Data-driven industrial equipment sign prediction method, device, equipment and medium

Abstract

The present application relates to the technical field of industrial equipment sign prediction, and discloses a data-driven industrial equipment sign prediction method, device and medium. The method comprises: real-time sampling of the operating status of industrial equipment to obtain multivariate time series data; encoding and decoding the multivariate time series data through a time series reconstruction model to obtain a reconstruction output; performing difference calculation on the multivariate time series data and the reconstruction output to obtain a reconstruction error set including multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors; scoring the reconstruction error set to obtain anomaly scores, wherein the anomaly scores include: multiple multivariate feature anomaly scores and multiple univariate feature anomaly scores; threshold estimation of the anomaly scores to obtain anomaly thresholds, comparing and judging the anomaly scores and the anomaly thresholds to obtain an anomaly determination result of the current state. The present application improves the accuracy of data-driven industrial equipment sign prediction.

Description

Technical Field

[0001] The present application relates to the field of industrial equipment sign prediction, and in particular to a data-driven industrial equipment sign prediction method, device, equipment and medium. Background Art

[0002] During long-term operation, the performance of industrial equipment will gradually degrade, which may lead to failures. Traditional equipment monitoring and maintenance methods mainly rely on threshold monitoring and rule constraints. These methods determine whether the equipment is in an abnormal state by setting fixed thresholds or predefined rules. For example, when a monitoring parameter exceeds a preset threshold, the system will issue an alarm. In addition, some advanced methods will combine expert experience to establish a more complex rule system to improve the accuracy of anomaly detection.

[0003] However, these traditional methods have some limitations. First, they are difficult to adapt to complex and changing industrial environments. Fixed thresholds and rules cannot effectively capture the dynamic changes of equipment under different working conditions, which can easily lead to false positives or false negatives. Second, these methods can usually only detect known types of faults, and the detection effect is poor for new or complex fault modes. Furthermore, traditional methods are often passive and can only issue alarms when a fault has occurred or is about to occur, and cannot predict potential problems in advance. Summary of the invention

[0004] The present application provides a data-driven industrial equipment sign prediction method, apparatus, device and medium for improving the accuracy of data-driven industrial equipment sign prediction.

[0005] In a first aspect, the present application provides a data-driven industrial equipment sign prediction method, the data-driven industrial equipment sign prediction method comprising: real-time sampling of the operating status of the industrial equipment to obtain multivariate time series data including multiple univariate features and multiple multivariate features; encoding and decoding the multivariate time series data through a preset time series reconstruction model to obtain a reconstruction output; performing difference calculation on the multivariate time series data and the reconstruction output to obtain a reconstruction error set including multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors; scoring the reconstruction error set to obtain an abnormality score, wherein the abnormality score includes: multiple multivariate feature abnormality scores and multiple univariate feature abnormality scores; threshold estimating the abnormality score to obtain an abnormality threshold, comparing and judging the abnormality score and the abnormality threshold to obtain an abnormality determination result of the current state; when the abnormality determination result is abnormal, respectively predicting multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors in a future time period to obtain a predicted reconstruction error time series set, specifically comprising: constructing a prediction model for a radial basis function neural network model to obtain a multivariate feature reconstruction error prediction model; A prediction model is constructed for an M-dimensional radial basis function neural network model to obtain a unary feature reconstruction error prediction model, wherein M is the number of unary features and M is a positive integer greater than 1; The multivariate feature reconstruction errors in the reconstruction error set are sorted in time series to obtain a multivariate feature reconstruction error time series; The multivariate feature reconstruction error prediction model is trained through the multivariate feature reconstruction error time series to obtain a trained multivariate feature prediction model; The multiple unary feature reconstruction errors in the reconstruction error set are sorted in time series respectively to obtain M unary feature reconstruction error time series; The unary feature reconstruction error prediction model is trained through the M unary feature reconstruction error time series to obtain a trained unary feature prediction model; The multivariate feature reconstruction errors in a future time period are predicted through the trained multivariate feature prediction model to obtain a predicted multivariate feature reconstruction error time series; The M unary feature reconstruction errors in a future time period are predicted through the trained unary feature prediction model to obtain M predicted unary feature reconstruction error time series; The predicted multivariate feature reconstruction error time series and the M predicted univariate feature reconstruction error time series are combined to obtain a predicted reconstruction error time series set; statistical feature calculation is performed on each time series in the predicted reconstruction error time series set to obtain a statistical feature result set; stationarity judgment is performed on each statistical feature result in the statistical feature result set to obtain a stationarity judgment result set; when there is an unsteady result in the stationarity judgment result set, anomaly score calculation is performed on the corresponding predicted reconstruction error time series to obtain an anomaly score set;Perform feature matching on the anomaly score set to obtain a sign feature set. ;

[0006] In a second aspect, the present application provides a data-driven industrial equipment sign prediction device, the data-driven industrial equipment sign prediction device comprising:

[0007] An acquisition module is used to sample the operating status of industrial equipment in real time to obtain multivariate time series data containing multiple univariate features and multiple multivariate features;

[0008] A decoding module, used for encoding and decoding the multivariate time series data through a preset time series reconstruction model to obtain a reconstructed output;

[0009] A calculation module, used for performing difference calculation on the multivariate time series data and the reconstructed output to obtain a reconstruction error set including a plurality of univariate feature reconstruction errors and a plurality of multivariate feature reconstruction errors;

[0010] A scoring module, used for scoring the reconstruction error set to obtain an anomaly score, wherein the anomaly score includes: a plurality of multivariate feature anomaly scores and a plurality of univariate feature anomaly scores;

[0011] A judgment module is used to perform a threshold estimation on the abnormal score to obtain an abnormal threshold, compare and judge the abnormal score and the abnormal threshold to obtain an abnormal judgment result of the current state;

[0012] A prediction module is used to predict multiple unary feature reconstruction errors and multiple multivariate feature reconstruction errors in a future time period respectively when the abnormality judgment result is abnormal, and obtain a predicted reconstruction error time series set, specifically including: constructing a prediction model for a radial basis function neural network model to obtain a multivariate feature reconstruction error prediction model; constructing a prediction model for an M-dimensional radial basis function neural network model to obtain a unary feature reconstruction error prediction model, wherein M is the number of unary features and M is a positive integer greater than 1; arranging the multivariate feature reconstruction errors in the reconstruction error set in time series to obtain a multivariate feature reconstruction error time series; training the multivariate feature reconstruction error prediction model through the multivariate feature reconstruction error time series to obtain a trained multivariate feature prediction model; arranging the multiple unary feature reconstruction errors in the reconstruction error set in time series respectively to obtain M unary feature reconstruction error time series; training the unary feature reconstruction error prediction model through the M unary feature reconstruction error time series to obtain a trained unary feature prediction model; Predicting the multivariate feature reconstruction error in a future time period by the trained multivariate feature prediction model to obtain a predicted multivariate feature reconstruction error time series; Predicting M univariate feature reconstruction errors in a future time period by the trained univariate feature prediction model to obtain M predicted univariate feature reconstruction error time series; Combining the predicted multivariate feature reconstruction error time series with the M predicted univariate feature reconstruction error time series to obtain a predicted reconstruction error time series set;

[0013] A statistical module, used for calculating the statistical characteristics of each time series in the predicted reconstruction error time series set to obtain a statistical characteristic result set;

[0014] An analysis module, used for performing a stationarity judgment on each statistical feature result in the statistical feature result set to obtain a stationarity judgment result set;

[0015] A screening module, used for calculating anomaly scores on corresponding prediction reconstruction error time series when there are non-stationary results in the stationarity judgment result set, to obtain anomaly score set;

[0016] The matching module is used to perform feature matching on the anomaly score set to obtain a symptom feature set.

[0017] The third aspect of the present application provides a data-driven industrial equipment sign prediction device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the data-driven industrial equipment sign prediction device executes the above-mentioned data-driven industrial equipment sign prediction method.

[0018] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned data-driven industrial equipment symptom prediction method.

[0019] In the technical solution provided by the present application, by real-time sampling of the operating status of industrial equipment, multivariate time series data containing multiple univariate features and multiple multivariate features are obtained, and these data are encoded and decoded using a preset time series reconstruction model to obtain a reconstructed output, which effectively captures the potential patterns and structures in the data. By performing difference calculation on the multivariate time series data and the reconstructed output, a reconstruction error set containing multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors is obtained, which can accurately quantify the deviation between the original data and the reconstructed data, and provide a key indicator for anomaly detection. The reconstruction error set is scored to obtain an anomaly score including multiple multivariate feature anomaly scores and multiple univariate feature anomaly scores. This multi-dimensional scoring method can fully reflect the degree of abnormality of the equipment state. Subsequently, the anomaly score is threshold estimated to obtain an anomaly threshold, and by comparing the anomaly score and the anomaly threshold, the anomaly determination result of the current state is obtained. This process realizes the accurate judgment of the equipment state. The advantage of this method is that it can simultaneously consider the anomaly of a single parameter (univariate feature) and the complex relationship between multiple parameters (multivariate feature), which greatly improves the accuracy and comprehensiveness of anomaly detection. Through real-time data processing and analysis, it is possible to promptly detect minor anomalies in equipment operation, making preventive maintenance possible. It can automatically adjust the abnormal threshold according to different equipment and working conditions, reducing human intervention and improving the intelligence level of the system. In general, this method achieves comprehensive monitoring and accurate prediction of the status of industrial equipment through data-driven methods, effectively improving the reliability and operating efficiency of equipment, while reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0021] Figure 1 A schematic diagram of an embodiment of a data-driven industrial equipment sign prediction method in an embodiment of the present application;

[0022] Figure 2 It is a schematic diagram of an embodiment of a data-driven industrial equipment sign prediction device in an embodiment of the present application. DETAILED DESCRIPTION

[0023] The embodiments of the present application provide a data-driven industrial equipment symptom prediction method, device, equipment and medium. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data passed in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the data-driven industrial equipment sign prediction method includes:

[0025] Step S101, sampling the operating status of industrial equipment in real time to obtain multivariate time series data including multiple univariate features and multiple multivariate features;

[0026] Step S102, encoding and decoding the multivariate time series data through a preset time series reconstruction model to obtain a reconstructed output;

[0027] Step S103, performing difference calculation on the multivariate time series data and the reconstructed output to obtain a reconstruction error set including a plurality of univariate feature reconstruction errors and a plurality of multivariate feature reconstruction errors;

[0028] Step S104: scoring the reconstruction error set to obtain an anomaly score, wherein the anomaly score includes: a plurality of multivariate feature anomaly scores and a plurality of univariate feature anomaly scores;

[0029] Step S105: perform threshold estimation on the abnormality score to obtain an abnormality threshold, perform comparison and judgment on the abnormality score and the abnormality threshold to obtain an abnormality judgment result of the current state.

[0030] It is understandable that the execution subject of the present application may be a data-driven industrial equipment sign prediction device, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0031] Specifically, the data-driven industrial equipment sign prediction method first obtains multivariate time series data containing multiple univariate features and multiple multivariate features by sampling the operating status of industrial equipment in real time. Univariate features refer to a single monitoring parameter, such as temperature, pressure, or vibration; multivariate features are comprehensive features that consider the relationship between multiple parameters. For example, in motor monitoring, the speed can be used as a univariate feature, and the relationship between the speed and the current can be used as a multivariate feature. The sampling frequency is set according to the equipment characteristics and prediction requirements, usually ranging from several times per second to several times per minute. Next, the collected multivariate time series data is encoded and decoded using a preset time series reconstruction model to obtain a reconstructed output. The time series reconstruction model is an algorithm that can capture the intrinsic structure and dynamic characteristics of the data, and common ones include autoencoders or recurrent neural networks. The encoding process compresses the original data into a low-dimensional space, while the decoding process attempts to reconstruct the original data from this low-dimensional representation. The purpose of this step is to learn the normal pattern of the data and lay the foundation for subsequent anomaly detection.

[0032] Subsequently, the difference between the original multivariate time series data and the reconstructed output is calculated to obtain a set of reconstruction errors including multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors. For univariate features, the reconstruction error is usually calculated using the absolute difference or relative difference between the original value and the reconstructed value. For multivariate features, the correlation between features needs to be considered, and methods such as the Mahalanobis distance are usually used to calculate the reconstruction error. The purpose of this step is to quantify the difference between the original data and the reconstructed data, so as to capture potential anomalies. Next, the reconstruction error set is scored to obtain anomaly scores. The anomaly score includes multiple multivariate feature anomaly scores and multiple univariate feature anomaly scores. The scoring process usually involves converting the reconstruction error into a standardized score for easy comparison and judgment. For univariate features, commonly used methods include z-score or percentile methods. For multivariate features, more complex statistical methods may need to be considered, such as scoring methods based on probability distribution. The purpose of this step is to convert the reconstruction error into an interpretable and comparable abnormality degree indicator.

[0033] Finally, the anomaly score is estimated by threshold, the anomaly threshold is obtained, and the anomaly score is compared with the anomaly threshold to obtain the anomaly judgment result of the current state. Threshold estimation is usually based on historical data and statistical methods, such as the triple standard deviation method or the quantile-based method. The comparison and judgment process is to compare the current anomaly score with the estimated threshold. If it exceeds the threshold, it is judged as an abnormal state.

[0034] For example, in an industrial cooling system, monitoring parameters include coolant temperature, flow rate, pressure, etc. The system collects data once a minute to obtain the time series of these parameters. After processing through the reconstruction model, it is found that the reconstruction error of the coolant temperature gradually increases, while the flow rate remains normal. After scoring processing, the temperature anomaly score reaches 0.95 (assuming the score range is 0 to 1), exceeding the preset 0.9 threshold. At the same time, the multivariate feature anomaly score of temperature and pressure also reaches 0.92. This indicates that there may be potential temperature anomalies in the cooling system.

[0035] In the embodiment of the present application, by real-time sampling of the operating status of industrial equipment, multivariate time series data containing multiple univariate features and multiple multivariate features are obtained, and these data are encoded and decoded using a preset time series reconstruction model to obtain a reconstruction output, which effectively captures the potential patterns and structures in the data. By performing difference calculation on the multivariate time series data and the reconstruction output, a reconstruction error set containing multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors is obtained, which can accurately quantify the deviation between the original data and the reconstructed data, and provide a key indicator for anomaly detection. The reconstruction error set is scored to obtain an abnormal score including multiple multivariate feature abnormal scores and multiple univariate feature abnormal scores. This multi-dimensional scoring method can fully reflect the degree of abnormality of the equipment state. Subsequently, the abnormal score is threshold estimated to obtain an abnormal threshold, and by comparing the abnormal score and the abnormal threshold, the abnormal judgment result of the current state is obtained. This process realizes the accurate judgment of the equipment state. The advantage of this method is that it can simultaneously consider the abnormality of a single parameter (univariate feature) and the complex relationship between multiple parameters (multivariate feature), which greatly improves the accuracy and comprehensiveness of anomaly detection. Through real-time data processing and analysis, it is possible to promptly detect minor anomalies in equipment operation, making preventive maintenance possible. It can automatically adjust the abnormal threshold according to different equipment and working conditions, reducing human intervention and improving the intelligence level of the system. In general, this method achieves comprehensive monitoring and accurate prediction of the status of industrial equipment through data-driven methods, effectively improving the reliability and operating efficiency of equipment, while reducing maintenance costs.

[0036] In a specific embodiment, the data-driven industrial equipment sign prediction method may further specifically include the following steps:

[0037] (1) When the anomaly judgment result is abnormal, multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors in the future time period are predicted respectively to obtain a set of predicted reconstruction error time series;

[0038] (2) Calculate the statistical characteristics of each time series in the predicted reconstruction error time series set to obtain a statistical characteristic result set;

[0039] (3) Performing a stationarity judgment on each statistical feature result in the statistical feature result set to obtain a stationarity judgment result set;

[0040] (4) When there are non-stationary results in the stationary judgment result set, the anomaly score is calculated for the corresponding prediction reconstruction error time series to obtain an anomaly score set;

[0041] (5) Perform feature matching on the anomaly score set to obtain the sign feature set.

[0042] Specifically, when the abnormality judgment result is abnormal, the data-driven industrial equipment sign prediction method first predicts multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors in the future time period respectively to obtain a predicted reconstruction error time series set. The radial basis function neural network (RBF neural network) is used to achieve the prediction. For the multivariate feature reconstruction error, the following formula is used for prediction:

[0043]

[0044] in, represents the output of the pth node of the output layer, i.e., the predicted multivariate feature reconstruction error, is the connection weight from the hidden layer to the output layer, is the number of nodes in the hidden layer, represents the radial basis function, x is the input vector, is the center of the basis function. The radial basis function uses the Gaussian function:

[0045]

[0046] in, is the Euclidean norm, is the variance of the basis function.

[0047] For the reconstruction error of univariate features, an M-dimensional RBF neural network prediction model is used, where M is the number of univariate features. The prediction formula for each univariate feature is similar to that for multivariate features, but the input and output dimensions are different. After the prediction is completed, the statistical characteristics of each time series in the predicted reconstruction error time series set are calculated to obtain a set of statistical characteristics results. This includes calculating the mean, variance, and covariance. The mean calculation formula is:

[0048]

[0049] in, Reconstruct the error mean for the predicted multivariate features, P is the length of the prediction period;

[0050] The variance calculation formula is:

[0051]

[0052] in, Reconstruct the error variance for the predicted multivariate features.

[0053] The covariance calculation formula is:

[0054]

[0055] in, Reconstruct the error covariance for the predicted multivariate features, For the time interval.

[0056] Next, a stationarity judgment is performed on each statistical feature result in the statistical feature result set to obtain a stationarity judgment result set. The stationarity judgment mainly examines whether the mean, variance and covariance change over time. The judgment criteria include: whether the mean fluctuates within a preset range, whether the variance remains within a certain range, and whether the covariance is only related to the time interval but not to the specific time point. When there are non-stationary results in the stationarity judgment result set, the anomaly score is calculated for the corresponding prediction reconstruction error time series to obtain an anomaly score set. For multivariate features, the anomaly score calculation formula is:

[0057]

[0058] in, is the abnormal score at the pth time point, F is the cumulative distribution function, and are the dynamic mean and dynamic standard deviation respectively. For one-element features, the predicted reconstruction error is directly used as the anomaly score:

[0059]

[0060] in, is the anomaly score of the ith unary feature at the pth time point. Finally, feature matching is performed on the anomaly score set to obtain the symptom feature set. The feature matching process includes setting a threshold to screen suspected symptom features, performing multi-dimensional evaluation on suspected symptom features (including prominence, frequency of occurrence, duration, severity level and confidence), using a radar chart to intuitively display the evaluation results, weighting and weighted calculation of the evaluation indicators to obtain a comprehensive score, and finally determining the credible symptom features through threshold judgment.

[0061] For example, the monitoring system of an industrial pump collects parameters such as bearing temperature, vibration, and current. After the abnormality is determined, the system predicts the reconstruction error for the next 24 hours. Using the RBF neural network prediction model, the reconstruction error data of the past 48 hours is input to predict the reconstruction error time series for the next 24 hours. The statistical characteristics of the prediction results are calculated, and the mean of the bearing temperature reconstruction error is 0.15 and the variance is 0.02; the mean of the vibration reconstruction error is 0.08 and the variance is 0.01. The stability judgment found that the variance of the bearing temperature showed an upward trend and was determined to be unstable. The abnormal score of the predicted reconstruction error time series of the bearing temperature was calculated, and the highest abnormal score was 0.92. After the feature matching process, the bearing temperature abnormality was determined to be the most likely sign feature, which provides an important reference for equipment maintenance.

[0062] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0063] (1) Calculate the Mahalanobis distance for each multivariate feature reconstruction error in the reconstruction error set to obtain multiple multivariate feature reconstruction error values;

[0064] (2) Calculate the absolute value of each unary feature reconstruction error in the reconstruction error set to obtain multiple unary feature reconstruction error values;

[0065] (3) Calculating the cumulative distribution function of the reconstruction error values ​​of multiple multivariate features to obtain multiple multivariate feature anomaly scores;

[0066] (4) Anomaly scores are calculated for the reconstruction error values ​​of multiple unigram features respectively to obtain multiple unigram feature anomaly scores.

[0067] Specifically, the Mahalanobis distance is calculated for each multivariate feature reconstruction error in the reconstruction error set to obtain multiple multivariate feature reconstruction error values. The reconstruction error of the multivariate feature is calculated as follows:

[0068]

[0069] in, is the reconstruction error of the multivariate features at the t-th time step, is the original input, To reconstruct the output, is the covariance matrix and satisfies:

[0070]

[0071] here, Indicates i The original features and j The covariance between the reconstructed features, Mis the total number of features. Next, the absolute value of each unary feature reconstruction error in the reconstruction error set is calculated to obtain multiple unary feature reconstruction error values. The reconstruction error calculation formula of the unary feature is:

[0072]

[0073] in, is the reconstruction error of the ith univariate feature at the tth time step, and |·| represents the absolute value. Then, the cumulative distribution function is calculated for the reconstruction error values ​​of multiple multivariate features to obtain multiple multivariate feature anomaly scores. The calculation formula for the anomaly score of the multivariate feature is:

[0074]

[0075] in, is the anomaly score of the multivariate feature, F is the cumulative distribution function, and Respectively i The dynamic mean and dynamic standard deviation of a univariate feature are calculated as follows:

[0076]

[0077]

[0078] Here, W is the adjustable sliding window length. Finally, the anomaly scores are calculated for the reconstruction error values ​​of multiple unigram features to obtain multiple unigram anomaly scores. The anomaly score of the unigram feature directly uses the reconstruction error value:

[0079]

[0080] in, is the anomaly score of the ith univariate feature. For example, suppose in an industrial pump monitoring system, we are concerned with three features: bearing temperature, vibration, and current. At a certain moment t , the original input data = [80, 0.5, 10] (representing temperature 80°C, vibration 0.5mm / s, and current 10A), reconstructed output = [78, 0.48, 9.8]. First, calculate the multivariate feature reconstruction error, assuming that the covariance matrix It is known that we can substitute it into the formula to get = 2.1. For the unary features, the reconstruction error is calculated separately:

[0081] temperature: = |80 - 78| = 2

[0082] vibration: = |0.5 - 0.48| = 0.02

[0083] Current: = |10 - 9.8| = 0.2

[0084] Next, we calculate the multivariate feature anomaly score. Assume that we use the data from the past 100 time steps to calculate the dynamic mean and standard deviation, and we get = 1.5, = 0.3. Substitute into the formula:

[0085]

[0086] For unigram features, the reconstruction error is directly used as the anomaly score:

[0087] Temperature Anomaly Score: = 2

[0088] Vibration Abnormality Score: = 0.02

[0089] Current Abnormality Score: = 0.2.

[0090] In a specific embodiment, the process of performing the step of respectively predicting a plurality of univariate feature reconstruction errors and a plurality of multivariate feature reconstruction errors in a future time period may specifically include the following steps:

[0091] (1) Construct a prediction model for the radial basis function neural network model to obtain a multivariate feature reconstruction error prediction model;

[0092] (2) constructing a prediction model for the M-dimensional radial basis function neural network model to obtain a univariate feature reconstruction error prediction model, where M is the number of univariate features and M is a positive integer greater than 1;

[0093] (3) Arranging the time series of the multivariate feature reconstruction errors in the reconstruction error set to obtain the multivariate feature reconstruction error time series;

[0094] (4) Training the multivariate feature reconstruction error prediction model through the multivariate feature reconstruction error time series to obtain a trained multivariate feature prediction model;

[0095] (5) sorting the time series of multiple one-element feature reconstruction errors in the reconstruction error set respectively to obtain M one-element feature reconstruction error time series;

[0096] (6) Training the univariate feature reconstruction error prediction model through M univariate feature reconstruction error time series to obtain a trained univariate feature prediction model;

[0097] (7) Predicting the multivariate feature reconstruction error in the future time period through the trained multivariate feature prediction model to obtain a predicted multivariate feature reconstruction error time series;

[0098] (8) Predicting the M univariate feature reconstruction errors in the future time period respectively through the trained univariate feature prediction model to obtain M predicted univariate feature reconstruction error time series;

[0099] (9) The predicted multi-element feature reconstruction error time series and the M predicted uni-element feature reconstruction error time series are combined to obtain a set of predicted reconstruction error time series.

[0100] Specifically, a prediction model is constructed for the radial basis function neural network model to obtain a multivariate feature reconstruction error prediction model. The output layer formula of the multivariate feature reconstruction error prediction model is as follows:

[0101]

[0102] in, The output layer p The output of the node, is the connection weight from the hidden layer to the output layer, is the number of nodes in the hidden layer, represents the radial basis function, is the input vector, is the center of the basis function. The radial basis function uses a Gaussian function:

[0103]

[0104] in, is the Euclidean norm, is the variance of the basis function. At the same time, the prediction model of the M-dimensional radial basis function neural network model is constructed to obtain a univariate feature reconstruction error prediction model, where M is the number of univariate features and M is a positive integer greater than 1. The structure of the univariate feature reconstruction error prediction model is similar to that of the multivariate feature model, but each univariate feature has its own independent prediction model.

[0105] Next, the multivariate feature reconstruction errors in the reconstruction error set are sorted in time series to obtain a multivariate feature reconstruction error time series. The discrete reconstruction error data are arranged in chronological order to form continuous time series data. Then, the multivariate feature reconstruction error prediction model is trained through the multivariate feature reconstruction error time series to obtain a trained multivariate feature prediction model. The training process mainly includes determining the basis function center and adjusting the connection weights. Similarly, the multiple univariate feature reconstruction errors in the reconstruction error set are sorted in time series to obtain M univariate feature reconstruction error time series. The univariate feature reconstruction error prediction model is trained through these M univariate feature reconstruction error time series to obtain a trained univariate feature prediction model.

[0106] After the training is completed, the multivariate feature prediction model is used to predict the multivariate feature reconstruction error in the future time period to obtain the predicted multivariate feature reconstruction error time series. At the same time, the trained univariate feature prediction model is used to predict the M univariate feature reconstruction errors in the future time period to obtain M predicted univariate feature reconstruction error time series. Finally, the predicted multivariate feature reconstruction error time series and the M predicted univariate feature reconstruction error time series are merged to obtain a set of predicted reconstruction error time series.

[0107] In a specific embodiment, the process of performing the feature matching step on the anomaly score set may specifically include the following steps:

[0108] (1) Design a threshold comparison for each unary feature anomaly score in the anomaly score set to obtain multiple suspected sign features;

[0109] (2) Conduct multi-dimensional evaluations on multiple suspected signs and characteristics, and obtain multiple evaluation results including prominence, frequency of occurrence, duration, severity, and confidence level;

[0110] (3) Draw radar charts for multiple evaluation results to obtain multiple intuitive display data for each evaluation indicator;

[0111] (4) weighting the evaluation indicators in the multiple intuitive display data of each evaluation indicator to obtain multiple weighted evaluation results;

[0112] (5) Comprehensively calculate multiple weighted evaluation results to obtain multiple comprehensive scores;

[0113] (6) Performing threshold judgments on multiple comprehensive scores to obtain multiple credible sign features;

[0114] (7) Integrate multiple credible sign features to obtain a sign feature set.

[0115] Specifically, a threshold is designed for each unary feature anomaly score in the anomaly score set to compare and obtain multiple suspected symptom features. This process sets a threshold and marks features with anomaly scores higher than the threshold as suspected symptom features. Next, multiple suspected symptom features are evaluated in multiple dimensions to obtain multiple evaluation results including prominence, frequency of occurrence, duration, severity and confidence. Prominence reflects the significance of the feature anomaly, frequency of occurrence indicates the number of times the feature anomaly occurs, duration describes the duration of the abnormal state, severity measures the potential impact of the anomaly on the device, and confidence reflects the degree of certainty that the feature is a symptom.

[0116] Then, radar charts are drawn for multiple evaluation results to obtain multiple intuitive display data for each evaluation indicator. Radar chart is a multidimensional data visualization tool. Each axis represents an evaluation indicator. The larger the value, the farther from the center. This intuitive display method helps to quickly identify the performance of each feature in different dimensions. Then, the evaluation indicators in the multiple intuitive display data of each evaluation indicator are weighted to obtain multiple weighted evaluation results. The weighting process takes into account the importance of different evaluation indicators, and usually uses expert experience or data-driven methods to determine the weights. Assume that the weight vector is W = (w1, w2, w3, w4, w5), corresponding to the weights of visibility, frequency of occurrence, duration, severity and confidence.

[0117] The multiple weighted evaluation results are calculated comprehensively to obtain multiple comprehensive scores. The comprehensive score can be calculated by weighted summation. Finally, the multiple comprehensive scores are judged by thresholds to obtain multiple credible sign features, and these credible sign features are integrated to obtain a sign feature set. The threshold judgment can use a fixed threshold or a dynamic threshold method to determine the features with a comprehensive score higher than the threshold as credible sign features.

[0118] For example, suppose that in an industrial pump monitoring system, the three features of interest are bearing temperature, vibration, and current. After the previous steps, the anomaly scores of these three features are 0.85, 0.62, and 0.78 respectively. If the threshold is set to 0.7, the bearing temperature and current are marked as suspected symptom features.

[0119] A multi-dimensional evaluation of these two suspected signs yielded the following results:

[0120] Bearing temperature: Obviousness 0.9, Occurrence frequency 0.8, Duration 0.7, Severity 0.9, Confidence 0.8

[0121] Electric current: Obviousness 0.8, Occurrence frequency 0.7, Duration 0.6, Severity 0.7, Confidence 0.7

[0122] After drawing the radar chart, the weight vector W = (0.25, 0.2, 0.15, 0.25,0.15) is determined through expert experience. Perform weighted calculation:

[0123] Bearing temperature comprehensive score: 0.9×0.25 + 0.8×0.2 + 0.7×0.15 + 0.9×0.25 + 0.8×0.15 = 0.84

[0124] Current comprehensive score: 0.8×0.25 + 0.7×0.2 + 0.6×0.15 + 0.7×0.25 + 0.7×0.15 = 0.72

[0125] When the comprehensive score threshold is set to 0.75, the bearing temperature is determined to be a reliable symptom feature and is finally included in the symptom feature set. The results show that abnormal bearing temperature is an important sign of potential problems in industrial pumps.

[0126] In a specific embodiment, the process of performing the step of weighting the evaluation indicators in the plurality of intuitive display data of each evaluation indicator may specifically include the following steps:

[0127] (1) Analyze the importance of the evaluation indicators in the multiple intuitive display data of each evaluation indicator to obtain the importance ranking of each evaluation indicator;

[0128] (2) According to the importance ranking of each evaluation indicator, a weight coefficient is assigned to each evaluation indicator to obtain a set of evaluation indicator weights;

[0129] (3) The multiple intuitive display data of each evaluation indicator are weighted and calculated through the evaluation indicator weight set to obtain multiple weighted evaluation results.

[0130] Specifically, the importance of the evaluation indicators in the multiple intuitive display data of each evaluation indicator is analyzed to obtain the importance ranking of each evaluation indicator, which is usually combined with expert experience and data statistical methods, such as principal component analysis or correlation analysis.

[0131] Evaluation indicators include prominence, frequency, duration, severity and confidence. Importance analysis can be performed by calculating the correlation between each indicator and the final abnormality judgment result. Assuming there are n samples, for the i-th evaluation indicator, its importance can be calculated using the following formula:

[0132]

[0133] in, It is i The importance score of each indicator, It isj The sample i The index value, It is i The average value of the indicators, It is j The final abnormal judgment result of samples is is the average value of the abnormal judgment results. After calculating the importance scores of all indicators, the indicators are sorted according to the scores to obtain the importance ranking. Next, according to the importance ranking of each evaluation indicator, a weight coefficient is assigned to each evaluation indicator to obtain the evaluation indicator weight set. The weight assignment can be normalized to ensure that the sum of all weights is 1. Finally, the multiple intuitive display data of each evaluation indicator are weighted through the evaluation indicator weight set to obtain multiple weighted evaluation results.

[0134] For example, suppose that in an industrial pump monitoring system, we focus on three features: bearing temperature, vibration, and current. Each feature has five evaluation indicators: visibility, frequency of occurrence, duration, severity, and confidence. Through the previous steps, we obtained the evaluation indicator data of 100 samples and the corresponding abnormal judgment results.

[0135] First, calculate the importance score of each indicator. Assume that the calculation results are as follows:

[0136] Obviousness: 0.85

[0137] Occurrence frequency: 0.72

[0138] Duration: 0.63

[0139] Severity level: 0.91

[0140] Confidence: 0.78

[0141] Based on these scores, the importance of indicators is ranked as follows: severity > visibility > confidence > frequency of occurrence > duration.

[0142] Next, calculate the weight of each indicator:

[0143] Severity level: 0.91 / (0.85 + 0.72 + 0.63 + 0.91 + 0.78) ≈ 0.234

[0144] Obviousness: 0.85 / 3.89 ≈ 0.219

[0145] Confidence: 0.78 / 3.89 ≈ 0.201

[0146] Occurrence frequency: 0.72 / 3.89 ≈ 0.185

[0147] Duration: 0.63 / 3.89 ≈ 0.162

[0148] Finally, for each sample, these weights are used for weighted calculation. For example, for a particular sample, its evaluation index value is:

[0149] Obviousness: 0.9, Occurrence: 0.8, Duration: 0.7, Severity: 0.95, Confidence: 0.85

[0150] The weighted evaluation results are calculated as follows:

[0151] S = 0.234 × 0.95 + 0.219 × 0.9 + 0.201 × 0.85 + 0.185 × 0.8 +0.162 × 0.7 ≈ 0.8573.

[0152] In a specific embodiment, the data-driven industrial equipment sign prediction method may further specifically include the following steps:

[0153] (1) Record and store the symptom feature set to obtain abnormal case data;

[0154] (2) Build a shared knowledge base of abnormal cases through abnormal case data to obtain a queryable abnormal case library.

[0155] Specifically, the symptom feature set is recorded and stored to obtain abnormal case data. This process involves systematically organizing and saving the symptom features and related information identified in the previous steps. The recorded content includes but is not limited to a detailed description of the abnormal event, abnormal threshold, abnormal frequency, start and end time, abnormal severity level, abnormal explanation, credible symptom features and false symptom features. These data are stored in a structured form, usually using a relational database or a non-relational database, to facilitate subsequent retrieval and analysis. Next, an abnormal case shared knowledge base is constructed through the abnormal case data to obtain a queryable abnormal case library. It is not just a simple data storage, but to convert the stored abnormal case data into a knowledge base that can be queried, analyzed and learned. The construction process includes data cleaning, standardization, index establishment and association analysis. Data cleaning ensures the quality of data stored in the knowledge base and removes redundant, erroneous or incomplete records. The standardization process unifies data from different sources or formats into a standard format to facilitate subsequent comparison and analysis. Index establishment is to improve query efficiency, and indexes are usually established for commonly used query fields such as device type, abnormal type, time, etc. Correlation analysis is about establishing connections between different abnormal cases, which helps to discover potential association patterns and causal relationships.

[0156] During the construction process, special attention is paid to the classification and labeling of abnormal cases. Cases are classified according to dimensions such as device type, abnormality type, and severity, and corresponding labels are added to each case. This classification and labeling not only facilitates subsequent retrieval, but also helps with statistical analysis and pattern recognition. In addition, the knowledge base needs to support multi-dimensional query and analysis functions, such as compound queries by time, device type, abnormality type, etc., or trend analysis and correlation analysis. In order to enable the knowledge base to have the ability to self-update and learn, a feedback mechanism needs to be designed. When new abnormal cases are successfully predicted and processed, relevant information will be added to the knowledge base in a timely manner, continuously enriching and updating the content of the knowledge base. At the same time, cases of prediction failure or false alarm will also be recorded for subsequent model optimization and improvement.

[0157] For example, in the equipment monitoring system of a large chemical plant, a centrifugal pump was detected to be abnormal during operation. Through the previous steps, it was identified that the bearing temperature abnormality was the main symptom feature. This case was recorded, including the following information: equipment ID (such as Pump-001), abnormality type (bearing temperature abnormality), abnormality occurrence time (2023-09-15 14:30:00), duration (2 hours), maximum temperature (85°C, the normal range should be below 60°C), vibration data (showing a slight increase but not exceeding the standard), and related operation records (last maintenance time, load changes, etc.). This information is stored in the abnormal case database. When building the knowledge base, this case was classified into categories such as "rotating equipment", "bearing failure", and "temperature abnormality". At the same time, tags such as "centrifugal pump", "bearing overheating", "inadequate lubrication", etc. are added to the case. Through this classification and tagging, when encountering similar situations in the future, engineers can quickly find relevant cases. For example, they can search for "all cases of bearing temperature abnormality that occurred in the last 3 months", and the system will quickly return all relevant records including this case. As time goes by, the knowledge base continues to accumulate similar cases. Through the analysis of these cases, it is found that after 48 hours of continuous high-load operation, there is an 80% probability that the bearing temperature of this model of centrifugal pump will be abnormal. This discovery was added to the knowledge base and became an important basis for predictive maintenance. When the system detects that a pump is about to reach this operating time, it will automatically issue an early warning to remind engineers to pay attention to monitoring the bearing temperature, thus achieving a transition from passive response to active prevention.

[0158] The above describes the data-driven industrial equipment sign prediction method in the embodiment of the present application. The following describes the data-driven industrial equipment sign prediction device in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a data-driven industrial equipment sign prediction device includes:

[0159] The acquisition module 201 is used to sample the operating status of the industrial equipment in real time to obtain multivariate time series data including multiple univariate features and multiple multivariate features;

[0160] A decoding module 202, used to encode and decode the multivariate time series data through a preset time series reconstruction model to obtain a reconstructed output;

[0161] A calculation module 203 is used to perform difference calculation on the multivariate time series data and the reconstructed output to obtain a reconstruction error set including a plurality of univariate feature reconstruction errors and a plurality of multivariate feature reconstruction errors;

[0162] A scoring module 204 is used to score the reconstruction error set to obtain an anomaly score, wherein the anomaly score includes: a plurality of multivariate feature anomaly scores and a plurality of univariate feature anomaly scores;

[0163] The judgment module 205 is used to perform a threshold estimation on the abnormality score to obtain an abnormality threshold, and compare and judge the abnormality score and the abnormality threshold to obtain an abnormality judgment result of the current state.

[0164] Through the cooperation of the above components, the multivariate time series data containing multiple univariate features and multiple multivariate features are obtained by real-time sampling of the operating status of industrial equipment. The preset time series reconstruction model is used to encode and decode these data to obtain the reconstructed output, which effectively captures the potential patterns and structures in the data. By calculating the difference between the multivariate time series data and the reconstructed output, a reconstruction error set containing multiple univariate feature reconstruction errors and multiple multivariate feature reconstruction errors is obtained, which can accurately quantify the deviation between the original data and the reconstructed data, providing a key indicator for anomaly detection. The reconstruction error set is scored to obtain an anomaly score including multiple multivariate feature anomaly scores and multiple univariate feature anomaly scores. This multi-dimensional scoring method can comprehensively reflect the degree of abnormality of the equipment status. Subsequently, the anomaly score is threshold estimated to obtain the anomaly threshold, and the anomaly judgment result of the current state is obtained by comparing the anomaly score and the anomaly threshold. This process realizes the accurate judgment of the equipment status. The advantage of this method is that it can simultaneously consider the anomaly of a single parameter (univariate feature) and the complex relationship between multiple parameters (multivariate feature), which greatly improves the accuracy and comprehensiveness of anomaly detection. Through real-time data processing and analysis, it is possible to promptly detect minor anomalies in equipment operation, making preventive maintenance possible. It can automatically adjust the abnormal threshold according to different equipment and working conditions, reducing human intervention and improving the intelligence level of the system. In general, this method achieves comprehensive monitoring and accurate prediction of the status of industrial equipment through data-driven methods, effectively improving the reliability and operating efficiency of equipment, while reducing maintenance costs.

[0165] The present application also provides a data-driven industrial equipment sign prediction device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the data-driven industrial equipment sign prediction method in the above-mentioned embodiments.

[0166] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein instructions are stored in the computer, and when the instructions are executed on a computer, the computer executes the steps of the data-driven industrial equipment sign prediction method.

[0167] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0168] If the integrated unit is implemented in the form of a software functional unit and sold or passed as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0169] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data-driven industrial equipment symptom prediction method, characterized in that: The data-driven industrial equipment symptom prediction method includes: The operating status of industrial equipment is sampled in real time to obtain multivariate time series data containing multiple univariate features and multiple multivariate features; Encoding and decoding the multivariate time series data through a preset time series reconstruction model to obtain a reconstructed output; Performing difference calculation on the multivariate time series data and the reconstruction output to obtain a reconstruction error set including a plurality of univariate feature reconstruction errors and a plurality of multivariate feature reconstruction errors; Performing scoring processing on the reconstruction error set to obtain an anomaly score, wherein the anomaly score includes: a plurality of multivariate feature anomaly scores and a plurality of univariate feature anomaly scores; Performing a threshold estimation on the abnormality score to obtain an abnormality threshold, and performing a comparison and judgment on the abnormality score and the abnormality threshold to obtain an abnormality judgment result of the current state; When the abnormality judgment result is abnormal, multiple unary feature reconstruction errors and multiple multivariate feature reconstruction errors in the future time period are predicted respectively to obtain a predicted reconstruction error time series set, which specifically includes: constructing a prediction model for the radial basis function neural network model to obtain a multivariate feature reconstruction error prediction model; constructing a prediction model for the M-dimensional radial basis function neural network model to obtain a unary feature reconstruction error prediction model, wherein M is the number of unary features and M is a positive integer greater than 1; sorting the multivariate feature reconstruction errors in the reconstruction error set in time series to obtain a multivariate feature reconstruction error time series; training the multivariate feature reconstruction error prediction model through the multivariate feature reconstruction error time series to obtain a trained multivariate feature prediction model; sorting the multiple unary feature reconstruction errors in the reconstruction error set in time series to obtain M unary feature reconstruction error time series; training the unary feature reconstruction error prediction model through the M unary feature reconstruction error time series to obtain a trained unary feature prediction model; predicting the multivariate feature reconstruction errors in the future time period through the trained multivariate feature prediction model to obtain a predicted multivariate feature reconstruction error time series; The M univariate feature reconstruction errors in the future time period are predicted respectively by the trained univariate feature prediction model to obtain M predicted univariate feature reconstruction error time series; The predicted multivariate feature reconstruction error time series and the M predicted univariate feature reconstruction error time series are combined to obtain a predicted reconstruction error time series set; Calculating statistical features for each time series in the predicted reconstruction error time series set to obtain a statistical feature result set; Performing a stationarity judgment on each statistical feature result in the statistical feature result set to obtain a stationarity judgment result set; When there is an unsteady result in the stationary judgment result set, an abnormal score is calculated for the corresponding prediction reconstruction error time series to obtain an abnormal score set; Feature matching is performed on the anomaly score set to obtain a symptom feature set.

2. The method according to claim 1, characterized in that The scoring process of the reconstruction error set to obtain an abnormality score includes: Performing Mahalanobis distance calculation on each multivariate feature reconstruction error in the reconstruction error set to obtain a plurality of multivariate feature reconstruction error values; Performing absolute value calculation on each unary feature reconstruction error in the reconstruction error set to obtain a plurality of unary feature reconstruction error values; Performing cumulative distribution function calculations on the multiple multivariate feature reconstruction error values ​​respectively to obtain multiple multivariate feature anomaly scores; Anomaly scores are calculated for the multiple unary feature reconstruction error values ​​respectively to obtain multiple unary feature anomaly scores.

3. The method according to claim 1, characterized in that The performing feature matching on the anomaly score set to obtain a sign feature set includes: Performing a designed threshold comparison on each unary feature anomaly score in the anomaly score set to obtain a plurality of suspected sign features; Performing multi-dimensional evaluations on the multiple suspected sign characteristics to obtain multiple evaluation results including prominence, frequency of occurrence, duration, severity level, and confidence level; Draw radar charts for the multiple evaluation results respectively to obtain multiple intuitive display data of each evaluation index; Performing weighting processing on the evaluation indicators in the multiple intuitive display data of each evaluation indicator to obtain multiple weighted evaluation results; Performing comprehensive calculations on the multiple weighted evaluation results respectively to obtain multiple comprehensive scores; Performing threshold judgment on the multiple comprehensive scores respectively to obtain multiple credible sign features; The multiple credible symptom features are integrated to obtain a symptom feature set.

4. The method according to claim 3, characterized in that The weighting process is performed on the evaluation indicators in the multiple intuitive display data of each evaluation indicator to obtain multiple weighted evaluation result sets, including: Performing importance analysis on the evaluation indicators in the plurality of intuitive display data of the evaluation indicators to obtain the importance ranking of the evaluation indicators; According to the importance ranking of the evaluation indicators, a weight coefficient is assigned to each evaluation indicator to obtain an evaluation indicator weight set; The plurality of intuitive display data of each evaluation index are weightedly calculated by the evaluation index weight set to obtain a plurality of weighted evaluation results.

5. The method according to claim 4, characterized in that Also includes: Recording and storing the symptom feature set to obtain abnormal case data; An abnormal case sharing knowledge base is constructed through the abnormal case data to obtain a queryable abnormal case base.

6. A data-driven industrial equipment sign prediction device, characterized in that: The data-driven industrial equipment symptom prediction device comprises: An acquisition module is used to sample the operating status of industrial equipment in real time to obtain multivariate time series data containing multiple univariate features and multiple multivariate features; A decoding module, used for encoding and decoding the multivariate time series data through a preset time series reconstruction model to obtain a reconstructed output; A calculation module, used for performing difference calculation on the multivariate time series data and the reconstructed output to obtain a reconstruction error set including a plurality of univariate feature reconstruction errors and a plurality of multivariate feature reconstruction errors; A scoring module, used for scoring the reconstruction error set to obtain an anomaly score, wherein the anomaly score includes: a plurality of multivariate feature anomaly scores and a plurality of univariate feature anomaly scores; A judgment module is used to perform a threshold estimation on the abnormal score to obtain an abnormal threshold, compare and judge the abnormal score and the abnormal threshold to obtain an abnormal judgment result of the current state; A prediction module is used to predict multiple unary feature reconstruction errors and multiple multivariate feature reconstruction errors in a future time period respectively when the abnormality judgment result is abnormal, and obtain a predicted reconstruction error time series set, specifically including: constructing a prediction model for a radial basis function neural network model to obtain a multivariate feature reconstruction error prediction model; constructing a prediction model for an M-dimensional radial basis function neural network model to obtain a unary feature reconstruction error prediction model, wherein M is the number of unary features and M is a positive integer greater than 1; arranging the multivariate feature reconstruction errors in the reconstruction error set in time series to obtain a multivariate feature reconstruction error time series; training the multivariate feature reconstruction error prediction model through the multivariate feature reconstruction error time series to obtain a trained multivariate feature prediction model; arranging the multiple unary feature reconstruction errors in the reconstruction error set in time series respectively to obtain M unary feature reconstruction error time series; training the unary feature reconstruction error prediction model through the M unary feature reconstruction error time series to obtain a trained unary feature prediction model; Predicting the multivariate feature reconstruction error in a future time period by the trained multivariate feature prediction model to obtain a predicted multivariate feature reconstruction error time series; Predicting M univariate feature reconstruction errors in a future time period by the trained univariate feature prediction model to obtain M predicted univariate feature reconstruction error time series; Combining the predicted multivariate feature reconstruction error time series with the M predicted univariate feature reconstruction error time series to obtain a predicted reconstruction error time series set; A statistical module, used for calculating the statistical characteristics of each time series in the predicted reconstruction error time series set to obtain a statistical characteristic result set; An analysis module, used for performing a stationarity judgment on each statistical feature result in the statistical feature result set to obtain a stationarity judgment result set; A screening module, used for calculating anomaly scores on corresponding prediction reconstruction error time series when there are non-stationary results in the stationarity judgment result set, to obtain anomaly score set; The matching module is used to perform feature matching on the anomaly score set to obtain a symptom feature set.

7. A data-driven industrial equipment symptom prediction device, characterized in that: The data-driven industrial equipment symptom prediction device comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory to enable the data-driven industrial equipment symptom prediction device to execute the data-driven industrial equipment symptom prediction method according to any one of claims 1 to 5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by a processor, the data-driven industrial equipment symptom prediction method according to any one of claims 1 to 5 is implemented.

Images