Industrial process depth monitoring method based on multi-sampling-rate data distillation

Abstract

The invention discloses an industrial process depth monitoring method based on multi-sampling-rate data distillation, and provides an industrial process depth monitoring method based on multi-sampling-rate data distillation under the framework of a depth residual error principal component analysis model. Under the condition that the number of training samples is not increased, multiple pieces of data sample information are compressed into a concentrated new data point through data distillation, and information extraction of multi-sampling-rate data is achieved. Therefore, compared with a traditional monitoring method, the method disclosed by the invention not only can effectively utilize all data information, but also ingeniously solves the problem of multiple sampling rates of industrial data by utilizing a data distillation technology. Through an actual industrial case, the feasibility and effectiveness of the method are well evaluated and confirmed.

Description

Technical Field

[0001] The present invention belongs to the field of safety monitoring and fault diagnosis of industrial production processes, and in particular relates to an industrial process in-depth monitoring method based on multi-sampling rate data distillation. Background Art

[0002] Data-driven process monitoring plays an important role in current industrial production and is the foundation of process control, operation optimization and other related applications. However, in the current industrial process monitoring methods, the problem of multiple sampling rates of data is often ignored by researchers, but it is a widely existing and very important data characteristic, which seriously restricts the performance of process monitoring models in terms of false alarm rate, monitoring sensitivity, time cost and human resources. With the accumulation of industrial process data and the introduction of various artificial intelligence methods into process monitoring applications, especially deep learning models, how to use data information more efficiently has become an important hot issue. Based on the widespread problem of multiple sampling rates of data in industrial processes, how to use all data information more effectively for process monitoring without significantly increasing the computational cost of the monitoring model is a very meaningful research topic, which has very important practical application value for the green and low-cost deployment of data-driven models in industrial systems and the stable and reliable operation of process monitoring models. Summary of the invention

[0003] The purpose of the present invention is to provide an industrial process in-depth monitoring method based on multi-sampling rate data distillation in order to address the problem that the performance of existing industrial process monitoring models is limited by data characteristics.

[0004] The objective of the present invention is achieved through the following technical solutions:

[0005] A method for in-depth monitoring of industrial processes based on multi-sampling rate data distillation comprises the following steps:

[0006] (1) Using the distributed control system deployed on a large scale in the industrial production process, the data of various measured variables under normal operating conditions are collected and uniformly recorded as: Where s = 1, 2, ..., S is the category of data with different sampling rates. Assuming that there are S categories of different sampling rates in the process variable, the data set is divided into S categories, n s ,s=1,2,...,S is the number of training samples for each type of data, m s ,s=1,2,...,S is the number of measurement variables for each type of data.

[0007] (2) Based on the sampling rates of different variable categories, the lowest common multiple is determined, and the data corresponding to the high sampling rate variables are distilled according to the corresponding proportions, and finally a concentrated modeling sample data set is obtained, which is recorded as: Among them, n c is the number of samples, and m is the number of all variables in the process.

[0008] (3) After necessary preprocessing of the distilled data sample set, it is normalized, that is, the mean and standard deviation of each variable are removed so that the mean of each variable is 0 and the variance is 1, and different variables are unified to the same dimensional scale.

[0009] (4) For a concentrated modeling sample set First, an initial principal component analysis model is established to extract the principal component information and calculate the residual data matrix of the model to form the first layer structure of the deep model.

[0010] (5) Based on the residual data matrix of the previous principal component analysis model, the second layer structure of the deep model is established, the principal component feature matrix is ​​extracted, and the residual data matrix corresponding to this layer is calculated. Similarly, we establish a deep principal component analysis model containing L layers in total, and store all the parameters obtained in the modeling process in the model library for future use.

[0011] (6) Finally, the principal component information extracted from different hidden layers in the deep principal component analysis model is integrated to obtain a comprehensive principal component feature, which is then subjected to principal component analysis again. On this basis, the process monitoring statistics are constructed and the corresponding confidence limits are determined.

[0012] (7) Collect new industrial process operation data and normalize them. Use the deep principal component analysis model to calculate the principal component feature information and residual information corresponding to each layer of the model, and obtain the value of the corresponding statistic.

[0013] (8) The principal component information extracted from the hidden layers of different depth models is integrated into a comprehensive principal component feature, which is input into the final principal component analysis model to calculate the monitoring statistic value and determine the operating status of the current process.

[0014] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the industrial process in-depth monitoring method based on multi-sampling rate data distillation is implemented.

[0015] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the industrial process in-depth monitoring method based on multi-sampling rate data distillation.

[0016] The beneficial effect of the present invention is that the present invention establishes a deep principal component analysis model for industrial process monitoring based on the distillation of multi-sampling rate data, thereby overcoming the data utilization bottleneck of the traditional data-driven monitoring model. Compared with the current commonly used methods, the present invention compresses multiple data sample information into a concentrated new data point through the data distillation technology of information accumulation without increasing the number of model training samples, realizes information extraction of multi-sampling rate data, greatly improves the utilization efficiency of data, and cleverly solves the multi-sampling rate problem of industrial data, improving the actual application effect of process monitoring from multiple aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of chemical production process;

[0018] Figure 2 is the monitoring result of the test data sample set in the method of the present invention;

[0019] Figure 3 It is the monitoring result of the test data sample set without data distillation method. DETAILED DESCRIPTION

[0020] The present invention aims at the problem of multi-sampling rate data characteristics that are widely present in industrial process monitoring. First, the data distillation technology of information accumulation is used to distill the data samples corresponding to the variables with different sampling rates, which greatly increases the utilization efficiency of data information without increasing the number of training samples of the data-driven model. Then, on the basis of normalizing the monitoring variables, a deep principal component analysis model including L hidden layers is established, and corresponding statistics are constructed in the principal component space and the residual space to monitor the process. Based on the constructed deep monitoring model, the new industrial process data is monitored in real time. First, the model information is used to normalize it, and then the deep monitoring model is used to calculate the statistics corresponding to the new data, so as to judge the operation status of the current process in real time.

[0021] The main steps of the technical solution adopted by the present invention are as follows: A method for in-depth monitoring of industrial processes based on multi-sampling rate data distillation, the method comprising the following steps:

[0022] The first step is to use the distributed control system deployed on a large scale in the industrial production process to collect data on various measurement variables under normal operating conditions, which are uniformly recorded as: Where s = 1, 2, ..., S is the category of data with different sampling rates. Assuming that there are S categories of different sampling rates in the process variable, the data set is divided into S categories, n s ,s=1,2,...,S is the number of training samples for each type of data, m s,s=1,2,...,S is the number of measurement variables for each type of data,

[0023] The second step is to determine the lowest common multiple based on the sampling rates of different variable categories, and distill the data corresponding to the high sampling rate variables according to the corresponding proportions, and finally obtain a concentrated modeling sample data set, which is recorded as: Among them, n c is the number of samples, and m is the number of all variables in the process. The data distillation process based on information accumulation is as follows:

[0024]

[0025] Among them, C s ,s=1,2,...,S is the number of samples of each variable used for data distillation,

[0026] The third step is to normalize the data sample set after necessary preprocessing, that is, to remove the mean and standard deviation of each variable, so that the mean of each variable is 0 and the variance is 1, and to unify different variables to the same dimension scale.

[0027] The fourth step is to focus on the concentrated modeling sample set First, an initial principal component analysis model is established to extract the principal component information and calculate the residual data matrix of the model to form the first layer structure of the deep model, namely

[0028]

[0029] in, is the principal component feature matrix corresponding to the principal component analysis model of the first layer of the deep model, is the load matrix of the model, is the corresponding residual data matrix,

[0030] The fifth step is to build the second layer structure of the deep model based on the residual data matrix of the previous layer principal component analysis model, extract the principal component feature matrix, and calculate the residual data matrix corresponding to this layer. Similarly, we build a deep principal component analysis model with L layers in total, and store all the parameters obtained in the modeling process in the model library for future use. Among them, the last layer structure of the deep model is as follows:

[0031]

[0032] Finally, in the sixth step, the principal component information extracted from different hidden layers in the deep principal component analysis model is integrated to obtain a comprehensive principal component feature, and principal component analysis is performed on it again. On this basis, process monitoring statistics are constructed and corresponding confidence limits are determined. The principal component information extracted from different hidden layers is integrated to obtain the following comprehensive principal component feature information matrix, and principal component analysis is performed on it:

[0033]

[0034] in, l=1,2,...L is the principal component matrix extracted at different levels. In summary, based on the hidden layer models of the deep model and the above-mentioned comprehensive principal component analysis model, we establish the following monitoring statistics and monitoring statistical confidence limits:

[0035]

[0036] Where b = 1, 2, ..., L, or uf, k b is the corresponding number of pivot elements, is the F distribution, and α is the confidence level. In the process of constructing the confidence limit of the SPE statistic, g b =var / (2mean) and h b =2(mean) 2 / var, where mean and var are the mean and variance of the SPE values ​​corresponding to each data in the training set.

[0037] The seventh step is to collect new industrial process operation data and normalize it. Using the deep principal component analysis model, calculate the principal component feature information and residual information corresponding to each layer of the model, and obtain the corresponding statistical value. new , the extraction of principal component information and residual information of each hidden layer of the deep model, and the calculation of the corresponding monitoring statistics are as follows:

[0038]

[0039] in, l=1,2,...L is the principal element vector extracted from the new data at different levels, l=1,2,...L is the residual vector obtained at the corresponding level, l=1,2,...L are T calculated at different levels 2 Statistical value, l=1,2,...L are the SPE statistics calculated at different levels.

[0040] In the eighth step, the principal component information extracted from the hidden layers of different depth models is integrated into a comprehensive principal component feature, which is input into the final principal component analysis model to calculate the monitoring statistic value and determine the operating status of the current process.

[0041] Among them, the new data corresponds to x new The final process monitoring statistics are calculated as follows:

[0042]

[0043] By comparing the values ​​of the two monitoring statistics with their respective statistical confidence limits, we can easily determine whether there is any abnormality in the operating status of the process.

[0044] Below we combine a specific chemical production process case to illustrate the effectiveness of the method of the present invention. The flow chart of the chemical production process is as follows Figure 1 As shown, it consists of several different operating units. In order to monitor the process, all 52 variables in the process are selected for monitoring, including 33 conventional high sampling rate variables such as temperature, pressure and flow, and 19 low sampling rate variables related to material composition. In order to establish a deep process monitoring model, we collected a total of 800 data points in the process. However, due to the multi-sampling rate problem of variables, the number of effective samples that can be used by conventional methods is only 80, which is 10% of the original samples. After using data distillation technology, the other 90% of the data sample information can be effectively utilized. Next, the implementation steps of the present invention are explained in detail in conjunction with this specific process:

[0045] 1. Collect chemical production process data, distill multi-sampling rate data, and establish a deep monitoring model

[0046] Based on the multi-sampling rate data distillation technology, the 800 collected normal operation status data samples are processed to compress the original data sample information into 80 final concentrated samples. Then, according to the specific implementation steps given by the method of the present invention, a deep model including 3 hidden layers is established to monitor the process.

[0047] 2. Integration of deep model principal component feature information and construction of comprehensive process monitoring statistics

[0048] The principal component features of each hidden layer of the deep principal component analysis model are integrated to establish the final principal component analysis model, determine the monitoring statistics corresponding to the principal component space and residual space, and determine their statistical confidence limits.

[0049] 3. Obtain the current monitoring data of the chemical production process and conduct online monitoring on the basis of normalization.

[0050] In order to test the effectiveness of the new method, the normal data samples and abnormal data samples collected in the process were tested respectively, including 96 sample points, of which the first 16 were normal data and the following 80 were abnormal data. On the basis of data normalization, the statistical values ​​corresponding to each new data sample were calculated to obtain the results of online monitoring. Figure 2 The monitoring results of the method of the present invention on the test data sample set are given. It can be seen that the new method has a good monitoring effect on both normal and abnormal data. In contrast, if we do not perform data distillation, the deep monitoring model will not only give a high false positive rate for normal data, but also a very high false negative rate for abnormal data. Figure 3 Therefore, from the monitoring results of this case, the method of the present invention is very effective and feasible.

[0051] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for in-depth monitoring of industrial processes based on multi-sampling rate data distillation, characterized in that: The following steps are involved: (1) Using the distributed control system deployed on a large scale in the industrial production process, the data of various measured variables under normal operating conditions are collected and uniformly recorded as: Where s = 1, 2, ..., S is the category of data with different sampling rates. Assuming that there are S categories of different sampling rates in the process variable, the data set is divided into S categories, n s ,s=1,2,...,S is the number of training samples for each type of data, m s ,s=1,2,...,S is the number of measurement variables for each type of data, (2) Based on the sampling rates of different variable categories, the lowest common multiple is determined, and the data corresponding to the high sampling rate variables are distilled according to the corresponding proportions, and finally a concentrated modeling sample data set is obtained, which is recorded as: Among them, n c is the number of samples, m is the number of all variables in the process, (3) After necessary preprocessing of the distilled data sample set, it is normalized, that is, the mean and standard deviation of each variable are removed so that the mean of each variable is 0 and the variance is 1, and different variables are unified to the same dimension scale. (4) For a concentrated modeling sample set First, an initial principal component analysis model is established to extract principal component information and calculate the residual data matrix of the model to form the first layer structure of the deep model. (5) Based on the residual data matrix of the previous principal component analysis model, the second layer structure of the deep model is established, the principal component feature matrix is ​​extracted, and the residual data matrix corresponding to this layer is calculated. Similarly, a deep principal component analysis model containing L layers is established, and all parameters obtained in the modeling process are stored in the model library for future use. (6) Finally, the principal component information extracted from different hidden layers in the deep principal component analysis model is integrated to obtain a comprehensive principal component feature, which is then subjected to principal component analysis again. On this basis, process monitoring statistics are constructed and the corresponding confidence limits are determined. (7) Collect new industrial process operation data and normalize them. Use the deep principal component analysis model to calculate the principal component feature information and residual information corresponding to each layer of the model to obtain the value of the corresponding statistic. (8) Integrate the principal component information extracted from the hidden layers of different deep models into a comprehensive principal component feature, input it into the final principal component analysis model, calculate the monitoring statistic value, and judge the operating status of the current process.

2. The industrial process in-depth monitoring method based on multi-sampling rate data distillation according to claim 1 is characterized in that: The step (2) is specifically as follows: based on the sampling rates of different variable categories, the lowest common multiple is determined, and the data corresponding to the high sampling rate variables are distilled according to the corresponding proportions, and finally a concentrated modeling sample data set is obtained, which is recorded as: Among them, n c is the number of samples, m is the number of all variables in the process, where the data distillation process based on information accumulation is as follows: Among them, C s ,s=1,2,...,S is the number of samples of each variable used for data distillation, m s is the number of variables of each category.

3. The industrial process in-depth monitoring method based on multi-sampling rate data distillation according to claim 1 is characterized in that: The step (4) is specifically as follows: for the concentrated modeling sample set First, an initial principal component analysis model is established to extract the principal component information and calculate the residual data matrix of the model to form the first layer structure of the deep model, namely in, is the principal component feature matrix corresponding to the principal component analysis model of the first layer of the deep model, is the load matrix of the model, is the corresponding residual data matrix.

4. The industrial process in-depth monitoring method based on multi-sampling rate data distillation according to claim 1 is characterized in that: The step (5) is specifically as follows: based on the residual data matrix of the previous principal component analysis model, the second layer structure of the deep model is established, the principal component feature matrix is ​​extracted, and the residual data matrix corresponding to the layer is calculated. By analogy, a deep principal component analysis model containing L layers is established in total, and all parameters obtained in the modeling process are stored in the model library for standby use. The last layer structure of the deep model is as follows:

5. The method for in-depth monitoring of industrial processes based on multi-sampling rate data distillation according to claim 1, characterized in that: The step (6) is specifically as follows: Finally, the principal component information extracted from different hidden layers in the deep principal component analysis model is integrated to obtain a comprehensive principal component feature, and principal component analysis is performed again. On this basis, process monitoring statistics are constructed, and corresponding confidence limits are determined. The principal component information extracted from different hidden layers is integrated to obtain the following comprehensive principal component feature information matrix, and principal component analysis is performed on it: in, The principal component matrices extracted at different levels are summarized as follows: Based on the hidden layer models of the deep model and the above-mentioned comprehensive principal component analysis model, the following monitoring statistics and monitoring statistical confidence limits are established respectively: Where b = 1, 2, ..., L, or uf, k b is the corresponding number of pivot elements, is the F distribution, α is the confidence level, and in the process of constructing the confidence limit of the SPE statistic, g b =var / (2mean) and h b =2(mean) 2 / var, where mean and var are the mean and variance of the SPE values ​​corresponding to each data in the training set.

6. The industrial process in-depth monitoring method based on multi-sampling rate data distillation according to claim 1, characterized in that: The step (7) is specifically as follows: collect new industrial process operation data, normalize it, use the deep principal component analysis model to calculate the principal component feature information and residual information corresponding to each layer of the model, obtain the value of the corresponding statistic, and new , the extraction of principal component information and residual information of each hidden layer of the deep model, and the calculation of the corresponding monitoring statistics are as follows: in, is the principal element vector extracted from the new data at different levels, is the residual vector obtained at the corresponding level, T calculated at different levels 2 Statistical value, SPE statistic values ​​calculated for different levels.

7. The method for in-depth monitoring of industrial processes based on multi-sampling rate data distillation according to claim 1, characterized in that: The step (8) is specifically as follows: integrating the principal component information extracted from the hidden layers of different depth models into a comprehensive principal component feature, inputting it into the final principal component analysis model, calculating the monitoring statistic value, and judging the operating status of the current process, wherein the new data corresponds to x new The final process monitoring statistics are calculated as follows: By comparing the values ​​of the two monitoring statistics with their respective statistical confidence limits, it is very convenient to determine whether there is any abnormality in the operating status of the process.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the industrial process in-depth monitoring method based on multi-sampling rate data distillation as described in any one of claims 1 to 7 above is implemented.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, an industrial process in-depth monitoring method based on multi-sampling rate data distillation as described in any one of claims 1 to 7 is implemented.

Images