Industrial fault detection method based on structured augmented dictionary learning

Abstract

The invention discloses an industrial fault detection method based on structured augmented dictionary learning, and the method comprises the steps: firstly processing a fault-free component through a basic dictionary, and introducing an additional low-dimensional sparse dictionary for precisely reconstructing a fault related part, thereby avoiding the decentralization of fault information. Secondly, in order to ensure that a fault-free part is in a statistical control state, a steady-state statistical characteristic and a manifold structure constraint are embedded into a basic dictionary coefficient, and meanwhile, a fault-related characteristic variable is accurately selected through a hard sparse constraint, so that a fault-free signal in an industrial process is effectively prevented from being interfered; therefore, accurate separation of the fault and the normal component of the signal is realized in sparse decoding, so that the fault mode is accurately revealed; in addition, an aggregation method based on a moving window is adopted, the sensitivity and detection efficiency of initial faults in the industrial process are remarkably improved, and an efficient and reliable fault diagnosis solution is provided for industrial application.

Description

Technical Field

[0001] The invention belongs to the technical field of fault detection, and in particular relates to an industrial fault detection method based on structured augmented dictionary learning. Background Art

[0002] The increasing complexity and automation of modern industrial processes have put forward higher requirements for real-time monitoring technology to ensure the stability and operation efficiency of the system. Most fault monitoring methods are based on data-driven models, such as principal component analysis, partial least squares regression, and canonical correlation analysis. These methods extract key features to monitor process status by projecting data into a low-dimensional space. However, low-dimensional space often cannot fully express the diversity and potential characteristics of process data, and it is easy to lose key information. For complex industrial processes, this is especially evident when dealing with high-dimensional and multimodal data.

[0003] In order to make up for this deficiency, by projecting the data into a high-dimensional space, the complex relationship between variables can be decoupled more efficiently, subtle data changes can be captured, and deep features can be extracted. This idea has promoted the widespread application of high-dimensional modeling techniques such as kernel methods, deep learning, and dictionary learning in the field of industrial fault detection. Among them, dictionary learning, as a powerful signal processing tool, has been widely used for its simple model structure and excellent monitoring performance. In practical applications, dictionary learning provides appropriate high-dimensional sparse representation for the original data through matrix theory and sparse induction technology. By linearly combining the coefficient matrix, the atoms of the complete dictionary can not only accurately reconstruct normal data, but also produce significant reconstruction errors for abnormal data, thereby achieving effective fault detection, showing great potential in the field of process monitoring.

[0004] Although dictionary learning-based fault monitoring techniques have made significant progress, they still have the following limitations: 1) Sparse coding technology (such as or Regularization) may suppress critical small anomalous changes in the coefficients that are important indicators of early / initial failure; 2) Improving detection sensitivity by enhancing the significance of reconstruction errors may weaken the ability of dictionary coefficients to distinguish initial faults, and vice versa.

[0005] 3) Fixing the fault direction or ignoring the interaction between variables when performing iterative reconstruction may destroy the internal structure and properties of the data, resulting in inaccurate isolation results.

[0006] For industrial processes with time-varying characteristics, existing methods have certain limitations in real-time monitoring and cannot fully meet the needs of fault detection in complex dynamic environments. Summary of the invention

[0007] The object of the present invention is to provide an industrial fault detection method based on structured augmented dictionary learning.

[0008] The present invention provides an industrial fault detection method based on structured augmented dictionary learning, which comprises the following steps: Step 1: collect the working conditions of the industrial process under normal operation through multi-dimensional sensors and construct a training sample data set, build a basic dictionary model based on the training sample data set, and obtain a basic dictionary; Step 2: Collect the real-time monitored working condition signals, and build a structured augmented dictionary learning model based on the collected real-time working condition signals to obtain a low-dimensional sparse dictionary and extended sparse coding; Step 3: Obtain monitoring statistics based on the low-dimensional sparse dictionary and extended sparse coding, and use the monitoring statistics to determine whether there is an industrial fault; if there is an industrial fault, execute step 4; Step 4: Construct an optimization model of the fault indication matrix based on monitoring statistics to obtain the fault indication matrix; indicate unqualified working conditions according to the fault indication matrix to complete accurate detection of industrial faults.

[0009] Preferably, in step 2, the structured augmented dictionary learning model constructed is: in, Collected by moving window L The data matrix consists of samples; for t The data matrix at each moment; L is the size of the sliding window; To expand the dictionary, ; is a low-dimensional sparse dictionary; As the basic dictionary; For and expand the dictionary The corresponding extended sparse coding, ; For and dictionary The corresponding main code, ; for t The master code of the moment; For low-dimensional sparse dictionary The corresponding extended code is ; for t Extended encoding of moments; is a low-dimensional sparse dictionary No. k atoms; For atoms The index of zero in ; For atoms The index set with zero in ; , and are model parameters; ; Represents the square of the Frobenius norm.

[0010] Preferably, in step 1, the basic dictionary model constructed is: in, is a K-dimensional sparse coefficient vector; Represents the basic dictionary No. k atoms; ; N is the number of samples; and Respectively Sum of squares of norm Norm.

[0011] As a preferred embodiment, in step 3, the monitoring statistic adopts the Hotelling statistic Sum squared prediction error .

[0012] Preferably, in step 3, the Hotelling statistic Sum squared prediction error The method to obtain is as follows: in, Reconstruction error The mean of ; An accuracy matrix constructed for data collected from an industrial process under normal operating conditions.

[0013] As a preferred embodiment, in the step 3, the basis for judging whether there is an industrial failure is: if the Hotelling statistic Greater than a preset Hotelling statistic threshold or squared prediction error If it is greater than the preset square prediction error threshold, there is an industrial fault; otherwise, there is no industrial fault.

[0014] Preferably, in step 4, the fault indication matrix constructed The optimization model is as follows: in, Used to indicate faults that affect monitoring statistics; and is the model parameter.

[0015] Preferably, in step 4, when obtaining the fault indication matrix Then, according to the fault indication matrix Calculating Failure Score To assess the severity of the fault, the fault score The expression is: in, is the fault indication matrix i Line j The value of the column; ; is the control limit; is the hyperparameter of a specific process; b To monitor statistics.

[0016] Preferably, in step 4, the basis for indicating the unqualified working condition is: obtaining the fault score of each working condition according to the fault scoring formula ; If there is a working condition whose corresponding fault score is greater than the preset fault score threshold, the working condition is unqualified.

[0017] Preferably, in the step 2, the collected operating condition signal is segmented and processed through a sliding window to construct a structured augmented dictionary learning model.

[0018] The present invention has the following beneficial effects: 1. The present invention decomposes the sensor signals in the industrial process into fault-free and fault-related dictionary components by sequentially expanding the basic dictionary into low-dimensional sparse dictionary components, thereby avoiding the decentralization of fault information; at the same time, an aggregation method based on a moving window is adopted to solve the problems of untimely detection, high false alarm rate and high missed alarm rate of the existing methods, significantly improving the sensitivity and detection efficiency of early faults in the industrial process, and providing an efficient and reliable fault diagnosis solution for industrial applications.

[0019] 2. The present invention effectively ensures that the fault-free signal in the industrial process is not disturbed by applying statistical and geometric constraints to the basic dictionary coefficients, thereby achieving accurate separation of the fault and normal components of the signal in sparse decoding. At the same time, applying hard sparse constraints to the extended dictionary can mine the potential patterns in the fault-related signals and significantly improve the accuracy and robustness of fault isolation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The figure is an overall flow chart of the method of the present invention.

[0021] Figure 2 This is a schematic diagram of key variable collection in Example 1 of the present invention.

[0022] Figure 3 Schematic diagram of rough fault detection results in Example 1 of the present invention; wherein (a) is a schematic diagram of Hotelling statistics; and (b) is a schematic diagram of Hotelling statistics.

[0023] Figure 4 Schematic diagram of rough fault detection results obtained based on traditional methods; among them, (a) is a schematic diagram of the PCA detection method; (b) is a schematic diagram of the Basic detection method; (c) is a schematic diagram of the MWRBC detection method; (d) is a schematic diagram of the MCC detection method; and (e) is a schematic diagram of the KL detection method.

[0024] Figure 5 Schematic diagram of the accurate fault detection result in Example 1 of the present invention; wherein, (a) is a schematic diagram of the Hotelling statistic; and (b) is a schematic diagram of the Hotelling statistic.

[0025] Figure 6 Schematic diagram of accurate fault detection results obtained based on traditional methods; among them, (a) is a schematic diagram of the MWRBC-SPE detection method; (b) is a schematic diagram of the SpCP detection method; (c) is a schematic diagram of the DL-RBC detection method; (d) is a schematic diagram of the DL-IRBC detection method. DETAILED DESCRIPTION

[0026] The present invention will be further described below in conjunction with the accompanying drawings.

[0027] like Figure 1 As shown, an industrial fault detection method based on structured augmented dictionary learning includes the following steps: Step 1: Build a basic dictionary The multi-dimensional sensors are used to collect the working conditions of the industrial process under normal operation and construct a training sample data set. The samples in the training sample data set are continuous under multiple working conditions. n Dimensional signal , and solve the basic dictionary based on the following basic dictionary model : (1) in, is a K-dimensional sparse coefficient vector (K>10n); Representation dictionary No. k atoms; ; N is the number of samples; and Respectively Sum of squares of norm norm; and is the model parameter.

[0028] Step 2: Collect multi-dimensional sensor process signals for online monitoring and process the signals in segments through a sliding window; construct a structured augmented dictionary learning model as follows: (2) in, Collected by moving window L The data matrix consists of samples; for t The data matrix at each moment; L is the size of the sliding window; To expand the dictionary, ; is a low-dimensional sparse dictionary; For and expand the dictionary The corresponding extended sparse coding, ; For and dictionary The corresponding main code, ; for t The master code of the moment; For low-dimensional sparse dictionary The corresponding extended code is ; for t Extended encoding of moments; is a low-dimensional sparse dictionary No. k atoms; For atoms The index of zero in ; For atoms The index set with zero in ; are model parameters; ; Represents the square of the Frobenius norm.

[0029] Counting atoms The cumulative variance contribution rate of each element in , and the index set is determined based on the cumulative variance contribution rate . According to the structured augmented dictionary learning model, a low-dimensional sparse dictionary is obtained and extended sparse coding .

[0030] and is the data fidelity term. The former resolves the signal decomposition of the process, and the latter aims to eliminate the zero coefficient problem that may exist in the decoding process and avoid information loss. In order to enhance the credibility of the fault-free part, the manifold geometry constraint term is introduced. , which optimizes the local geometric relationship of the fault-free signal based on the characteristics of k-nearest neighbors. In order to enable the atoms of the sparse dictionary to robustly match the fault mode, the constraint term is introduced Adjust the difference between the sparse dictionary and its mean matrix. The model solving process involves the main encoding and extended encoding , which corresponds to the corresponding sub-problem in the structured augmented dictionary learning model, can be regarded as a LASSO problem to be solved. The column-by-column update method is used to improve the computational efficiency. Low-dimensional sparse dictionary The update involves two main steps: unconstrained optimization and convex projection.

[0031] Step 3: Rough fault detection According to the monitoring statistics, it is judged whether there is an industrial fault. In this embodiment, the Hotelling statistic is selected. Sum squared prediction error As the basis for judgment, the expression is: (3) in, Reconstruction error The mean of ; An accuracy matrix constructed for data collected from an industrial process under normal operating conditions.

[0032] Hotelling statistic Greater than a preset Hotelling statistic threshold or squared prediction error If the error is greater than the preset square prediction error threshold, it is considered that there is an industrial fault and step 4 is executed; otherwise, it is considered that there is no industrial fault. The presence of an industrial fault is determined by rough fault detection.

[0033] Compared with traditional monitoring statistics, the statistics of the present invention have the following characteristics: (1) The separability of sparse coding can prevent statistical constructions from interfering with each other; (2) Combining the variable correlations of sparse atoms and precision matrices, the important factors of process variation can be comprehensively integrated.

[0034] Step 4: Accurate fault detection Since the structured augmented dictionary learning model explicitly contains fault information, it can be simply obtained by checking the low-dimensional sparse dictionary or by comparing the reconstruction error To achieve accurate fault detection. In order to suppress abnormal disturbances in the process, the present invention uses a The regularized feature selection model removes non-fault features and robustly retains the relevant features of potential fault variables. At the same time, the similarity of fault modes between adjacent windows is considered to construct the fault indication matrix The optimization model is as follows: (4) in, Used to indicate faults that affect monitoring statistics; and is the model parameter.

[0035] Get the Hotelling statistic separately Sum squared prediction error Corresponding fault indication matrix ( or ); In order to facilitate the assessment of the severity of the fault, the fault scores of each working condition at different time points are obtained , whose expression is: in, is the fault indication matrix i Line j The value of the column; ; is the control limit; is the hyperparameter of a specific process; b To monitor statistics.

[0036] According to the fault scoring formula, the fault scores of each working condition at different time points are obtained respectively. If the fault score corresponding to a working condition is greater than the preset fault score threshold, the working condition is considered unqualified, thereby achieving accurate detection of faults in the industrial process. The cause of industrial faults can be determined through accurate fault detection.

[0037] Example 1 This embodiment is applied to a multi-dimensional sensor network at an industrial site. Multiple sensors can collect multi-dimensional signals of key variables in the industrial glass melting process in real time. The fault detection method includes the following steps: Step 1: Figure 2 As shown in Figure 1, in order to monitor the process status, multiple sensors are arranged in the container to collect data of 14 process variables, including multiple furnace temperatures ( T 1 ~ T 8 ) and coil power ( P 1 ~ P4 ) and glass viscosity μ , coil voltage v Under normal operation of industrial glass melting, data is recorded every 5 minutes, thus constructing a continuous training sample data set. X i , and solve the basic dictionary based on the basic learning dictionary model of formula (1) .

[0038] Step 2: Collect multiple sensor process signals for online monitoring and process the signals in segments, each segment is generated by a sliding window. The low-dimensional sparse dictionary is obtained based on the structured augmented dictionary learning model of formula (2). And the corresponding extended dictionary extension code .

[0039] Step 3: Based on Hotelling statistics Sum squared prediction error Determine whether there is an industrial failure; if the Hotelling statistic Greater than a preset Hotelling statistic threshold or squared prediction error If the square prediction error is greater than the preset threshold, it is considered that there is an industrial fault and step 4 is executed; otherwise, it is considered that there is no industrial fault. The fault detection results based on these two monitoring statistics are as follows: Figure 3 As shown in Figure 2, the fault detection rates are as high as 73.48% and 95.65% respectively; compared with the results obtained by traditional fault detection methods, Figure 4 As shown in the fault detection results, the fault detection rate of the present invention is significantly improved. This comparison fully highlights the excellent potential and advantages of the present invention in dealing with gradually developing or initial faults.

[0040] Step 4: Solve the fault indication matrix based on the constructed optimization model of the fault indication matrix , and calculate the fault score of each variable , thereby achieving accurate fault detection; the fault score calculation formula is used to obtain The score vector The results are as follows Figure 5 As shown. Figure 5 It can be seen that the present invention focuses precisely on the key problem variables in the initial stage: , and Further analysis showed that the initial cracks in the furnace vessel were clearly reflected in the readings of the 7th temperature sensor and the 1st and 4th coil sensors. The color depth directly reflects the degree of the fault, where the darker the color, the more serious the fault. Compared with the traditional fault detection method, Figure 6The score vectors shown in the figure cannot accurately detect the variable that has failed, and the false alarm rate is high. This result shows that the present invention has excellent accuracy and reliability in the detection and isolation of early faults.

Claims

1. An industrial fault detection method based on structured augmented dictionary learning, characterized in that: The following steps are involved: Step 1: collect the working conditions of the industrial process under normal operation through multi-dimensional sensors and construct a training sample data set, build a basic dictionary model based on the training sample data set, and obtain a basic dictionary; Step 2: Collect the real-time monitored working condition signals, and build a structured augmented dictionary learning model based on the collected real-time working condition signals to obtain a low-dimensional sparse dictionary and extended sparse coding; Step 3: Obtain monitoring statistics based on low-dimensional sparse dictionary and extended sparse coding, and use the monitoring statistics to determine whether there is an industrial fault; If there is an industrial fault, proceed to step 4; Step 4: construct an optimization model of the fault indication matrix based on monitoring statistics to obtain the fault indication matrix; According to the fault indication matrix, unqualified working conditions are indicated to complete the accurate detection of industrial faults.

2. The industrial fault detection method based on structured augmented dictionary learning according to claim 1 is characterized in that: In the step 2, the structured augmented dictionary learning model constructed is: in, Collected by moving window L The data matrix consists of samples; for t The data matrix at each moment; L is the size of the sliding window; To expand the dictionary, ; is a low-dimensional sparse dictionary; is the basic dictionary; For and expand the dictionary The corresponding extended sparse coding, ; For and dictionary The corresponding main code, ; for t The master code of the moment; For low-dimensional sparse dictionary The corresponding extended code is ; for t Extended encoding of moments; is a low-dimensional sparse dictionary No. k atoms; For atoms The index of zero in ; For atoms The index set with zero in ; , and are model parameters; ; Represents the square of the Frobenius norm.

3. The industrial fault detection method based on structured augmented dictionary learning according to claim 1 is characterized in that: In the step 1, the basic dictionary model constructed is: in, is a K-dimensional sparse coefficient vector; Represents the basic dictionary No. k atoms; ; N is the number of samples; and Respectively Sum of squares of norm Norm.

4. The industrial fault detection method based on structured augmented dictionary learning according to claim 1 is characterized in that: In step 3, the monitoring statistic adopts Hotelling statistic Sum squared prediction error .

5. The industrial fault detection method based on structured augmented dictionary learning according to claim 4 is characterized in that: In step 3, the Hotelling statistic Sum squared prediction error The method to obtain is as follows: in, Reconstruction error The mean of ; An accuracy matrix constructed for data collected from an industrial process under normal operating conditions.

6. The industrial fault detection method based on structured augmented dictionary learning according to claim 5 is characterized in that: In step 3, the basis for judging whether there is an industrial failure is: if the Hotelling statistic Greater than a preset Hotelling statistic threshold or squared prediction error If it is greater than the preset square prediction error threshold, there is an industrial fault; otherwise, there is no industrial fault.

7. The industrial fault detection method based on structured augmented dictionary learning according to claim 1 is characterized in that: In step 4, the fault indication matrix constructed The optimization model is as follows: in, Used to indicate faults that affect monitoring statistics; and is the model parameter.

8. The industrial fault detection method based on structured augmented dictionary learning according to claim 1 is characterized in that: In the step 4, when obtaining the fault indication matrix Then, according to the fault indication matrix Calculating Failure Score To assess the severity of the fault, the fault score The expression is: in, is the fault indication matrix i Line j The value of the column; ; is the control limit; is the hyperparameter of a specific process; b To monitor statistics.

9. The industrial fault detection method based on structured augmented dictionary learning according to claim 8, characterized in that: In the step 4, the basis for indicating the unqualified working condition is: the fault score of each working condition is obtained according to the fault scoring formula ; If there is a working condition whose corresponding fault score is greater than the preset fault score threshold, the working condition is unqualified.

10. The industrial fault detection method based on structured augmented dictionary learning according to claim 1, characterized in that: In the step 2, the collected operating condition signal is processed in segments through a sliding window to construct a structured augmented dictionary learning model.

Images