Early fault detection method for incomplete complex industrial system data

Abstract

The application discloses an early fault detection method for incomplete complex industrial system data and belongs to the technical field of industrial process monitoring and fault diagnosis.The early fault detection method comprises the following steps: constructing a missing indication matrix and a neighborhood probability mask matrix for incomplete process variable data, repairing the data by using a pre-trained position weighting model, obtaining low-dimensional potential features under the constraint of a pre-trained multi-scale structure graph by guiding feature mapping through a regularization term, calling a pre-constructed quality-related / quality-unrelated subspace statistical model, calculating monitoring statistics and comparing the monitoring statistics with a control threshold, and determining that an early fault exists if any statistic exceeds the limit.The application considers data repair reliability and fault weak feature reservation, improves early fault detection sensitivity and accuracy, reduces the false alarm rate, and is suitable for complex industrial systems such as chemical production and wind power generation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial process monitoring and fault diagnosis, and particularly relates to an early fault detection method for incomplete complex industrial system data. BACKGROUND

[0002] With the continuous improvement of industrial automation, informatization and networking level, complex industrial systems show the characteristics of large equipment scale, large number of variables, strong process coupling and complex running mechanism. Industrial Internet of Things (IIoT) continuously collects industrial process running state and product quality related indicators by deploying a large number of sensors and communication devices in the industrial field, which provides a data basis for industrial process monitoring, fault diagnosis and quality control. Under this background, data-driven industrial process fault detection methods have been widely studied and applied. In the prior art, traditional statistical process monitoring methods (such as principal component analysis, partial least squares analysis, etc.) establish statistical models based on multivariate data to detect and analyze abnormal states of the system. Such methods usually assume that process data is complete and available, and the statistical relationship between variables remains stable under normal working conditions. However, in actual industrial environments, due to sensor failure, communication interruption, network congestion or abnormal acquisition equipment, process data often has different forms of missing, discontinuity or local unobservability. When missing data is directly involved in modeling or monitoring, it is easy to destroy the dynamic correlation between variables, resulting in failure of the statistical model, thereby significantly reducing the reliability of fault detection.

[0003] To cope with the problem of data missing, some existing technologies introduce missing data filling methods to repair the missing data through interpolation, low-rank reconstruction or model prediction, and then perform fault detection based on the repaired data. However, the above methods usually take minimizing the global reconstruction error as the goal, without distinguishing the reliability difference of data at different missing positions, and without fully considering the influence of missing data on the preservation of early fault features. In complex industrial systems, early faults often manifest as weak abnormal signals with small amplitude and slow evolution, which are easy to be smoothed or covered in the filling process, thereby causing delay or even missing detection of early faults.

[0004] In recent years, deep learning-based fault detection methods have gradually attracted attention. Autoencoders, denoising autoencoders and their variants improve the modeling ability and robustness of complex industrial processes to some extent through nonlinear feature learning capabilities. However, this kind of method usually focuses on overall data reconstruction or feature compression in design, lacks explicit constraints on the multi-scale structure characteristics and time sequence correlation of industrial processes, and is difficult to stably retain the local structure disturbance features corresponding to early faults. Under the condition of incomplete data, the above problems are further aggravated, so that the information related to early faults in the deep features is easily overwhelmed by normal fluctuations and noise. In addition, after the abnormality of the industrial system occurs, it is of great significance to accurately identify the key variables related to the fault to guide equipment maintenance and operation decision-making.

[0005] In summary, under the condition of incomplete data of complex industrial systems, the prior art generally has the following deficiencies: first, there is a lack of processing mechanism that can balance data repair reliability and early fault weak feature retention; second, there is a lack of effective modeling means for the multi-scale structure characteristics and time sequence correlation of industrial processes. Therefore, it is necessary to propose a new technical scheme to maintain the structure characteristics of industrial processes under incomplete observation conditions and realize reliable detection of early faults of complex industrial systems. SUMMARY

[0006] The technical purpose of the present application is to solve the above problems in the prior art. The present application proposes a quality-related early fault detection method for incomplete industrial Internet of Things data, which maintains the structure characteristics of industrial processes under incomplete observation conditions and realizes reliable detection of early faults of complex industrial systems.

[0007] To achieve the above technical purpose, the embodiments of the present application adopt the following technical scheme.

[0008] The embodiment provides an early fault detection method for incomplete complex industrial system data, comprising:

[0009] Online acquisition of process variable data of a complex industrial system, the process variable data being incomplete;

[0010] Based on a pre-set missing indicator matrix construction rule, a missing indicator matrix is constructed for the process variable data, the missing indicator matrix being used to represent the observable state of each process variable data at each time point in the process variable data;

[0011] Based on a pre-determined time neighborhood set prior probability distribution and the missing indicator matrix, a neighborhood probability mask matrix is determined, the neighborhood probability mask matrix being used to measure the reliability of the missing position in the process variable data;

[0012] Call a pre-trained position-weighted data repair model to repair the incomplete process variable data in combination with the neighborhood probability mask matrix;

[0013] Determine the numerical similarity and time sequence correlation between the process variables, and perform weighted fusion to form a multi-scale structure graph, and under the constraint of the multi-scale structure graph, feature mapping is performed on the repaired process variable data by introducing a structure graph regularization term based on the structure graph, to obtain low-dimensional latent features of the process variable data;

[0014] Call a pre-constructed quality-related subspace and quality-independent subspace statistical monitoring model to calculate the monitoring statistics corresponding to the low-dimensional latent features, and compare the monitoring statistics with the pre-set control threshold; if any monitoring statistics exceeds the corresponding control threshold, it is determined that there is an early fault in the system.

[0015] Compared with the prior art, the present application has the following beneficial effects: (1) by introducing a neighborhood probability guided data repair mechanism, the time continuity and variable correlation of the industrial process can be maintained under the condition of structured missing data, and the data repair reliability is improved; (2) by multi-scale structure graph regularization auto-encoding feature learning, the early fault features of weak, small and slow evolution are effectively preserved; (3) by using a pre-constructed quality-related subspace and quality-independent subspace statistical monitoring model, quality-related and non-quality-related changes are clearly separated, and the early fault detection sensitivity is improved; (4) by introducing a confidence-weighted fault contribution index, the fault variable identification accuracy is significantly improved, and the false alarm rate is reduced.

[0016] It should be understood that the summary section is not intended to identify key or essential features of embodiments of the disclosure, nor is it intended to limit the scope of the disclosure. Other features of the disclosure will become readily apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0017] The drawings described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure in any way. In addition, the shapes and scale sizes of the components in the drawings are only illustrative and are used to help understand the application, and are not specific limitations on the shapes and scale sizes of the components of the application. Those skilled in the art can select various possible shapes and scale sizes to implement the application according to specific circumstances under the teaching of the application. In the drawings:

[0018] Figure 1 The early fault detection method flowchart for incomplete complex industrial system data provided for the embodiments. DETAILED DESCRIPTION

[0019] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0020] The embodiments provide an early fault detection method for incomplete and complex industrial system data, including:

[0021] S1: Online acquisition of process variable data from complex industrial systems; process variable data is incomplete.

[0022] S2: Based on the pre-defined missing indicator matrix construction rules, a missing indicator matrix is ​​constructed for the process variable data. The missing indicator matrix is ​​used to characterize the observable state of each process variable data at each time point in the process variable data.

[0023] S3: Determine the neighborhood probability mask matrix based on the prior probability distribution on the predetermined time neighborhood set and the missing indication matrix. The neighborhood probability mask matrix is ​​used to measure the reliability of missing positions in process variable data.

[0024] S4: Call the pre-trained position-weighted data repair model and combine it with the neighborhood probability mask matrix to repair incomplete process variable data;

[0025] S5: Determine the numerical similarity and temporal correlation between process variables, and perform weighted fusion to form a multi-scale structure graph. Under the constraints of the multi-scale structure graph, feature mapping is performed on the repaired process variable data by introducing a structure graph-based regularization term to obtain the low-dimensional potential features of the process variable data.

[0026] S6: Call the pre-built statistical monitoring model of the quality-related subspace and the quality-independent subspace, calculate the monitoring statistics corresponding to the low-dimensional potential features, and compare the monitoring statistics with the pre-set control thresholds; if any monitoring statistic exceeds its corresponding control threshold, it is determined that the system has an early fault.

[0027] In the embodiments, such as Figure 1 As shown, this includes the offline modeling stage, specifically:

[0028] Step 101: Collect multiple process variable data of the complex industrial system under normal operating conditions to form a process variable data matrix. Simultaneously, a data matrix of quality variables related to product quality was collected. .

[0029] Step 102: Constructing the missing indicator matrix , which is used to represent the observable state (observable value or missing value) of each process variable data corresponding to each time point in the process data, wherein the element in the missing indicator matrix is used to indicate whether the corresponding position data is missing.

[0030] (1)

[0031] wherein t represents the sampling time index, t = 1, 2,..., T t , T t is the total number of sampling points; d represents the process variable index, d = 1, 2,..., D, D is the total number of process variables, constitutes the missing indicator matrix , which is used to represent the observable state of the dth process variable in the time neighborhood, and the value is 1 when the data is observable, otherwise the value is 0.

[0032] Step 103: Constructing the neighborhood probability mask matrix based on the time neighborhood relationship. Let the time neighborhood centered at the current sampling time t be:

[0033] (2)

[0034] wherein represents the sampling time index in the time neighborhood, is the time range covered by the neighborhood; is the time neighborhood set centered at the t sampling time.

[0035] According to the distribution of observable data of the dth process variable in the time neighborhood, the normalized weight coefficient in the neighborhood is defined as:

[0036] (3)

[0037] can be regarded as the prior probability distribution of the th element in the time neighborhood set centered at the t sampling time, prior probability distribution of the th element in the time neighborhood set centered at the t sampling time.

[0038] Further, the data reliability probability of the missing position (t, d) is defined as:

[0039] (4)

[0040] wherein ​​denotes the expected probability value of the dth process variable being in observable state within a time neighborhood centered at the tth sampling time, which is used to measure the data reliability of the missing position (t, d). The matrix thus formed is denoted as The set forms a neighborhood probability mask matrix , which is used for weight constraint in subsequent data repair and fault analysis process, is an element in the missing indication matrix .

[0041] Step 104: Based on the neighborhood probability mask matrix constructed in step S103 , a position-weighted reconstruction loss function is constructed , and a position-weighted data repair model is trained. The position-weighted data repair model repairs incomplete industrial process data according to the autoencoder method and the position-weighted reconstruction loss function . By imposing a higher weight on the missing data position, the model is guided to focus on the reconstruction accuracy of the missing data, thereby maintaining the time continuity and variable correlation of the industrial process in the repair process and avoiding the over-smoothing of weak features of early faults.

[0042] In some embodiments, the expression of the position-weighted reconstruction loss function is as follows:

[0043] (5)

[0044] wherein is the repaired data. is the original data. The feature space is composed of the values of the repaired process data in the time dimension, T t denotes the total number of sampling points.

[0045] Step 105: Based on the repaired data, a multi-scale structure graph model is constructed. The numerical similarity and temporal correlation between process variables are calculated respectively at different time scales, and the various similarities are weighted and fused to form a multi-scale structure graph, which is used to depict the correlation characteristics of industrial process variables at different time scales.

[0046] Let the numerical similarity and temporal correlation between process variable i and process variable j at the mth time scale be defined as:

[0047] Numerical similarity (feature space proximity):

[0048] (6)

[0049] wherein is the geometric neighborhood similarity at scale m, is the set neighborhood radius, measuring the Euclidean distance of two process variables in the feature space, is the ith process variable, is the jth process variable, .

[0050] Temporal correlation (time proximity):

[0051] (7)

[0052] where, and are the temporal index positions of the ith process variable and the jth process variable, respectively. is the set temporal radius. is the temporal neighborhood similarity at scale m, The smaller the difference is, the smaller the difference is.

[0053] (8)

[0054] where is the weighted similarity at scale m, and β is the set balance parameter for adjusting the relative influence of numerical similarity and temporal correlation in structure modeling. is the degree matrix, which is a diagonal matrix whose diagonal elements are ; is the weighted similarity matrix at scale m, is the structure graph at the mth scale (graph Laplacian matrix).

[0055] Step 106: Perform feature learning under the constraint of multi-scale structure graph. By introducing a structure graph-based regularization term, process variables with high similarity in the original data space can maintain a neighborhood relationship in the low-dimensional feature space, thereby reducing the feature dimension while preserving the structural stability and local abnormal disturbance characteristics of the industrial process.

[0056] Let the representation of the sample in the low-dimensional latent space be and , is the feature vector of the ith sample in the low-dimensional latent space, is the feature vector of the jth sample in the low-dimensional latent space; define the graph regularization term as:

[0057] (9)

[0058] The graph regularization term is combined with the position-weighted reconstruction loss function of the data repair model to construct the overall optimization objective function:

[0059] (10)

[0060] wherein, is a regularization parameter, used to balance the data reconstruction accuracy and the structure preserving ability. is the final multi-scale structure graph (graph Laplacian matrix). By optimizing the above objective function, the structural stability of the industrial process and the local abnormal disturbance features corresponding to the early faults are effectively preserved while reducing the feature dimension.

[0061] Step S107: Based on the low-dimensional latent feature representation obtained in step S106, further mine the internal correlation between the latent features and the quality variables, and perform subspace decomposition on the latent feature space. Perform joint decomposition on (Z, Y) by generalized singular value decomposition method:

[0062] (11)

[0063] Based on the generalized singular value decomposition method, a set of interrelated transformation matrices are obtained. The process variable data matrix Z and the quality variable data matrix Y are projected into a joint mathematical space, so that the internal correlation between them can be analyzed. P and Q are the left and right singular matrices, respectively, and characterize the diagonal matrix.

[0064] Further, the quality-related subspace features of the latent features are obtained.

[0065] (12)

[0066] wherein r is the number of quality variables. represents the low-dimensional latent feature vector of the t-th sample in the quality-related subspace. represents the quality-related subspace feature of the latent features, represents the submatrix composed of the first r columns extracted from the transformation matrix . represents the projection matrix constructed, and the number of matrix columns is determined according to the number of quality variables.

[0067] According to the subspace decomposition form, the quality-related subspace statistic and the quality-unrelated subspace statistic are obtained, and the corresponding statistical monitoring models (as shown in formula 13) are constructed, wherein different monitoring models (quality-related subspace and quality-unrelated subspace statistical monitoring models) include the corresponding subspace statistics and the calculation method of the threshold value. The two statistical monitoring models are used to distinguish abnormal changes affecting product quality from non-quality-related disturbances.

[0068] (13)

[0069] wherein, is the control threshold of the quality-independent subspace; is the covariance matrix representing the potential characteristics of the quality-dependent subspace. is the Hotelling’s T 2 statistic based on the quality-dependent subspace. is the statistic of the quality-independent subspace. is the significance level, N is the number of normal operating condition samples (process variable data). The F distribution with degrees of freedom r and N-r corresponds to the quantile value at the confidence level . is the control threshold of the quality-dependent subspace. D is the total number of process variables. is the correction coefficient for approximating the residual statistical distribution shape, ; the intermediate variable , , , respectively represent the first, second and third order statistics of the unmodeled subspace eigenvalues, , is the inverse cumulative distribution function representing the standard normal distribution, represents the upper quantile of the F distribution with degrees of freedom r and N-r at the significance level .

[0070] In the embodiment, the process variable data collected online is subjected to structured missing data repair and feature mapping processing consistent with the offline stage by steps S2, S3, S4 and S5, to obtain the potential feature representation of the process variable data.

[0071] The construction rule of the missing indicator matrix in the online stage is that the matrix element takes value 1 or 0, wherein 1 represents that the corresponding position data is observable, and 0 represents that the corresponding position data is missing; the construction rule is consistent with the rule of constructing the missing indicator matrix in the offline stage, such as formula (1).

[0072] (14)

[0073] wherein, represents the feature deviation of the dth process variable in the potential feature space, which is calculated as the difference between the potential feature of the process variable data and the mean value of the normal operating condition process variable data. is the mean value of the potential feature of the process variable data. is the sample mean of the dth process variable potential eigenvalue in the historical normal operating condition process variable data, The expected probability value of the dth process variable being in an observable state in a time neighborhood centered at the tth sampling time can be used as a data reliability weight of the missing position (t, d).

[0074] In step S6, based on the quality-related subspace and the quality-independent subspace statistical monitoring models constructed on the normal operating condition data in the offline modeling stage, the corresponding monitoring statistics are calculated, and the control thresholds of the monitoring statistics are set in advance according to the statistical distribution of the normal operating condition process variable data.

[0075] In the online monitoring process, the monitoring statistics calculated in real time are compared with the corresponding control thresholds; when any monitoring statistic exceeds the corresponding control threshold, it is determined that the abnormal state of the industrial system occurs, thereby realizing the real-time operation monitoring and early fault detection of the complex industrial system, which is represented as:

[0076] If or , the system is determined to be abnormal. Hotelling’s T 2 statistic of the quality-related subspace of the tth online process variable data, and the quality-independent subspace statistic of the tth online process variable data.

[0077] At present, it is difficult to realize stable and accurate fault variable positioning under incomplete data conditions after a fault occurs. The existing fault variable positioning methods are mostly based on statistical contribution analysis or feature deviation measurement, but under incomplete data conditions, the variable contribution results are easily disturbed due to the uncertainty introduced by the data repair process and the existence of deep feature aliasing problem, resulting in unstable positioning results, and even introducing variables unrelated to the actual fault, reducing the engineering application value.

[0078] Therefore, in some embodiments, in order to realize accurate positioning of early faults of complex industrial systems, the method further comprises: when the monitoring statistics exceed the corresponding control thresholds, based on the reliability weights obtained in the data repair stage, the deviations of the process variables in the potential feature space are analyzed, and a fault contribution index reflecting the abnormal contribution degree of the variables is constructed ; according to the size of the fault contribution index , the process variables are threshold discriminated, thereby identifying the key fault variables highly related to the current abnormal state, and realizing accurate positioning of the fault variables.

[0079] (15)

[0080] When , judge the dth process as a fault-related variable, d = 1, 2, …, D, D is the total number of process variables, represents the mean of the potential feature of the dth variable under normal working condition; represents the contribution index of the dth variable in the nth sample; represents the set of contribution indexes of all samples.

[0081] In formula (15), the threshold function is determined in advance based on the statistical distribution characteristics of the process variable data under normal working condition, and is used to distinguish normal fluctuation variables from abnormal contribution variables.

[0082] The embodiment aims at the problems that early faults are difficult to detect and fault variables are difficult to accurately locate in complex industrial systems under incomplete data conditions, and proposes an early fault detection and positioning method for incomplete complex industrial system data. The method takes multivariate time series data of industrial processes as the research object, constructs the overall framework around the technical main line of "structured missing data repair-multiple scale structure feature preservation-quality related change separation-reliable fault positioning", guarantees the reliability of data repair, maximizes the preservation of weak evolution features corresponding to early faults, and realizes high-sensitivity detection and accurate positioning of early faults in complex industrial systems.

[0083] The early fault detection method for incomplete complex industrial system data provided by the present application is described in detail above, specific examples are applied in this paper to explain the principles and implementation modes of the present application, and the above description of the embodiments is only used to help understand the concept of the present application, and should not be understood as limiting the protection scope of the present application.

Claims

1. A method for early fault detection oriented to incomplete complex industrial system data, characterized in that, The method comprises the following steps: Collecting process variable data of a complex industrial system online, the process variable data being incomplete; Constructing a missing indicator matrix for the process variable data based on a pre-set construction rule, the missing indicator matrix being used to represent the observable state of each process variable data at each time point; Determining a neighborhood probability mask matrix based on a pre-determined prior probability distribution on a time neighborhood set and the missing indicator matrix, the neighborhood probability mask matrix being used to measure the reliability of the missing position in the process variable data; Invoking a pre-trained position-weighted data repair model to repair the incomplete process variable data in combination with the neighborhood probability mask matrix; Determining the numerical similarity and time sequence correlation between process variables, and performing weighted fusion to form a multi-scale structure graph, and then introducing a structure graph regularization term to perform feature mapping on the repaired process variable data under the constraint of the multi-scale structure graph, to obtain low-dimensional latent features of the process variable data; Invoking a pre-constructed quality-related subspace and quality-independent subspace statistical monitoring model to calculate monitoring statistics corresponding to the low-dimensional latent features, and comparing the monitoring statistics with pre-set control thresholds; If any monitoring statistic exceeds the corresponding control threshold, it is determined that there is an early fault in the system; The method further comprises the following steps: calculating a fault contribution index of each process variable data in the latent feature space based on a pre-determined reliability weight rule, the fault contribution index reflecting the contribution degree of each variable to the fault; and discriminating the fault contribution index based on a pre-set threshold function to screen out key fault variables and realize fault localization. The training target of the position-weighted data repair model is to minimize a position-weighted reconstruction loss function, which applies a higher weight to the missing data position than to the complete data position. The expression of the position-weighted reconstruction loss function is as follows: ； wherein is the repaired process variable data, is the original process variable data; represents the expected probability value that the dth process variable is in an observable state within a time neighborhood centered at the tth sampling time, d = 1, 2,..., D, D is the total number of process variables; t = 1, 2,..., T t , T t represents the total number of sampling points; The set forms a neighborhood probability mask matrix.

2. The early fault detection method of claim 1, wherein, The fault contribution index is calculated as the product of the deviation of each process variable data in the latent feature space and the corresponding reliability weight, the deviation being the difference between the latent feature of each process variable data and the mean value of the latent feature of the process variable data under normal working conditions.

3. The early fault detection method of claim 1, wherein, The method further comprises the following step: the missing indicator matrix construction rule is that the matrix element takes a value of 1 or 0, where 1 represents that the corresponding position data is observable, and 0 represents that the corresponding position data is missing.

4. The method of claim 1, wherein, The monitoring statistics include statistics of a quality-related subspace and statistics of a quality-independent subspace; the control threshold corresponding to the statistics of the quality-related subspace is an F distribution quantile value corresponding to the degrees of freedom of the number of quality variables r and the number of offline normal process variable data N-r; and the control threshold corresponding to the statistics of the quality-independent subspace is calculated by the inverse cumulative distribution function of the standard normal distribution and a correction coefficient.

Images