Multi-working-condition process monitoring method and system based on zero-forgetting continuous dictionary learning

Abstract

The invention discloses a multi-working-condition process monitoring method and system based on zero-forgetting continuous dictionary learning, and the method comprises the steps: carrying out the offline modeling: firstly, carrying out the dictionary learning through employing an initial working condition data set, obtaining an initial dictionary, carrying out the decomposition, obtaining a low-rank matrix of an initial working condition, and calculating the control limit of the initial working condition through calculating a sample reconstruction error; performing incremental updating on the low-rank matrix obtained by learning the old working condition by using the new working condition data set to obtain a low-rank matrix of a new working condition, and calculating a control limit of the new working condition; on-line monitoring comprises the following steps: firstly, a weight selector is used for distributing weights for a low-rank matrix of a learned working condition according to monitoring data, and a monitoring dictionary is constructed in a self-adaptive manner; reconstructing the monitoring data by using the monitoring dictionary, solving a reconstruction error, and further judging the working condition of the monitoring data; and finally, calculating a fault detection statistic according to the reconstruction error and the control limit of the corresponding working condition, and judging whether the monitoring data is abnormal or not according to the statistic. According to the invention, accurate monitoring of a multi-working-condition process is realized.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process monitoring, and in particular relates to a multi-operating-condition process monitoring method and system based on zero-forgetting continuous dictionary learning. Background Art

[0002] As modern industry continues to expand, production processes are becoming increasingly complex, with high coupling between subsystems. In highly complex systems, a failure in any link can impact the operation of the entire system, leading to serious production accidents. Process monitoring technology can capture the current state of industrial processes in real time, ensuring their safe and stable operation. However, as industrial systems become increasingly complex, accurate physical and knowledge models are difficult to obtain, making many model- and knowledge-based approaches difficult to apply. Data-driven approaches, rather than relying on first principles and process knowledge, instead build models based on massive amounts of process data to achieve precise monitoring of industrial processes. With the rapid development of the Industrial Internet and smart sensors in recent years, the collection of process data has become increasingly convenient, and data-driven approaches have gradually become the mainstream of process monitoring.

[0003] Although data-driven monitoring methods have achieved certain success, they often require data to obey a single distribution characteristic. However, with changes in production materials, production loads, etc., actual industrial processes often operate under a variety of different working conditions. There are large differences in the data distribution of different working conditions, making it difficult for traditional data-driven methods to adapt to complex multi-condition process monitoring. In recent years, in order to solve the problem of multi-condition process monitoring, many advanced methods have been proposed. Among them, the multi-condition process monitoring method based on online update uses historical working condition data to train the initial model, and adaptively updates the model for new working conditions, so that the model can continuously learn the representation of new working conditions, thereby realizing multi-condition process monitoring. However, the online update method has the problem of "catastrophic forgetting", that is, when the model learns the representation of new working conditions, it often loses the ability to represent historical working conditions.

[0004] To address the problem of catastrophic forgetting, many continuous learning concepts have been applied to process monitoring. However, most current methods require sufficient similarity between different operating conditions to ensure that the model's shared parameter space can effectively represent them. In actual industrial processes, however, different operating conditions often differ significantly. Furthermore, as new operating conditions continue to emerge, the interference between knowledge from different conditions exacerbates model forgetting, making it difficult to adapt to long-term tasks with a continuously growing number of operating conditions. Summary of the Invention

[0005] The present invention provides a multi-operating process monitoring method and system based on zero-forgetting continuous dictionary learning, which enables the dictionary to memorize and integrate knowledge of new and old operating conditions, thereby realizing accurate monitoring of multi-operating processes.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] A multi-condition process monitoring method based on zero-forgetting continuous dictionary learning, comprising:

[0008] Offline modeling: First, dictionary learning is performed using the dataset of the initial operating condition. The resulting initial dictionary is decomposed into a low-rank matrix and weight selector for the initial operating condition. The reconstruction error of each training sample is further calculated to calculate the control limits of the initial operating condition. Then, using the dataset of the new operating condition, the low-rank matrix obtained from the old operating condition is incrementally updated to obtain the low-rank matrix and weight selector for the new operating condition. The reconstruction error of each training sample is further calculated to calculate the control limits of the new operating condition.

[0009] Online monitoring: First, a weight selector is used to assign weights to the low-rank matrix of the learned operating conditions based on the monitoring data, and a monitoring dictionary is adaptively constructed. Then, the monitoring data is reconstructed using the monitoring dictionary and the reconstruction error is solved. The operating condition of the monitoring data is determined based on the reconstruction error. Finally, the fault detection statistic is calculated based on the reconstruction error and control limits of the corresponding operating condition, and the size of the statistic is used to determine whether the monitoring data is abnormal.

[0010] Furthermore, the obtained initial dictionary is decomposed into a low-rank matrix and a weight selector of the initial working condition, and the initial dictionary is decomposed using the SVD method, and the front The largest singular values ​​construct a diagonal matrix and record it as the weight selector of the initial working condition , the corresponding left and right singular vectors and As the low-rank matrix of the initial working condition .

[0011] Furthermore, the reconstruction error of each training sample is calculated using the low-rank matrix and weight selector of the working condition, and then the control limit of the working condition is calculated, specifically:

[0012] First, use the weight selector to target the working condition Training dataset Each training sample , build the corresponding dictionary:

[0013]

[0014] in, Based on working conditions No. training samples The weight of the selection, Represents the corresponding constructed dictionary, For training data set The number of training samples;

[0015] Then, using the constructed dictionary For the corresponding training samples Perform reconstruction and calculate reconstruction error :

[0016]

[0017] Finally, according to the working conditions The reconstruction error of all training samples is calculated using the KDE method to calculate the working condition control limits.

[0018] Furthermore, the low-rank matrix obtained by learning the old working conditions is incrementally updated using the data set of the new working condition, specifically:

[0019] First, the low-rank matrix incremental update is formulated as the following optimization problem:

[0020]

[0021] in, represents the training data set for the new working condition, represents two incremental low-rank matrices trained for new working conditions, Two low-rank matrices representing the old working conditions, and Respectively represent the weight selectors for assigning weights to the low-rank matrices of the new and old working conditions, For the dataset Sparse coding matrix based on the new working condition dictionary, Represents the hyperparameter of the sparsity constraint term.

[0022] Then, the optimization problem of incremental update of the above low-rank matrix is ​​equivalent to the following form, and the alternating iterative method is used to solve the new working condition dictionary :

[0023]

[0024] Then the new working condition dictionary is analyzed by SVD method. Decompose to solve the following optimization problem, and obtain two incremental low-rank matrices of the new working condition :

[0025]

[0026]

[0027] in, Indicates the new working condition dictionary Remove the remainder of the incremental portion;

[0028] Finally, according to the incremental low-rank matrix of the new working condition , expand the low-rank matrix of the old working condition to obtain the low-rank matrix of the new working condition :

[0029] .

[0030] Furthermore, the method of using a weight selector to assign weights to the low-rank matrix of the learned working condition based on the monitoring data and adaptively constructing a monitoring dictionary includes:

[0031] First, construct the following optimization problem for the weight selector:

[0032] (01)

[0033] in, represents the weight selector, which is a diagonal matrix; Indicates monitoring data; It means sparse coding; A low-rank matrix representing the learned operating conditions;

[0034] Then, the optimization variable is solved by alternating iteration: weight selector and sparse coding :Fix first optimization ; Then fix And use OMP algorithm to solve Among them, fixed optimization The specific optimization process is:

[0035] Convert the F norm in formula (01) into the following trace form:

[0036] (02)

[0037] Based on formula (02) Taking the partial derivative, we get:

[0038] (03)

[0039] Let the intermediate matrix ,get:

[0040] (04)

[0041] Respectively and Perform orthogonal diagonal decomposition and let the intermediate matrix ,get:

[0042] (05)

[0043] in and The orthogonal diagonal decomposition results are and , and They are and The matrix composed of the corresponding singular value vectors, and They are and The diagonal matrix composed of the corresponding singular values; since and are all real symmetric matrices, so the matrix ; Representation matrix No. Rank Elements of the column, Intermediate Matrix No. Rank Elements of the column;

[0044] Use the following formula (06) to Fit the diagonal elements of :

[0045] (06)

[0046] in express the number of elements on the diagonal, express The diagonal elements;

[0047] Finally, the weight selector obtained by fitting is used The monitoring dictionary for each working condition is adaptively constructed as follows:

[0048]

[0049] in, For working conditions The monitoring dictionary, For working conditions The low-rank matrix, For working conditions The incremental low-rank matrix, is the number of working conditions that have been learned.

[0050] Furthermore, the use of the monitoring dictionary to reconstruct the monitoring data and solve the reconstruction error, and judging the working condition of the monitoring data according to the reconstruction error, includes:

[0051] Using the monitoring dictionary for each working condition, the corresponding reconstruction error is calculated for each monitoring data:

[0052]

[0053] in, For working conditions The monitoring dictionary, Indicates monitoring data, represents sparse coding, For monitoring data In monitoring dictionary The reconstruction error under is the number of working conditions that have been learned;

[0054] Starting from the initial working condition and following the working condition learning order, the monitoring data are The reconstruction error Compare with the preset threshold value, when the reconstruction error under a certain working condition is less than the preset threshold value for the first time, the working condition is determined as monitoring data working conditions.

[0055] Furthermore, the calculation of the fault detection statistic based on the reconstruction error and control limit of the corresponding working condition and judging whether the monitoring data is abnormal based on the size of the statistic include:

[0056] Set monitoring data The working condition is , the reconstruction error is , calculate the fault detection statistics as:

[0057]

[0058] in, For monitoring data Based on working conditions The fault detection statistics, For working conditions control limits;

[0059] Only when Only when the monitoring data Normal, otherwise the monitoring data abnormal.

[0060] A multi-operating process monitoring system based on zero-forgetting continuous dictionary learning includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor implements the method described above.

[0061] The present invention proposes a multi-operating process monitoring method and system based on zero-forgetting continuous dictionary learning. First, the intrinsic characteristics of various operating conditions are accurately captured through low-rank matrices, significantly reducing the redundant parameters of the over-complete dictionary; then, through an incremental learning framework based on low-rank matrix expansion, while maintaining the "zero forgetting" of historical operating condition representations, continuous learning of new operating condition representations is achieved; finally, through an adaptive weight selector, adaptive fusion of different operating condition knowledge is achieved, and accurate monitoring of multi-operating process is achieved on the basis of accurate identification of operating conditions. The method of the present invention integrates the incremental learning framework based on low-rank matrix expansion and the adaptive weight selector, so that the dictionary can memorize and integrate the knowledge of new and old operating conditions, thereby achieving accurate monitoring of multi-operating process. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a diagram of the overall architecture of the multi-condition process monitoring method described in an embodiment of the present invention.

[0063] Figure 2 It is a structural diagram of the roasting process of the zinc smelter according to an embodiment of the present invention.

[0064] Figure 3 : This is a comparison of the experimental results described in the embodiments of the present invention; among them, (a)~(c), (d)~(f), (g)~(i) are the monitoring results of all working condition test data of the trained dictionaries in the first, second and third stages respectively, using the traditional dictionary learning method (DL), the dictionary learning method based on modal matching and similarity preservation (JMSDL) and the method of the present invention (IRCDL); among them, 0~500, 500~1000, 1000~1500, and 1500~2000 are the test data of working conditions 1, 2, 3 and abnormal working conditions respectively.

[0065] Figure 4 This is the working condition recognition result of three normal working condition samples described in the embodiment of the present invention. DETAILED DESCRIPTION

[0066] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, provides a detailed implementation method and a specific operation process, and further explains the technical solution of the present invention.

[0067] The embodiment of the present invention provides a multi-operating process monitoring method based on zero-forgetting continuous dictionary learning, which includes two parts: offline modeling and online monitoring. In the offline modeling part, the rank-increasing continuous dictionary learning method is used to establish the process monitoring model; in the online monitoring part, the process monitoring model trained in the offline modeling stage is used to realize online monitoring of multi-operating processes. Overall architecture Figure 1 shown.

[0068] 1. Offline modeling.

[0069] The process monitoring model is established using the rank-increasing continuous dictionary learning method. First, the process data needs to be collected and labeled. The data is represented as: ,in Indicates working conditions The data contains The basic principle of dictionary learning is to build a dictionary To characterize the data , where each column element of the dictionary is called a dictionary atom. The optimization problem of dictionary learning is expressed as follows:

[0070] (1)

[0071] in Represents the reconstruction error of the dictionary. By minimizing the reconstruction error, the dictionary can effectively learn the representation of the data. Mathematically, it is usually used Norm replacement norm, thus converting (1) into an unconstrained problem:

[0072] (2)

[0073] Dictionary learning can effectively represent single-condition data. However, due to the large differences in the distribution of multi-condition process data and the complexity of the underlying features, traditional dictionary learning methods have difficulty effectively learning the representation of multi-condition data. In recent years, people have attempted to update dictionaries to continuously adapt to the representation of new conditions. However, online dictionary update methods suffer from the problem of "catastrophic forgetting", that is, the updated dictionary's ability to represent old conditions is significantly reduced, which seriously affects the overall monitoring performance. Although some regularization methods currently mitigate "catastrophic forgetting" by protecting important model parameters, as the number of conditions increases, the interference between knowledge will gradually increase, leading to a decline in model performance. To eliminate the interference between knowledge, an intuitive approach is to establish a dedicated dictionary for each condition. However, since each overcomplete dictionary usually contains redundant information, this method usually requires a large amount of model storage space and also lacks scalability.

[0074] To address these issues, this paper proposes a rank-increasing continuous dictionary learning method. This method cleverly establishes a dedicated low-rank matrix for each operating condition. This method has low storage overhead and effectively suppresses interference between different knowledge points, achieving "zero forgetting" continuous learning. The proposed method is based on two basic frameworks: a "dictionary learning method based on low-rank matrix weighting" and an "incremental learning framework based on low-rank matrix amplification." These two basic frameworks are described in detail below.

[0075] ① Dictionary learning method based on low-rank matrix weighting:

[0076] The present invention will characterize the Working Condition Dictionary Expressed as a low-rank matrix weighted form:

[0077] (3)

[0078] in Contains low-rank matrix , the superscript of the low-rank matrix indicates that it is based on the The data of each working condition is trained; Indicates the adaptation The weight selector of each working condition, its component Represented as a low-rank matrix The weight of the assignment.

[0079] The above method represents the dictionary as a collection of information about the core features of all learned working conditions. On the one hand, the representation of new working conditions can be continuously learned by expanding the low-rank matrix. Expand the low-rank matrix of each working condition , to learn Characterization information of each working condition, where the low-rank matrix Based on the Data of each working condition On the other hand, due to the update process, the low-rank matrix containing historical working condition information remains unchanged, so the historical working conditions can still be characterized by formula (3) Dictionary , enabling the dictionary to continuously learn the representation of new working conditions with “zero forgetting”. It is worth mentioning that the low-rank matrix discards redundant information, so it has fewer parameters than an over-complete dictionary and accurately captures the intrinsic characteristics of different working conditions with lower storage consumption. In addition, the parameter space of the low-rank matrix is ​​independent and will not be interfered with by historical working condition knowledge during the process of learning working condition representation. Therefore, the present invention has stronger scalability than traditional methods.

[0080] ② Incremental learning framework based on low-rank matrix augmentation:

[0081] Based on (2) and (3), the optimization problem of low-rank matrix initialization is expressed as follows:

[0082] (4)

[0083] in represents the training set of the initial working condition, represents the initial low-rank matrix, Denotes the weight assigned to the initial low-rank matrix. The equivalent form of the optimization problem in (4) is as follows:

[0084] (5)

[0085] This problem can be solved by the method of alternating iterative optimization. First, fix , use OMP algorithm to solve the current optimal ; Then, fix , use KSVD algorithm to solve the current optimal By iterating the above steps alternately until the convergence condition is reached, the initial dictionary can be obtained. .

[0086] After getting the initial dictionary After that, we need to further solve the following optimization problem to obtain the corresponding low-rank matrix:

[0087] (6)

[0088] In order to solve the above optimization problem, Perform SVD decomposition and select the front The largest singular value is used as the weight , the corresponding left and right singular vectors and As low-rank matrices The above steps complete the initialization process of the low-rank matrix. Selected manually, usually the initial working condition should be a larger Ensure accurate construction of the initial working condition to avoid error transmission; smaller To reduce memory overhead growth.

[0089] When learning the representation of a new working condition, it is necessary to incrementally update the learned historical low-rank matrix. The optimization problem of incremental updating of the low-rank matrix is ​​expressed as follows:

[0090] (7)

[0091] in represents the training data set for the new working condition, represents the incremental low-rank matrix trained for the new working condition, It represents the historical low-rank matrix, and Represent the weights assigned to the new and old low-rank matrices respectively. The equivalent form of the optimization problem (7) is as follows:

[0092] (8)

[0093] The above optimization problem can also be solved by the alternating iterative method. , which includes the characterization of the new working condition Then, in order to obtain the incremental low-rank matrix, the following optimization problem needs to be further solved:

[0094] (9)

[0095] (10)

[0096] in Represents the remainder of the dictionary minus the increment. In order to capture Key new working condition information in Perform SVD decomposition and select the front The largest singular value is , the corresponding left and right singular vectors are used as incremental low-rank matrices To compensate for the information To reduce the loss, a weight selector is used to assign the optimal weight to the historical low-rank matrix. The principle of the weight selector is introduced in detail in the "Online Monitoring" section.

[0097] Therefore, the offline modeling part of the present invention includes: first, using the data set of the initial working condition to perform dictionary learning, and decomposing the obtained initial dictionary into the low-rank matrix and weight selector of the initial working condition; then, using the data set of the new working condition, the low-rank matrix obtained by learning the old working condition is incrementally updated to obtain the low-rank matrix and weight selector of the new working condition.

[0098] Each time the low-rank matrix is ​​updated for a new operating condition, the corresponding "component control limits" must be calculated, while the control limits for historical operating conditions do not need to be calculated. Therefore, after learning the low-rank matrix and weight selector for each operating condition through offline modeling, the reconstruction error of each training sample under that operating condition is further calculated to calculate the control limits for that operating condition. This assigns unique "component control limits" to each operating condition, effectively improving the accuracy of anomaly identification.

[0099] In this embodiment, any The process of solving the control limits for each operating condition includes:

[0100] First, use the weight selector to target the working condition Training data Each sample Build a common Dictionaries:

[0101] (11)

[0102] in, and Indicates based on working conditions No. training samples The weights are chosen and the dictionary is constructed.

[0103] Then, all samples are reconstructed using the constructed dictionary and the reconstruction error is calculated:

[0104] (12)

[0105] Finally, the component control limits are calculated using the KDE method:

[0106] (13)

[0107] (14)

[0108] in, Indicates the bandwidth that can be automatically selected. and Represent the probability density function and Gaussian kernel function respectively. By selecting the appropriate confidence coefficient , according to formula (14), the working condition can be calculated Component control limits .

[0109] 2. Online monitoring.

[0110] First, the present invention does not directly use the fixed weights obtained in the offline stage, but uses a weight selector to assign weights to the low-rank matrix of the learned working conditions based on the monitoring data, and adaptively constructs the monitoring dictionary, thereby achieving better generalization capability.

[0111] The optimization problem that the weight selector needs to solve is as follows:

[0112] (15)

[0113] in Represents the weight selector, which is a diagonal matrix. Indicates monitoring data, Then it represents sparse coding. Equation (11) contains two optimization variables and can be solved by alternating iteration. First, fix optimization ; Next, fix , using the OMP algorithm to solve By iterating the above steps alternately until the stopping condition is reached, the optimal parameters can be obtained. The following will introduce The optimization process first converts the F norm in (15) into the trace form:

[0114] (16)

[0115] Based on the above formula Taking partial derivatives we can get:

[0116] (17)

[0117] make We can get:

[0118] (18)

[0119] Respectively and Performing orthogonal diagonal decomposition yields:

[0120] (19)

[0121] in and The orthogonal diagonal decomposition results are and , and They are and The matrix composed of the corresponding singular value vectors, and They are and The diagonal matrix composed of the corresponding singular values; since and are all real symmetric matrices, so the matrix . Representation matrix No. Rank Elements of the column, Intermediate Matrix No. Rank To ensure that The diagonal property of the matrix is ​​calculated using the following formula: Fit the diagonal elements of :

[0122] (20)

[0123] in express the number of elements on the matrix diagonal, express The diagonal It should be noted that if the dictionary has not learned the representation of the working conditions corresponding to the data, or the data is abnormal, the fitting result will have a large deviation, which will cause the constructed dictionary to have a large reconstruction error for the data.

[0124] Then, the monitoring data is reconstructed using the monitoring dictionary and the reconstruction error is solved, and the working condition of the monitoring data is judged based on the reconstruction error.

[0125] The weight selector realizes the adaptive fusion of different working condition knowledge. On this basis, by judging the contribution of each working condition knowledge, the working condition of the data can be accurately identified. Monitoring data When arriving, the specific process of working condition identification is as follows: First, the weight is calculated by solving formula (15) and the dictionary is constructed as follows:

[0126] (twenty one)

[0127] in, Indicates the total number of learned working conditions. Indicates the working condition Dictionary, which is used before The low-rank matrix of the working conditions Then, the reconstruction error is solved based on the constructed dictionary:

[0128] (twenty two)

[0129] when When, due to its lack of Partial information, so it cannot effectively characterize the working conditions , which leads to reconstruction error Very big. Time, dictionary Reconstruction error calculated for the data Compared to The data condition can be judged by judging the corresponding dictionary number when the reconstruction error is significantly reduced. Comparing with the reconstruction error, when the reconstruction error is less than the threshold for the first time, it can be judged that the working condition of the data is the corresponding reconstruction error number, that is:

[0130] (twenty three)

[0131] After completing the working condition identification, you can further determine whether the data is abnormal.

[0132] Finally, the fault detection statistic is calculated based on the reconstruction error and control limit of the corresponding working condition, and the abnormality of the monitoring data is judged by the size of the statistic. In order to facilitate the drawing of the monitoring diagram, the monitoring statistics are defined as follows in this embodiment:

[0133] (twenty four)

[0134] Only when The monitoring data is considered normal only when Corresponding working conditions "According to the working condition identification process, The working condition identification process is based on right The reconstruction error is calculated as, It is calculated during the offline modeling process, and the corresponding control limits are selected based on the working conditions of the online monitoring data.

[0135] In order to verify the effectiveness of the present invention in actual industrial processes, the following section takes the roasting process of a zinc smelter as an example. The roasting process converts insoluble zinc sulfide into zinc oxide soluble in weak acid by high-temperature oxidation of zinc concentrate in a fluidized bed roaster. The roasting process is a typical multi-condition process, including healthy conditions, high-efficiency conditions, over-decomposition conditions, etc. The structure of the roasting process is as follows: Figure 2 shown.

[0136] This experiment collected 25 variables from a real roasting process, including temperature, pressure, and flow rate. Three operating conditions were selected as normal conditions: Condition 1 for healthy conditions, Condition 2 for high-efficiency conditions, and Condition 3 for underoxidation conditions. Each condition included 1,000 training samples and 500 test samples. Following the test samples, 500 samples were added to represent abnormal bed deposition conditions, for a total of 3,000 training samples and 2,000 test samples.

[0137] First, in order to verify the superiority of the present invention in multi-condition process monitoring performance, a process monitoring experiment was conducted. Using the error detection rate (FDR), false alarm rate (FAR) and average monitoring accuracy as indicators, the present invention method, the traditional dictionary learning method (DL), the label consistent dictionary learning method (LCDL) and the dictionary learning method based on modal matching and similarity preservation (JMSDL) were compared. In the offline modeling stage, based on the four methods, the training samples of working conditions 1 to 3 were used to train the models in stages to obtain the corresponding models. Among them, DL, LCDL and JMSDL were obtained in turn according to the training stage. Three dictionary models, the present invention obtains Three sets of low-rank matrices; in the online monitoring phase, all methods are used to obtain the model in the last training phase to monitor the test data. The monitoring results are shown in Table 1.

[0138]

[0139] The experimental results show that the present invention has a higher FDR and the lowest FAR, and the average monitoring accuracy is far superior to other methods, which verifies the superior multi-condition process monitoring performance of the present invention.

[0140] Next, in order to further verify the continuous learning ability of the present invention, the model obtained by training the training sets of working conditions 1, 2, and 3 in sequence using DL and JMSDL is , and the model obtained by sequentially training the training sets based on working conditions 1, 2, and 3 of the present invention The test data were monitored and compared, and the experimental results are as follows: Figure 3 shown.

[0141] In addition, this example also calculates the average monitoring accuracy and average forgetting rate of the update process, as shown in Table 2:

[0142]

[0143] The experimental results show that compared to the other two methods, the forgetting rate of the proposed method remains at 0%. Therefore, it can effectively maintain the monitoring performance of historical working conditions while continuously learning the representation of new working condition data. Therefore, the proposed method has more advanced continuous learning capabilities and multi-working condition process monitoring performance.

[0144] Finally, the working condition recognition results of the present invention for three normal working condition samples are also recorded, such as Figure 4 shown.

[0145] It can be seen from the experimental results that the present invention can achieve accurate working condition identification results and further enhance the monitoring performance of multiple working condition processes.

[0146] The above embodiments are preferred embodiments of the present invention. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present invention, these changes or improvements should fall within the scope of protection required by the present invention.

Claims

1. A multi-process monitoring method based on zero-forgetting continuous dictionary learning, characterized in that: include: Offline modeling: First, dictionary learning is performed using the dataset of the initial operating condition. The resulting initial dictionary is decomposed into a low-rank matrix and a weight selector for the initial operating condition. The reconstruction error of each training sample is then calculated to calculate the control limits for the initial operating condition. Then, using the dataset of the new working condition, the low-rank matrix obtained by learning the old working condition is incrementally updated to obtain the low-rank matrix and weight selector of the new working condition. The reconstruction error of each training sample is further calculated to calculate the control limit of the new working condition. Online monitoring: First, a weight selector is used to assign weights to the low-rank matrix of the learned operating conditions based on the monitoring data, and a monitoring dictionary is adaptively constructed. Then, the monitoring data is reconstructed using the monitoring dictionary and the reconstruction error is solved. The operating condition of the monitoring data is determined based on the reconstruction error. Finally, the fault detection statistic is calculated based on the reconstruction error and control limits of the corresponding operating condition, and the size of the statistic is used to determine whether the monitoring data is abnormal.

2. The multi-operating process monitoring method based on zero-forgetting continuous dictionary learning according to claim 1 is characterized in that: The initial dictionary is decomposed into a low-rank matrix of the initial working condition and a weight selector, and the initial dictionary is decomposed using the SVD method, and the front The largest singular values ​​construct a diagonal matrix and record it as the weight selector of the initial working condition , the corresponding left and right singular vectors and As the low-rank matrix of the initial working condition .

3. The multi-operating process monitoring method based on zero-forgetting continuous dictionary learning according to claim 1 is characterized in that: The reconstruction error of each training sample is calculated using the low-rank matrix and weight selector of the working condition, and then the control limit of the working condition is calculated, which is specifically: First, use the weight selector to target the working condition Training dataset Each training sample , build the corresponding dictionary: ； in, Based on working conditions No. training samples The weight of the selection, Represents the corresponding constructed dictionary, For training data set The number of training samples; Then, using the constructed dictionary For the corresponding training samples Perform reconstruction and calculate reconstruction error : ； Finally, according to the working conditions The reconstruction error of all training samples is calculated using the KDE method to calculate the working condition control limits.

4. The multi-operating process monitoring method based on zero-forgetting continuous dictionary learning according to claim 1 is characterized in that: The low-rank matrix obtained by learning the old working conditions is incrementally updated using the new working condition data set, specifically: First, the low-rank matrix incremental update is formulated as the following optimization problem: ； in, represents the training data set for the new working condition, represents two incremental low-rank matrices trained for new working conditions, Two low-rank matrices representing the old working conditions, and Respectively represent the weight selectors for assigning weights to the low-rank matrices of the new and old working conditions, For the dataset Sparse coding matrix based on the new working condition dictionary, Hyperparameters representing the sparsity constraint; Then, the optimization problem of incremental update of the above low-rank matrix is ​​equivalent to the following form, and the alternating iterative method is used to solve the new working condition dictionary : ； Then the new working condition dictionary is analyzed by SVD method. Decompose to solve the following optimization problem, and obtain two incremental low-rank matrices of the new working condition : ； ； in, Indicates the new working condition dictionary Remove the remainder of the incremental portion; Finally, according to the incremental low-rank matrix of the new working condition , expand the low-rank matrix of the old working condition to obtain the low-rank matrix of the new working condition : 。 5. The multi-operating process monitoring method based on zero-forgetting continuous dictionary learning according to claim 1 is characterized in that: The method of using a weight selector to assign weights to a low-rank matrix of a learned working condition based on monitoring data and adaptively constructing a monitoring dictionary includes: First, construct the following optimization problem for the weight selector: （01） in, represents the weight selector, which is a diagonal matrix; Indicates monitoring data; It means sparse coding; A low-rank matrix representing the learned operating conditions; Then, the optimization variable is solved by alternating iteration: weight selector and sparse coding :Fix first optimization ; Then fix And use OMP algorithm to solve Among them, fixed optimization The specific optimization process is: Convert the F norm in formula (01) into the following trace form: (02) Based on formula (02) Taking the partial derivative, we get: (03) Let the intermediate matrix ,get: (04) Respectively and Perform orthogonal diagonal decomposition and let the intermediate matrix ,get: (05) in and The orthogonal diagonal decomposition results are and , and They are and The matrix composed of the corresponding singular value vectors, and They are and The diagonal matrix composed of the corresponding singular values; since and are all real symmetric matrices, so the matrix ; Representation matrix No. Rank Elements of the column, Represents the intermediate matrix No. Rank Elements of the column; Use the following formula (06) to Fit the diagonal elements of : (06) in express the number of elements on the diagonal, express The diagonal elements; Finally, the weight selector obtained by fitting is used The monitoring dictionary for each working condition is adaptively constructed as follows: ； in, For working conditions The monitoring dictionary, For working conditions The low-rank matrix, For working conditions The incremental low-rank matrix, is the number of working conditions that have been learned.

6. The multi-operating process monitoring method based on zero-forgetting continuous dictionary learning according to claim 1 is characterized in that: The method of reconstructing the monitoring data using the monitoring dictionary and solving the reconstruction error, and judging the working condition of the monitoring data according to the reconstruction error, includes: Using the monitoring dictionary for each working condition, the corresponding reconstruction error is calculated for each monitoring data: ； in, For working conditions The monitoring dictionary, Indicates monitoring data, represents sparse coding, For monitoring data In monitoring dictionary The reconstruction error under is the number of working conditions that have been learned; Starting from the initial working condition and following the working condition learning order, the monitoring data are The reconstruction error Compare with the preset threshold value, when the reconstruction error under a certain working condition is less than the preset threshold value for the first time, the working condition is determined as monitoring data working conditions.

7. The multi-operating process monitoring method based on zero-forgetting continuous dictionary learning according to claim 1 is characterized in that: The calculation of the fault detection statistic based on the reconstruction error and control limit of the corresponding working condition and judging whether the monitoring data is abnormal based on the size of the statistic include: Set monitoring data The working condition is , the reconstruction error is , calculate the fault detection statistics as: ； in, For monitoring data Based on working conditions The fault detection statistics, For working conditions control limits; Only when Only when the monitoring data Normal, otherwise the monitoring data abnormal.

8. A multi-operating process monitoring system based on zero-forgetting continuous dictionary learning, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that: When the computer program is executed by the processor, the processor is caused to implement the method according to any one of claims 1 to 7.

Images