A method for complex industrial process product quality detection integrating prior knowledge
By adopting a method of fusion of prior knowledge in complex industrial processes, combined with multi-scale feature extraction and feature fusion, the problem that traditional methods are difficult to characterize complex relationships between multiple processes is solved, and the accuracy and efficiency of product quality inspection is improved.
Patent Information
- Application Number
- CN202411000499.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-07-25
AI Technical Summary
In complex industrial processes, traditional data-driven modeling methods are difficult to accurately characterize complex relationships between multiple processes, resulting in the inability to obtain better feature representations, which affects the performance of product quality detection models.
Using a method of fusion prior knowledge, a feature extraction network and feature fusion module are combined with multi-scale feature extraction and prior knowledge fusion to build a product quality detection model for complex industrial processes, which is used for feature extraction within processes and feature fusion between multiple processes.
It improves the accuracy and efficiency of product quality inspection, and can more comprehensively extract the characteristic representation within and between multiple processes, achieving effective quality monitoring of complex industrial processes.
Smart Images

Figure CN118521013B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of complex industrial process modeling and product quality detection, and in particular, relates to a complex industrial process product quality detection method integrating prior knowledge. Background Art
[0002] Complex industrial processes are an important pillar of economic development, covering production areas such as chemical industry, oil refining, metallurgy, mineral processing, and steel production. These processes usually consist of multiple continuous production processes. The raw materials are gradually processed to obtain the final product, and the product quality is closely related to each process. Therefore, product quality is an important indicator reflecting the production status. However, the current detection of product quality usually relies on sampling and manual inspection, which needs to be carried out after production is completed, which leads to high costs and a certain degree of lag, limiting the real-time monitoring and quality control of the production process.
[0003] In recent years, with the development of informatization, digitization and artificial intelligence technology of complex industrial processes, data-driven modeling methods have been widely used in the field of product quality inspection. This method predicts product quality by establishing a relationship model between production process data and product quality. However, complex industrial processes have the production characteristics of multiple processes, and the processes influence each other and have complex relationships. Traditional data-driven modeling methods often find it difficult to accurately characterize this complex relationship, resulting in the inability to obtain better feature representation, which in turn affects the performance of the product quality inspection model.
[0004] Since complex industrial production processes involve the transformation of raw materials into products, raw materials undergo different processing forms in different processes before obtaining the final product, resulting in differences in process data within each process, and the need to extract the features of each process separately. A large number of sensor monitoring devices are used for real-time monitoring and data collection, accumulating a large amount of time series data. These time series data usually have multivariate and multi-scale characteristics. Traditional feature extraction methods often only consider features on a single time scale and ignore the correlation between multiple variables. In addition, traditional modeling methods usually only extract features on a single scale and cannot accurately capture the changes in data at different scales, making it difficult to extract more complete data features.
[0005] Invention patent CN112560337A discloses "intelligent modeling method, device, equipment and storage medium for digital twin system of complex industrial process", which combines the system identification method based on mechanism model with the deep learning method based on big data, and adopts the end-edge-cloud collaboration mode to establish an intelligent model of digital twin system of complex industrial process. Patent publication number CN108776831A discloses "a complex industrial process data modeling method based on dynamic convolutional neural network", which establishes a data prediction model of complex industrial process indicators based on dynamic convolutional neural network, and effectively considers the correlation of sensitive variables. However, invention patent CN112560337A does not consider the problem of time series data feature extraction under complex industrial processes. In the modeling process, invention patent CN112560337A and patent publication number CN108776831A did not consider the multi-process problem faced by process data modeling in complex industrial process scenarios. Summary of the invention
[0006] In view of the shortcomings of the existing technology, the present invention provides a complex industrial process product quality detection method that integrates prior knowledge to solve the problem of feature fusion between multiple processes and feature extraction of time series data within a process faced by complex industrial process modeling, and is used for product quality detection.
[0007] The technical solution of the present invention is:
[0008] A complex industrial process product quality detection method integrating prior knowledge comprises the following steps:
[0009] S1. Collect relevant data on product production in complex industrial processes as samples.
[0010] The relevant data includes the production process data of each step collected by each detection sensor during the production process and the product quality label marked after the product is produced.
[0011] S2. Preprocess the collected samples and use the preprocessed samples to construct a training data set; the preprocessing includes data cleaning, data alignment and normalization.
[0012] S2.1. Use the data cleaning algorithm to clean the collected samples, remove the samples containing sampling anomalies and non-numeric values in the collected production process data, and obtain the cleaned samples.
[0013] S2.2. Perform data alignment and normalization on the cleaned samples and construct a training dataset.
[0014] Specifically, firstly, the production process data in the cleaned samples are aligned so that the production process data of each sample have the same length. Then the aligned production process data is normalized to finally obtain the production process data corresponding to each process after preprocessing. ,in is the number of the process and , is the number of corresponding processes in the production process; therefore, each sample is represented by , is the sample number, and finally the training data set is obtained ,in, is the number of samples, For sample Corresponding product quality label.
[0015] Preferably, the data alignment in S2.2 adopts a linear interpolation algorithm.
[0016] S3. Construct a complex industrial process product quality detection model that integrates prior knowledge.
[0017] The complex industrial process product quality detection model integrating prior knowledge includes several feature extraction networks, a feature fusion module and a quality detection module.
[0018] The feature extraction network is used to extract features from the production process data in each process to obtain feature representations corresponding to the production process data of each process. The number of feature extraction networks is equal to the number of corresponding processes in the production process, and one feature extraction network corresponds to one process.
[0019] The feature fusion module is used to splice the feature representations corresponding to the input production process data of each process to obtain a spliced feature representation, then take the mutual influence between the processes as prior knowledge, and use the prior knowledge to fuse the spliced feature representation, and finally obtain the feature representation of the sample.
[0020] The prior knowledge is represented by an adjacency matrix, where A=\left [ {{a}_{pq}} \right ]_{n\times n} Indicates that during production The influence relationship between processes, among which, Indicates Is the first process The value of 0 indicates no impact and 1 indicates impact.
[0021] The quality detection module is a two-layer fully connected network, which is used to classify the feature representation of the input sample and output the predicted product quality.
[0022] Furthermore, each feature extraction network includes a number of multi-scale feature extraction modules connected in sequence, and each multi-scale feature extraction module consists of two branches. The first branch first extracts the feature representation corresponding to the input production process data in time series through a one-dimensional convolutional network, and then extracts the feature in the variable dimension through a one-dimensional convolutional network with the same size to obtain the first branch feature. The second branch has the same structure as the first branch, and the convolution kernel size of the convolutional network is different. The input is the feature representation corresponding to the production process data, and the second branch feature is output. Then, the first branch feature and the second branch feature are added and fused as the output feature of the multi-scale feature extraction module and input to the next cascaded multi-scale feature extraction module, until the last multi-scale feature extraction module outputs the feature representation of the sample.
[0023] S4. Use the training data set to train a complex industrial process product quality detection model that integrates prior knowledge to obtain a trained complex industrial process product quality detection model that integrates prior knowledge.
[0024] S4.1. Input the production process data of each process in the samples in the training data set into the corresponding feature extraction network respectively to obtain the feature representation corresponding to the production process data of each process.
[0025] The feature extraction process of the feature extraction network is as follows:
[0026]
[0027] in, For the The characteristics of each process are represented by Indicates The feature extraction network corresponding to each process.
[0028] The multi-scale feature extraction module in the feature extraction network uses convolution kernels of various sizes to extract features in parallel. Specifically, , The convolution kernel and Sliding extracts features on different time scales, using , The convolution kernel and Extract features at different variable scales, and is the convolution kernel size.
[0029] The multi-scale feature extraction module is as follows:
[0030]
[0031] in, represents the input features of the feature extraction module, represents the output features, is the convolution kernel of the first convolutional network in the first branch, is the convolution kernel of the second convolutional network in the first branch, is the convolution kernel of the first convolutional network in the second branch, is the convolution kernel of the second convolutional network in the second branch.
[0032] S4.2. Analyze the production process in complex industrial processes and obtain the mutual influence relationship between the working steps as prior knowledge.
[0033] S4.3. Use the acquired prior knowledge to perform feature fusion on the feature representation corresponding to the extracted production process data of each process, and finally obtain the feature representation of the sample, and use the quality inspection module to predict the product quality to obtain the predicted product quality.
[0034] Specifically, the feature fusion based on prior knowledge in S4.3 is shown in the following formula:
[0035]
[0036] in, represents the predicted value of product quality by the complex industrial process product quality detection model integrating prior knowledge, Indicates the splicing operation of the feature representation of each process. Represents the quality inspection module. represents the adjacency matrix, Indicates Characteristic representation of a process.
[0037] S4.4. The training is terminated after reaching the set training rounds or the loss function converges, and a trained complex industrial process product quality detection model integrating prior knowledge is obtained.
[0038] Specifically, the loss function in S4.4 is a cross entropy loss function, using cross entropy loss Calculate the loss between the predicted value of product quality and the actual value of product quality as shown in the following formula:
[0039]
[0040] in, To indicate the quantity of product quality types, The number of the product quality category; is the true value of product quality, The predicted value of product quality.
[0041] S5. Deploy the trained complex industrial process product quality detection model that integrates prior knowledge and use it for product quality detection in complex industries to obtain product quality.
[0042] Compared with the prior art, the beneficial effects of the present invention are:
[0043] The present invention provides a complex industrial process product quality detection method that integrates prior knowledge, which is used to solve the problem of intra-process data feature extraction and feature fusion between multiple processes in complex industrial production processes. Aiming at the problem of data feature extraction collected by sensors in each process, the present invention designs a multi-scale feature extraction network using convolution kernels of different sizes, which supports the extraction of features at multiple scales, thereby obtaining a more comprehensive feature representation within the process; Aiming at the multi-process problem faced in complex production processes, the present invention designs a multi-process feature fusion method using prior knowledge, which fuses the extracted features of each process to obtain a more accurate feature representation. The present invention provides an effective product quality detection method for complex industrial processes and improves production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 The figure is a flow chart of a complex industrial process product quality detection method integrating prior knowledge according to an embodiment of the present invention.
[0045] Figure 2 It is a structural block diagram of a complex industrial process product quality detection model integrating prior knowledge according to an embodiment of the present invention. DETAILED DESCRIPTION
[0046] The present invention is described in detail below with reference to the accompanying drawings and embodiments.
[0047] In order to solve the problem of extracting features from data collected by sensors in each process, the present invention designs a multi-scale feature extraction network using convolution kernels of different sizes, which supports the extraction of features at multiple scales, thereby obtaining a more comprehensive feature representation within the process. In order to solve the problem of multi-process problems faced in complex production processes, the present invention designs a multi-process feature fusion method using prior knowledge, which fuses the extracted features of each process to obtain a more accurate feature representation.
[0048] The present invention provides a complex industrial process product quality detection method that integrates prior knowledge, and uses the collected process data to train the model to achieve product quality detection. In this embodiment, it is used to detect the plate quality of thick plate products during the production process, such as Figure 1 As shown, the following steps are included:
[0049] S1. Using the monitoring sensors deployed in the production process of thick plate products, the production process data of each process in the production process of thick plate products are collected as samples, and the thick plate shape quality label information is obtained at the same time;
[0050] In this embodiment, in step S1, the collected production process data includes the monitoring data of each sensor in the heating process, rolling process and cooling process in the thick plate production process. Specifically, the collected data includes the slab temperature in the heating process in the heating process; the rolling force and bending roll force in the rolling process; the temperature of the steel plate at various positions in the cooling process, such as the final rolling temperature, the start cooling temperature and the final cooling temperature. The thick plate shape quality label information is obtained after the main production process is completed through the plate shape detection instrument combined with manual marking detection.
[0051] S2. Preprocess the collected samples and use the preprocessed samples to construct a training data set; the preprocessing includes data cleaning, data alignment and normalization.
[0052] S2.1. Use the data cleaning algorithm to clean the collected samples, remove samples containing sampling anomalies and non-number (NaN) values in the collected production process data, and obtain cleaned samples.
[0053] S2.2. Perform data alignment and normalization on the cleaned samples and construct a training dataset.
[0054] First, the data is aligned and linear interpolation is used to make the data have the same length. In this example, the length of the aligned samples is =36, and then normalize the interpolated data to finally obtain the production process data corresponding to each process. ,in is the number of the process and , is the number of corresponding processes in the production process; in this example, the main production processes involved are heating, rolling and cooling, and the corresponding , so each sample is represented as , is the sample number, and finally the training data set is obtained ,in, is the number of samples, For sample The corresponding product quality label. Preferably, the data alignment in S2.2 adopts a linear interpolation algorithm.
[0055] S3. Construct a complex industrial process product quality detection model that integrates prior knowledge.
[0056] like Figure 2As shown, the complex industrial process product quality detection model integrating prior knowledge includes several feature extraction networks, a feature fusion module and a quality detection module.
[0057] The feature extraction network is used to extract features from the production process data in each process to obtain feature representations corresponding to the production process data of each process. The number of the feature extraction networks is equal to the number of corresponding processes in the production process, and one feature extraction network corresponds to one process.
[0058] The feature fusion module is used to splice the feature representations corresponding to the input production process data of each process to obtain a spliced feature representation, then take the mutual influence between the processes as prior knowledge, and use the prior knowledge to fuse the spliced feature representation, and finally obtain the feature representation of the sample.
[0059] The prior knowledge is represented by an adjacency matrix, where A=\left [ {{a}_{pq}} \right ]_{n\times n} Indicates that during production The influence relationship between processes, among which, Indicates Is the first process The value of 0 indicates no impact and 1 indicates impact.
[0060] The quality detection module is a two-layer fully connected network, which is used to classify the feature representation of the input sample and output the predicted product quality.
[0061] Furthermore, each feature extraction network includes a number of multi-scale feature extraction modules connected in sequence, and each multi-scale feature extraction module consists of two branches. The first branch first extracts the feature representation corresponding to the input production process data in time series through a one-dimensional convolutional network, and then extracts the feature in the variable dimension through a one-dimensional convolutional network with the same size to obtain the first branch feature. The second branch has the same structure as the first branch, and the convolution kernel size of the convolutional network is different. The input is the feature representation corresponding to the production process data, and the second branch feature is output. Then, the first branch feature and the second branch feature are added and fused as the output feature of the multi-scale feature extraction module and input to the next cascaded multi-scale feature extraction module until the multi-scale feature extraction module outputs the feature representation of the sample.
[0062] S4. Use the training data set to train a complex industrial process product quality detection model that integrates prior knowledge to obtain a trained complex industrial process product quality detection model that integrates prior knowledge.
[0063] S4.1. Input the production process data of each process in the samples in the training data set into the corresponding feature extraction network respectively to obtain the feature representation corresponding to the production process data of each process.
[0064] Specifically, the feature extraction network in S4.1 is a multi-scale feature extraction network based on a convolutional neural network designed according to the characteristics of the data within the process.
[0065] The feature extraction process of the feature extraction network is as follows:
[0066]
[0067] in, For the The characteristics of each process are represented by Indicates The feature extraction network corresponding to each process, in order to obtain the features of data at multiple scales, the multi-scale feature extraction module designed by the present invention adopts convolution kernels of various sizes to perform parallel feature extraction on the data.
[0068] Specifically, use , The convolution kernel and Sliding extracts features on different time scales, using , The convolution kernel and Extract features at different variable scales, and is the convolution kernel size.
[0069] The multi-scale feature extraction module is as follows:
[0070]
[0071] in, represents the input features of the feature extraction module, represents the output features, is the convolution kernel of the first convolutional network in the first branch, is the convolution kernel of the second convolutional network in the first branch, is the convolution kernel of the first convolutional network in the second branch, is the convolution kernel of the second convolutional network in the second branch. In this embodiment, , .
[0072] S4.2. Analyze the production process of thick plate production and obtain the mutual influence relationship between the processes as prior knowledge; the production process and technology of thick plate products are relatively fixed, which makes it easy to extract relatively common prior knowledge in the production process.
[0073] Specifically, in this example, considering the impact of the front, back and internal processes in the production of thick plates, there are .
[0074] S4.3. Use the acquired prior knowledge to perform feature fusion on the feature representation corresponding to the extracted production process data of each process, and finally obtain the feature representation of the sample, and use the quality inspection module to predict the product quality to obtain the predicted product quality.
[0075] Specifically, the feature fusion based on prior knowledge in S4.3 is shown in the following formula:
[0076]
[0077] in, represents the predicted value of product quality by the complex industrial process product quality detection model integrating prior knowledge, Indicates the splicing operation of the feature representation of each process. Represents the quality inspection module, represents the adjacency matrix, Indicates Characteristic representation of a process.
[0078] S4.4. The training is terminated after reaching the set training rounds or the loss function converges, and a trained complex industrial process product quality detection model integrating prior knowledge is obtained.
[0079] Specifically, the loss function in S4.4 is a cross entropy loss function, using cross entropy loss Calculate the loss between the predicted value of product quality and the actual value of product quality as shown in the following formula:
[0080]
[0081] in, To indicate the quantity of product quality types, The number of the product quality category; is the true value of product quality, The predicted value of product quality.
[0082] S5. Deploy the trained complex industrial process product quality detection model that integrates prior knowledge and use it for product quality detection in complex industries to obtain product quality.
[0083] In this example, the trained complex industrial process product quality detection model integrating prior knowledge is deployed in the server of the digital factory. When production starts, first, step S1 is executed to collect data information in the production process; then, step S2 is executed to preprocess the collected data; finally, the preprocessed data is input into the trained product quality detection model in step S3 to predict the final product quality.
[0084] In order to demonstrate the effect of the present invention in thick plate detection, the corresponding plate quality detection model is trained and tested on the data set D collected in S2. In this example, the recall rate, precision and F1-Score indicators are used to evaluate the effect of the model, and WA represents the weighted average result of each indicator. The results are shown in the following table.
[0085] Table 1 Results of the method proposed in this invention in thick plate shape quality inspection (%)
[0086] index bending Horizontal Wave Left wave Right wave normal WA Accuracy 74.66 76.03 86.85 79.17 90.74 86.48 Recall 63.79 60.11 91.65 49.84 93.52 86.90 F1-Socre 68.79 67.14 89.19 61.17 92.11 86.48
[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.
Claims
1. A complex industrial process product quality detection method integrating prior knowledge, characterized in that: The steps include: S1. Collect relevant data on product production in complex industrial processes as samples; The relevant data includes the production process data of each process collected by each detection sensor during the production process and the product quality label marked after the product is produced; S2, preprocessing the collected samples and using the preprocessed samples to construct a training data set; The preprocessing includes data cleaning, data alignment and normalization; S2 specifically includes: S2.
1. Use data cleaning algorithms to clean the collected samples, remove samples containing sampling anomalies and non-numeric values in the collected production process data, and obtain cleaned samples; S2.2, align and normalize the cleaned samples and construct a training data set; Specifically, firstly, the production process data in the cleaned samples are aligned so that the production process data of each sample have the same length m, and then the aligned production process data are normalized to finally obtain the production process data X corresponding to each process after preprocessing. p , where p is the number of the process and p = 1, 2, ..., n, n is the number of corresponding processes in the production process; therefore, each sample is represented by X i ={X 1 ,X 2 ,…,X n }, i is the sample number, and finally the training data set is obtained Where N is the number of samples, y i For sample X i Corresponding product quality label; S3. Construct a complex industrial process product quality detection model that integrates prior knowledge; The complex industrial process product quality detection model integrating prior knowledge includes several feature extraction networks, a feature fusion module and a quality detection module; The feature extraction network is used to extract features from the production process data in each process to obtain feature representations corresponding to the production process data of each process; the number of the feature extraction networks is equal to the number of corresponding processes in the production process, and one feature extraction network corresponds to one process; The feature fusion module is used to splice the feature representations corresponding to the input production process data of each process to obtain a spliced feature representation, then take the mutual influence between the processes as prior knowledge, and use the prior knowledge to fuse the spliced feature representations, and finally obtain the feature representation of the sample; The prior knowledge is represented by an adjacency matrix, where A = [a pq ] n×n Represents the influence relationship between n processes in the production process, where a pq ∈{0,1} indicates whether the p-th process has an impact on the production of the q-th process, 0 indicates no impact, and 1 indicates impact; The quality detection module is a two-layer fully connected network, which is used to classify the feature representation of the input sample and output the predicted product quality; S4. Using the training data set to train a complex industrial process product quality detection model integrating prior knowledge, to obtain a trained complex industrial process product quality detection model integrating prior knowledge; S4 specifically includes: S4.
1. Input the production process data of each process in the samples in the training data set into the corresponding feature extraction network to obtain the feature representation corresponding to the production process data of each process; The feature extraction process of the feature extraction network is as follows: z p =f p (X p ) Among them, z p is the characteristic representation of the pth process, f p (·) represents the feature extraction network corresponding to the p-th process; The multi-scale feature extraction module in the feature extraction network uses convolution kernels of various sizes for parallel feature extraction. Specifically, convolution kernels W of 1×k1 and 1×k2 are used. t1 and W t2 Sliding extracts features at different time scales, using convolution kernels W of k1×1 and k2×1 respectively. s1 and W s2 Extract features at different variable scales, k1 and k2 are the convolution kernel sizes; The multi-scale feature extraction module is as follows: With p_out =In s1 ×(In t1 ×z p_in )+W s2 ×(In t2 ×z p_in ) Among them, z p_in represents the input features of the feature extraction module, z p_out represents the output feature, W t1 is the convolution kernel of the first convolutional network in the first branch, W s1 is the convolution kernel of the second convolutional network in the first branch, W t2 is the convolution kernel of the first convolutional network in the second branch, W s2 is the convolution kernel of the second convolutional network in the second branch; S4.
2. Analyze the production process in complex industrial processes and obtain the mutual influence relationship between the processes as prior knowledge; S4.
3. Using the acquired prior knowledge, feature fusion is performed on the feature representations corresponding to the extracted production process data of each process, and finally the feature representation of the sample is obtained. The quality detection module is used to predict the product quality to obtain the predicted product quality. Specifically, the feature fusion based on prior knowledge in S4.3 is shown in the following formula: in, represents the predicted value of product quality by the complex industrial process product quality detection model integrating prior knowledge, concat(·) represents the concatenation operation of the feature representations of each process, and f c (·) represents the quality detection module; A represents the adjacency matrix, z n The characteristic representation of the nth process; S4.
4. The training is terminated after reaching the set training rounds or the loss function converges, and a complex industrial process product quality detection model integrating prior knowledge is obtained after training; Specifically, the loss function in S4.4 is a cross entropy loss function, using the cross entropy loss L CE Calculate the loss between the predicted value of product quality and the actual value of product quality as shown in the following formula: Where c represents the number of product quality types, j is the number of product quality types; y j is the true value of product quality, is the predicted value of product quality; S5. Deploy the trained complex industrial process product quality detection model that integrates prior knowledge and use it for product quality detection in complex industries to obtain product quality.
2. A complex industrial process product quality detection method integrating prior knowledge according to claim 1, characterized in that: The data alignment in S2.2 uses a linear interpolation algorithm.
3. A complex industrial process product quality detection method integrating prior knowledge according to claim 1, characterized in that: Each feature extraction network includes a number of multi-scale feature extraction modules connected in sequence, and each multi-scale feature extraction module consists of two branches. The first branch first extracts the feature representation of the input production process data corresponding to the time series through a one-dimensional convolutional network, and then extracts the feature in the variable dimension through a one-dimensional convolutional network with the same size to obtain the first branch feature. The second branch has the same structure as the first branch, and the convolution kernel size of the convolutional network is different. The input is the feature representation corresponding to the production process data, and the second branch feature is output. Then, the first branch feature and the second branch feature are added and fused as the output feature of the multi-scale feature extraction module and input to the next cascaded multi-scale feature extraction module, until the last multi-scale feature extraction module outputs the feature representation of the sample.
