A residual double-space industrial process fault detection method and system based on IForest-LOF

Abstract

This invention relates to the field of industrial process fault detection technology, specifically disclosing a residual dual-space industrial process fault detection method and system based on IFOres-LOF. The method collects multivariate operational data of industrial processes and performs standardized preprocessing. Principal component analysis is used to construct principal component spaces and residual spaces. In the principal component space, an isolated forest model is used to identify global sparse anomalies, while in the residual space, a local outlier model is used to capture local fine-grained density deviations. The scores of the two types of anomalies are successively normalized and augmented with posterior probabilities. Adaptive weighted fusion is then performed based on the statistical uncertainty of the scores to obtain a comprehensive anomaly score. The fault judgment threshold is determined based on the comprehensive score distribution of normally operating samples, thus completing the fault identification of the test samples. This invention takes into account both global faults and local weak faults, does not rely on fault sample training, and is suitable for online fault monitoring of complex industrial processes.

Description

Technical Field

[0001] This invention relates to the field of industrial process fault detection technology, and in particular to a residual dual-space industrial process fault detection method and system based on IFOres-LOF. Background Technology

[0002] As industrial processes evolve towards automation, digitalization, and intelligence, production facilities continuously generate vast amounts of multivariate process data. Timely identification of potential faults from this data is crucial for ensuring the safe and stable operation of facilities and minimizing downtime losses. Data-driven process monitoring methods, because they do not rely on complex mechanistic models, have become an important direction for industrial fault detection.

[0003] Principal Component Analysis (PCA) is a commonly used multivariate statistical method in industrial process monitoring. Existing PCA methods typically use T0... 2 Statistics describe anomalies in the principal component space, while the squared prediction error (SPE) statistic describes anomalies in the residual space, or a simple combination of the two types of statistics can be used. This approach has the following limitations: First, T... 2 First, since SPEs are mostly scalar statistics, they easily compress the directional and structural information in the residual vectors. Second, the data distribution and anomaly patterns of the principal component space and the residual space are different, and it is difficult to fully utilize the complementary information of the two spaces by using the same type of detection logic or simple threshold combination. Third, although supervised learning methods such as support vector machines and neural networks can directly analyze residual vectors, they require a large number of fault samples, while fault samples are scarce and class imbalanced in real industrial scenarios.

[0004] Isolation Forest (IForest) is an unsupervised anomaly detection algorithm suitable for identifying global anomalies in high-dimensional data that are sparsely distributed and deviate from the overall normal pattern. Local Outlier Factor (LOF) is an unsupervised anomaly detection algorithm based on local density differences, suitable for capturing hidden anomalies with significant decreases in local neighborhood density. Current technologies lack an industrial process fault detection scheme that separately models the principal component space and residual space of PCA and adaptively fuses them according to the statistical characteristics of anomaly scores in each space. Summary of the Invention

[0005] This invention aims to solve the problems in existing industrial process fault detection, such as the imbalance between global features and local biases, insufficient utilization of residual information, poor adaptability of single detection models to different anomaly patterns, and the dependence of supervisory models on fault samples.

[0006] To achieve the above objectives, this invention provides a residual dual-space industrial process fault detection method based on IFOres-LOF, comprising the following steps: S1. Collect multivariate operational data of industrial processes and perform standardized preprocessing to form a sample matrix; S2. Establish a principal component analysis model based on normal operating condition training data and construct a projection matrix according to the principal components; project the sample matrix onto the principal component space to obtain the principal component space features; reconstruct the sample matrix using the principal component space features and the projection matrix to obtain the reconstructed data, calculate the difference between the sample matrix and the reconstructed data, and obtain the residual space features; S3. Train the isolated forest detection model on the principal component space features and output the first anomaly score; train the local outlier detection model on the residual space features and output the second anomaly score. S4. After normalizing the first abnormal score and the second abnormal score and performing posterior probabilistic augmentation processing, adaptive weighted fusion is performed to obtain a comprehensive abnormal score. S5. Using the operating data during the normal operation phase, perform steps S1-S4 to obtain the comprehensive anomaly score for normal operation. Determine the fault judgment threshold based on the distribution of the comprehensive anomaly score for normal operation. Perform steps S1-S4 on the sample to be tested to obtain the comprehensive anomaly score for the sample to be tested. Compare the comprehensive anomaly score of the sample to be tested with the fault judgment threshold and output the industrial process fault detection result.

[0007] Furthermore, in step S2, the method for obtaining the principal components is as follows: Calculate the covariance matrix of the training data under normal operating conditions and perform eigenvalue decomposition. Select the top k principal components according to the cumulative variance contribution rate; the cumulative variance contribution rate is 95%, 97%, or 99%.

[0008] Furthermore, in step S3, the isolated forest detection model constructs multiple random isolated trees and calculates the first anomaly score based on the average path length of the sample in each isolated tree. This score is used to identify global sparse anomalies in the principal component space that deviate from the normal global change pattern.

[0009] Furthermore, in step S3, the local outlier detection model obtains a second outlier score by calculating the reachability distance, local reachability density, and relative density deviation between the sample and its k nearest neighbors, which is used to identify local fine-grained density deviations in the residual space.

[0010] Furthermore, in step S4, the normalization process is Min-Max normalization; the posterior probabilistic enhancement process is implemented based on the Sigmoid function.

[0011] Furthermore, in step S4, the weights of the adaptive weighted fusion are determined based on the statistical uncertainty of the abnormal scores after the two feature spaces are enhanced. The feature space with lower statistical uncertainty corresponds to a higher fusion weight. The statistical uncertainty is characterized by the variance or standard deviation of the abnormal scores.

[0012] Furthermore, in step S5, the fault judgment threshold is determined based on the distribution of the comprehensive anomaly score during normal operation, and the values ​​corresponding to the top 5 percentiles of the comprehensive anomaly score during normal operation, arranged from high to low, are selected as the fault judgment threshold. When the overall anomaly score of the sample under test is greater than the fault judgment threshold, it is judged as a fault; when the overall anomaly score of the sample under test is less than or equal to the fault judgment threshold, it is judged as normal.

[0013] Furthermore, the isolated forest detection model is replaced with a global anomaly detection model that includes a class of support vector machines, elliptical envelopes, and autoencoders; the local outlier detection model is replaced with an unsupervised detection model that includes k-nearest neighbor distance, density peaks, and local kernel density estimation; and the fault judgment threshold adopts quantile threshold, mean plus standard deviation, extreme value theory, or sliding window adaptive threshold.

[0014] This invention also provides a residual dual-space industrial process fault detection system based on IFOrest-LOF, which adopts the above method and includes a data acquisition module, a data preprocessing module, a PCA dual-space construction module, a principal component space isolated forest detection module, a residual space local outlier factor detection module, an anomaly score fusion module, and a fault discrimination output module. The data acquisition module is used to collect multivariate operational data of industrial processes; The data preprocessing module is used to perform preprocessing on the collected multivariate runtime data; The PCA dual-space construction module is used to generate principal component space features through principal component analysis and to obtain residual space features through data reconstruction calculation. The principal component space isolated forest detection module has a built-in isolated forest detection model, which is used to calculate and output the first anomaly score based on the principal component space features. The residual space local outlier detection module has a built-in local outlier detection model, which is used to calculate and output the second outlier score based on the residual space features. The anomaly score fusion module is used to perform normalization, enhancement and adaptive weighted fusion on the first anomaly score and the second anomaly score in sequence to obtain a comprehensive anomaly score; The fault detection output module is used to compare the comprehensive anomaly score with the fault judgment threshold and output the fault detection result.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for residual dual-space industrial process fault detection based on IFOres-LOF.

[0016] The present invention employs the above-mentioned residual dual-space industrial process fault detection method and system based on IFOres-LOF, and its beneficial effects are as follows: (1) The present invention makes full use of the complementary information of principal component space and residual space; principal component space is used to characterize the global change pattern of industrial process, and residual space is used to capture local deviations and hidden anomalies that are not retained by principal component, thereby reducing the information loss caused by single space detection. (2) This invention employs different detection models for different spaces to improve the adaptability of anomaly patterns. IForest is suitable for identifying global sparse anomalies, while LOF is suitable for identifying local density anomalies. The combination of the two can take into account both global faults and local weak faults. (3) This invention does not require a large number of fault samples. The IFOres and LOF used are both unsupervised anomaly detection models, which can be mainly trained based on normal operating condition data, making them suitable for the problem of scarce fault samples in actual industrial scenarios; (4) This invention reduces the impact of differences in the score scales of different models on the detection results by score normalization, posterior enhancement and adaptive weight fusion, and improves the separability between fault samples and normal samples. Attached Figure Description

[0017] Figure 1 This is a flowchart of a residual dual-space industrial process fault detection method and system based on IFOrest-LOF according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the TF process of a residual dual-space industrial process fault detection method based on IFOrest-LOF according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the dual-space model structure of a residual dual-space industrial process fault detection method based on IFOrest-LOF according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the fault detection results of a residual dual-space industrial process fault detection method based on IFOrest-LOF according to an embodiment of the present invention; wherein, (a) is the detection result of Fault 05, (b) is the detection result of Fault 10, and (c) is the detection result of Fault 19. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0019] like Figure 1 As shown in the figure, this invention provides a residual dual-space industrial process fault detection method based on IFOres-LOF.

[0020] This embodiment uses Tennessee-Eastman (TE) process baseline data for illustration. TE, for example... Figure 2 As shown; this process involves multiple continuous chemical process variables, and this embodiment selects 52 variables for process monitoring. Normal operating condition data d00 is used as training data, containing 500 samples; d00 te As normal validation data, it contains 960 samples; using d01 te up to d21 te The test data consists of 21 types of faults, each containing 960 samples, with the 161st sample serving as the fault initiation point.

[0021] Includes the following steps: S1. Standardize the 52-dimensional process variables to obtain the training sample matrix X; Standardization involves handling missing values, removing obviously erroneous values, and standardizing or normalizing to make different variables comparable.

[0022] S2. Calculate the covariance matrix of the training data under normal operating conditions and perform eigenvalue decomposition. Select k principal components according to the rule that the cumulative variance contribution rate is not less than 99% to form the projection matrix W; so as to suppress the influence of noise while retaining the main process information.

[0023] Projecting X onto the principal component space (low-dimensional space) yields the principal component space feature matrix Z; using Z and W to reconstruct X, we obtain the reconstructed data matrix X. hat The difference between the sample matrix and the reconstructed data is calculated to obtain the residual space feature matrix R=XX. hat R represents residual information not explained by the principal component space.

[0024] S3, RIFL dual-space model structure as follows Figure 3As shown, an isolation forest detection model is trained on the principal component space feature matrix Z. The isolation forest detection model constructs multiple random isolation trees by randomly selecting features and split points, calculates the average path length of each sample in each isolation tree, and calculates the principal component space anomaly score based on the average path length. This score is used as the first anomaly score to identify global sparse anomalies in the principal component space that deviate from the global change pattern of normal operation. Samples that deviate from the overall change pattern of normal operating conditions are more likely to be isolated, have shorter path lengths, and higher anomaly scores.

[0025] A local outlier detection model is trained on the residual space feature matrix R. For each sample, the reachability distance, local reachability density, and relative density deviation among its k nearest neighbors are calculated to obtain the residual space outlier score, which is used as the second outlier score to identify local fine-grained density deviations in the residual space. When the local density of a sample in the residual space is significantly lower than that of its neighboring samples, the LOF score increases, indicating that it may correspond to a local hidden anomaly.

[0026] S4. Perform Min-Max normalization on the first and second abnormal scores respectively, based on the mean and standard deviation of the normalized scores of the training or validation samples; use the Sigmoid function to perform posterior enhancement on the abnormal scores, with a sensitivity parameter λ of 0.1, to enhance the response of samples close to the abnormal boundary; amplify abnormal scores that are close to or above the normal level, and suppress low abnormal scores.

[0027] Calculate the variance or standard deviation of the anomaly scores after enhancement of the two detection spaces, and determine the principal component space weights and residual space weights according to the rule that lower uncertainty has higher weights, so that the more stable and reliable detection space has a higher proportion in the fusion. The anomaly scores of the two spaces are then weighted and summed to obtain the comprehensive anomaly score S.

[0028] S5. Finally, the threshold θ is determined based on the comprehensive anomaly score during normal operation. In this embodiment, the threshold can be the value corresponding to the highest 5% of the comprehensive anomaly scores in the samples during the normal operation phase before the fault is introduced. When the sample S > θ, it is determined to be a fault; when S ≤ θ, it is determined to be normal. The detection result can be output to the industrial monitoring system in real time for alarm, fault recording, or subsequent diagnostic analysis.

[0029] The test results of Fault 05, Fault 10, and Fault 19 were used as examples to illustrate the experimental results on 21 types of fault data in the TE process. The curves showing the change of the comprehensive anomaly score before and after fault injection with the sample are as follows: Figure 4 As shown.

[0030] On 21 types of fault data in the TE process, the average false alarm rate of the method in this embodiment is 2.62%, and the average false alarm rate is 18.67%. Specifically, the average false alarm rate of LOF is approximately 3.57%, and the average false alarm rate is approximately 26.52%; the average false alarm rate of PCA is approximately 1.37%, and the average false alarm rate is approximately 37.12%; and the average false alarm rate of IFOres is approximately 4.26%, and the average false alarm rate is approximately 44.22%. Therefore, the method in this embodiment significantly reduces the false alarm rate while maintaining a low false alarm level, making it suitable for fault monitoring in complex industrial processes.

[0031] In another optional embodiment, the number of principal components can be determined based on different cumulative variance contribution rates such as 95%, 97%, and 99%; the principal component space detector (isolated forest detection model) can be replaced by a type of support vector machine, elliptical envelope, autoencoder, or other global anomaly detection model; the residual space detector (local outlier detection model) can be replaced by an unsupervised detection model based on k-nearest neighbor distance, density peak, or local kernel density estimation; the fusion weights can be determined by variance, entropy, validation set error, stability index, or expert rules; the threshold can be a quantile threshold, mean doubled standard deviation, extreme value theory, or sliding window adaptive threshold.

[0032] This invention provides a residual dual-space industrial process fault detection system based on IFOrest-LOF, comprising a data acquisition module, a data preprocessing module, a PCA dual-space construction module, a principal component space isolated forest detection module, a residual space local outlier factor detection module, an anomaly score fusion module, and a fault discrimination output module. The data acquisition module is used to collect multivariate operational data of industrial processes; The data preprocessing module is used to perform preprocessing on the collected multivariate runtime data; The PCA dual-space construction module is used to generate principal component space features through principal component analysis and to obtain residual space features through data reconstruction calculation. The principal component space isolated forest detection module has a built-in isolated forest detection model, which is used to calculate and output the first anomaly score based on the principal component space features; The residual space local outlier detection module has a built-in local outlier detection model, which is used to calculate and output the second outlier score based on residual space features; The anomaly score fusion module is used to perform normalization, enhancement, and adaptive weighted fusion on the first anomaly score and the second anomaly score in sequence to obtain a comprehensive anomaly score. The fault detection output module compares the comprehensive anomaly score with the fault judgment threshold and outputs the fault detection result.

[0033] Therefore, this invention adopts the above-mentioned residual dual-space industrial process fault detection method and system based on IForest-LOF, which 1) fully utilizes the complementary information of the principal component space and the residual space; the principal component space is used to characterize the global change pattern of the industrial process, and the residual space is used to capture local deviations and hidden anomalies not retained by the principal components, reducing the information loss caused by single-space detection; 2) different detection models are used for different spaces to improve the adaptability of anomaly patterns. IForest is suitable for identifying global sparse anomalies, and LOF is suitable for identifying local density anomalies. The combination of the two can take into account both global faults and local weak faults; 3) no large number of fault samples are required. IForest and LOF are both unsupervised anomaly detection models, which can be mainly trained based on normal operating condition data, suitable for the problem of scarce fault samples in actual industrial scenarios; 4) by score normalization, posterior enhancement and adaptive weight fusion, the influence of the difference in score scale of different models on the detection results is reduced, and the separability between fault samples and normal samples is improved.

[0034] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0035] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A residual double space industrial process fault detection method based on IForest-LOF, characterized in that, Includes the following steps: S1. Collect multivariate operational data of industrial processes and perform standardized preprocessing to form a sample matrix; S2. Establish a principal component analysis model based on training data under normal operating conditions and construct a projection matrix according to the principal components; project the sample matrix onto the principal component space to obtain the features of the principal component space. The sample matrix is ​​reconstructed using the principal component space features and the projection matrix to obtain the reconstructed data. The difference between the sample matrix and the reconstructed data is calculated to obtain the residual space features. S3. Train the isolated forest detection model on the principal component space features and output the first anomaly score; train the local outlier detection model on the residual space features and output the second anomaly score. S4. After normalizing the first abnormal score and the second abnormal score and performing posterior probabilistic augmentation processing, adaptive weighted fusion is performed to obtain a comprehensive abnormal score. S5. Using the operating data during the normal operation phase, perform steps S1-S4 to obtain the comprehensive anomaly score for normal operation. Determine the fault judgment threshold based on the distribution of the comprehensive anomaly score for normal operation. Perform steps S1-S4 on the sample to be tested to obtain the comprehensive anomaly score for the sample to be tested. Compare the comprehensive anomaly score of the sample to be tested with the fault judgment threshold and output the industrial process fault detection result.

2. The residual bi-space industrial process fault detection method based on IForest-LOF according to claim 1, wherein, In step S2, the principal components are obtained as follows: Calculate the covariance matrix of the training data under normal operating conditions and perform eigenvalue decomposition. Select the top k principal components according to the cumulative variance contribution rate; the cumulative variance contribution rate is 95%, 97%, or 99%.

3. The residual bi-space industrial process fault detection method based on IForest-LOF according to claim 1, wherein, In step S3, the isolated forest detection model constructs multiple random isolated trees and calculates the first anomaly score based on the average path length of the sample in each isolated tree. This score is used to identify global sparse anomalies in the principal component space that deviate from the normal global change pattern.

4. The residual bi-space industrial process fault detection method based on IForest-LOF according to claim 1, wherein, In step S3, the local outlier detection model calculates the reachability distance, local reachability density, and relative density deviation between a sample and its k nearest neighbors to obtain a second outlier score, which is used to identify local fine-grained density deviations in the residual space.

5. The residual dual-space industrial process fault detection method based on IFOres-LOF according to claim 1, characterized in that, In step S4, the normalization process is Min-Max normalization; the posterior probabilistic enhancement process is implemented based on the Sigmoid function.

6. The residual dual-space industrial process fault detection method based on IFOres-LOF according to claim 1, characterized in that, In step S4, the weights of the adaptive weighted fusion are determined based on the statistical uncertainty of the abnormal scores after the two feature spaces are enhanced. The feature space with lower statistical uncertainty corresponds to a higher fusion weight. The statistical uncertainty is characterized by the variance or standard deviation of the abnormal scores.

7. The residual dual-space industrial process fault detection method based on IFOres-LOF according to claim 1, characterized in that, In step S5, the fault judgment threshold is determined based on the distribution of the comprehensive anomaly score during normal operation. The values ​​corresponding to the top 5 percentiles of the comprehensive anomaly score during normal operation, arranged from high to low, are selected as the fault judgment threshold. When the overall anomaly score of the sample under test is greater than the fault judgment threshold, it is judged as a fault; when the overall anomaly score of the sample under test is less than or equal to the fault judgment threshold, it is judged as normal.

8. The residual dual-space industrial process fault detection method based on IFOres-LOF according to claim 1, characterized in that, The isolated forest detection model was replaced with a global anomaly detection model that includes a class of support vector machines, elliptical envelopes, and autoencoders; the local outlier detection model was replaced with an unsupervised detection model that includes k-nearest neighbor distance, density peaks, and local kernel density estimation; the fault judgment threshold adopted quantile threshold, mean plus standard deviation, extreme value theory, or sliding window adaptive threshold.

9. A residual dual-space industrial process fault detection system based on IFOres-LOF, employing the method described in any one of claims 1-8, characterized in that, It includes a data acquisition module, a data preprocessing module, a PCA dual-space construction module, a principal component spatial isolation forest detection module, a residual spatial local outlier detection module, an anomaly score fusion module, and a fault discrimination output module; The data acquisition module is used to collect multivariate operational data of industrial processes; The data preprocessing module is used to perform preprocessing on the collected multivariate runtime data; The PCA dual-space construction module is used to generate principal component space features through principal component analysis and to obtain residual space features through data reconstruction calculation. The principal component space isolated forest detection module has a built-in isolated forest detection model, which is used to calculate and output the first anomaly score based on the principal component space features. The residual space local outlier detection module has a built-in local outlier detection model, which is used to calculate and output the second outlier score based on the residual space features. The anomaly score fusion module is used to perform normalization, enhancement and adaptive weighted fusion on the first anomaly score and the second anomaly score in sequence to obtain a comprehensive anomaly score; The fault detection output module is used to compare the comprehensive anomaly score with the fault judgment threshold and output the fault detection result.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the residual dual-space industrial process fault detection method based on IFOrest-LOF as described in any one of claims 1 to 8.

Images