters, or performing equipment maintenance.
[0072] Through the above steps, potential problems can be discovered in advance through predictive alarms, avoiding downtime losses caused by equipment failures; and through the processing suggestions provided, maintenance personnel can quickly locate problems and take measures to shorten fault handling time.
[0073] Furthermore, early warning can avoid a large number of alarms that may affect on-site tracing when a fault occurs. In addition, by recording alarms and processing historical data, it helps factories optimize equipment maintenance strategies.
[0074] In a specific implementation, a factory equipment detection method for a large time series prediction model of the present invention uses a hybrid training technique when training the model. In addition to the versatility of the Transformer architecture itself, dynamic normalization and dynamic scaling are used in data processing, and time-related attention weighting is used in attention calculation, which can improve general capabilities and make the method of the present invention applicable to various types of industrial equipment without being restricted by equipment type and operating environment.
[0075] In this embodiment, by using a factory equipment anomaly detection method based on a time series prediction large model, the downtime caused by equipment failure is reduced, the operating efficiency and production capacity of the production line are improved; the safety and stability of equipment operation are improved, and the cost of equipment maintenance and replacement is reduced.
[0076] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.
[0077] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0078] In the present invention, unless otherwise clearly specified and limited, when a first feature is “on” or “below” a second feature, it may be that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Moreover, when a first feature is “above”, “above” or “above” a second feature, it may be that the first feature is directly above or obliquely above the second feature, or it may simply mean that the first feature is higher in level than the second feature. When a first feature is “below”, “below” or “below” a second feature, it may be that the first feature is directly below or obliquely below the second feature, or it may simply mean that the first feature is lower in level than the second feature.
[0079] In the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.
[0080] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.
Claims
1. A method for detecting abnormality of plant equipment based on a time series prediction model, characterized in that: include: S100, obtaining operation status data of the plant equipment to be detected, and performing data preprocessing on the operation status data of the plant equipment to be detected to obtain preprocessed data; S200, inputting the preprocessed data into a pre-trained time series prediction large model to obtain high-dimensional features, anomaly detection results and data prediction values corresponding to the preprocessed data; The time series prediction model includes: a high-dimensional feature acquisition unit and an anomaly detection and prediction unit.
2. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 1 is characterized in that: The method further comprises: S300, when the result of the anomaly detection is no anomaly, determine whether the data prediction value is within the upper and lower limit thresholds, if not, report the anomaly; if yes, proceed to step S400; S400, obtaining a prediction curve according to the data prediction value, and determining whether the prediction curve deviates from the historical curve. If so, proceeding to step S500; The historical curve is a curve corresponding to the preprocessed data; S500, linearly mapping the high-dimensional features corresponding to the preprocessed data to obtain a vector of a specific dimension, matching the vector of the specific dimension with the corresponding feature vector in the feature library constructed in advance one by one, obtaining the cosine similarity of the two vectors, and reporting an abnormality if the cosine similarity is less than a threshold set in advance; The vector of the specific dimension is a vector of dimension [1, 1024].
3. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 2 is characterized in that: The upper and lower thresholds are obtained according to parameter limit boundaries, expert experience boundaries or through dynamic sliding window statistics.
4. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 2, characterized in that: The pre-built feature library includes features of dimensions [1,1024] corresponding to key data of different plant equipment under different working conditions and different time periods; The key data of the equipment in different plant areas under different working conditions and different time periods are determined by clustering method and specific needs of the plant areas; Among them, the key data of the different factory equipment under different working conditions and different time periods are input into the high-dimensional feature acquisition unit to obtain the corresponding high-dimensional features; the corresponding high-dimensional features are linearly mapped to obtain the [1,1024] dimensional features corresponding to the key data under different working conditions and different time periods.
5. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 1, characterized in that: The S100 also includes: S000, using a training data set to train the time series prediction big model to obtain a trained time series prediction big model; The training data set includes: a normal operating state data training set and an abnormal operating state data set; The normal operating status data training set includes the normal operating data of each device in each factory; the abnormal operating status data set includes the abnormal operating data of each device in each factory.
6. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 5 is characterized in that: The S000 specifically includes: S010, inputting the data of each time series in the normal operating status data training set into the corresponding high-dimensional feature acquisition unit, using the MLM technology, randomly selecting a certain period of time in each iteration for mask processing, and performing model forward prediction; using the Adam optimizer, repeating the training for multiple rounds until the MSE loss is minimized; S020. Input the data of each time series in the normal operating status data training set and the abnormal operating status data training set into the corresponding abnormal data training set, use the mean square error as the prediction loss, the cross entropy as the anomaly detection loss, use Adam as the optimizer, and repeat the training for multiple rounds until the mean square error and the cross entropy are minimized.
7. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 1, characterized in that: The S200 specifically includes: S210, inputting the preprocessed data into a high-dimensional feature acquisition unit to acquire high-dimensional features corresponding to the preprocessed data; S220: Input the high-dimensional features corresponding to the preprocessed data into an anomaly detection and prediction unit to obtain an anomaly detection result and a predicted data value.
8. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 7, characterized in that: in, The high-dimensional feature acquisition unit uses the temporal Transformer architecture as the basic architecture, and the high-dimensional feature acquisition unit includes: a temporal data processing layer, a temporal data embedding layer, a learnable prompt embedding layer, a time correlation attention mechanism layer, and a spatial correlation attention mechanism layer.
9. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 7, characterized in that: The anomaly detection and prediction unit includes: a high-dimensional time series feature extraction layer and a multi-task head layer.
10. The method for detecting abnormality of plant equipment based on a time series prediction model according to claim 1, characterized in that: In S100, preprocessing the operating status data of the plant equipment to be detected includes: Invalid data in a non-operating state is deleted from the operating status data of the plant equipment to be detected, and the operating status data of the plant equipment from which the invalid data in the non-operating state is deleted is converted into a normal distribution interval to obtain preprocessed data.
Citation Information
Patent Citations
Abnormality detection method, system and equipment for time series data and storage medium
CN114386521A
Transform anomaly detection method based on spatio-temporal characteristics
CN115618196A
Equipment anomaly detection method and device, electronic equipment and storage medium
CN116628563A
Method and device for determining equipment state and electronic equipment
CN116661397A
Time sequence anomaly detection method based on self-supervised learning
CN116994772A
