Project quality prediction method based on process variable

Abstract

The invention provides a project quality prediction method based on process variables, which comprises a quality prediction model building process and a quality prediction process, and the quality prediction model building process comprises the following steps: providing training data which comprises process variables and quality variables which are in one-to-one correspondence, preprocessing the input process variable to obtain initial graph structure data; inputting the initial graph structure data into a multilayer graph neural network composed of a multilayer convolutional network and a graph attention network, and extracting graph feature data; taking the graph feature data as input of a long short-term memory network, and capturing time sequence information of the graph feature data; and image feature data time sequence information is output through a full connection layer, corresponding learning is carried out on the image feature data time sequence information and a corresponding quality variable result, a training process is completed, and a quality prediction model is established. According to the method, a process variable is added into a training process, an industrial process prediction model is constructed, and the reliability of a quality prediction model is improved.

Description

Technical Field

[0001] The invention relates to the field of data prediction, and in particular to an engineering quality prediction method based on process variables. Background Art

[0002] The continuous iteration and advancement of technology have promoted the development of industrial processes towards intelligence. However, this has also made industrial processes more and more complex. In order to monitor process quality in a timely and effective manner, reduce costs, and ensure process safety and energy efficiency, the monitoring of quality variables is particularly important. In actual processes, online hardware devices or offline laboratory analyzers are usually used to monitor quality variables. However, these instruments and equipment are expensive, require regular maintenance, are extremely susceptible to adverse environmental conditions, and have some lags in analysis. Therefore, in most industrial processes, it is challenging to implement the monitoring of relevant quality variables. In order to solve these problems, soft measurement technology has been widely developed and applied to predict key quality variables in industrial processes. Soft measurement mainly uses easy-to-measure process variables such as temperature, pressure, and flow to establish mathematical models and predict key quality variables.

[0003] In order to cope with the increasingly complex industrial processes, data-driven methods that do not require detailed knowledge of industrial processes or rich experience have gradually become the mainstream methods of soft sensors. At the same time, the widespread adoption of distributed control systems ensures the feasibility of soft sensor modeling in the context of big data. Typical data-driven methods include principal component analysis (PCA), partial least squares (PLS), support vector machine (SVM) and artificial neural network (ANN). In particular, artificial neural networks have the unique advantage of extracting data features layer by layer, making them very suitable for processing high-dimensional nonlinear data. However, artificial neural networks are a shallow learning structure. As the soft sensor model deepens, problems such as gradient disappearance and gradient explosion continue to plague soft sensor research based on artificial neural networks. Deep learning is a specialized application of artificial neural networks. It overcomes the training difficulties of deep networks by utilizing techniques such as unsupervised hierarchical pre-training and fine-tuning, thereby providing a more effective solution. Therefore, providing a prediction model that applies process variables to the quality prediction process has become an urgent problem to be solved. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting engineering quality based on process variables.

[0005] The technical solution of the present invention is: a method for predicting engineering quality based on process variables, including a process of establishing a quality prediction model and a quality prediction process, wherein the process of establishing a quality prediction module includes the steps of: S1. Provide training data, which includes one-to-one corresponding process variables and quality variables, and obtain initial graph structure data after preprocessing the input process variables; S2, after inputting the initial graph structure data into a multi-layer graph neural network composed of a multi-layer convolutional network and a graph attention network, extracting graph feature data; S3, using the graph feature data as the input of the long short-term memory network to capture the temporal information of the graph feature data; S4. Output the time series information of the graph feature data through the fully connected layer, perform corresponding learning with the corresponding quality variable results, complete the training process, and establish a quality prediction model.

[0006] Preferably, in step S1, a knowledge domain correction table is provided to correct the initial adjacency matrix, and the corrected adjacency matrix is ​​combined with the process variables for preprocessing to obtain initial graph structure data.

[0007] Preferably, in step S2, the multi-layer graph neural network adopts a two-layer structure, each layer of the structure is composed of a multi-layer convolutional network and a graph attention network.

[0008] Preferably, in step S3, the graph feature data is processed continuously by two long short-term memory networks to obtain the time series information of the graph feature data.

[0009] Preferably, in the dimensionality reduction process of step S4, the residual connection data of the original graph structure data is introduced and passed through two fully connected layers together with the time series information to obtain the final result.

[0010] Preferably, during the training process, the Adam algorithm is used as the optimizer.

[0011] Preferably, the quality prediction process includes the steps of: S5. Provide data to be detected, analyze process variables in the data to be detected, obtain an initial adjacency matrix to be detected and an initial graph structure to be detected, correct the initial adjacency matrix to be detected, and refine and adjust the initial graph structure to be detected with the corrected adjacency matrix to be detected; S6. Input the initial graph structure to be detected into the quality prediction model, extract graph features using a multi-layer graph neural network structure consisting of a multi-layer graph convolutional network and a graph attention network, connect the outputs of the two layers and input them into a long short-term memory network to capture the time dependency in the sequential data; S7. Pass the results captured by the long short-term memory network layer through the fully connected layer to obtain the prediction results of the quality variables.

[0012] Preferably, in step S5, the maximum information coefficient method is used to refine and adjust the initial graph structure to be detected.

[0013] Preferably, the initial adjacency matrix to be detected is corrected using a domain knowledge correction table.

[0014] In the above method, process variables are added to the training process to construct an industrial process prediction model, which improves the reliability of the quality prediction model. In addition, the domain knowledge (i.e., expert experience and process mechanism knowledge to determine process variables) and quality information are fully utilized, i.e., a knowledge-guided supervised hybrid graph prediction model. After performing correlation analysis on the original process data and exploring the relationship between variables, the knowledge domain correction table is used to optimize the graph structure to achieve high-precision industrial process quality prediction. During the training process, the Adam algorithm is used as the optimizer, which has the advantage of adaptive learning rate adjustment and helps the model converge to the optimal solution faster. During the prediction process, the maximum information coefficient method is used to refine and adjust the initial structure graph to be detected, that is, in the case of prior knowledge, the influence between two nodes may be relatively small, but after MIC calculation, it is found that there is a strong correlation between them. Therefore, the domain knowledge correction table can be used to correct this relationship. This correction process of the graph structure enhances the interpretability of the constructed adjacency matrix and improves the accuracy of graph model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The embodiments of the present invention will be described more fully with reference to the attached drawings, which are provided for illustration and description only and are not intended to limit the scope of the present invention.

[0016] Figure 1 The following is a flow chart of a method for predicting engineering quality based on process variables according to the present invention. Figure 1 ; Figure 2 The following is a flow chart of a method for predicting engineering quality based on process variables according to the present invention. Figure 2 ; Figure 3 This is an example diagram of correction of the adjacency matrix in the engineering quality prediction method based on process variables of the present invention. DETAILED DESCRIPTION

[0017] The following is a detailed description of an engineering quality prediction method based on process variables of the present invention in conjunction with the accompanying drawings and specific embodiments.

[0018] In the description of the present invention, it should be understood that the orientations or positional relationships indicated by the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc. are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. At the same time, the terms "first", "second", etc. are only used to distinguish the names of the components, and there is no primary or secondary relationship, and therefore cannot be understood as a limitation on the present invention. Figure 1 As shown, a method for predicting engineering quality based on process variables includes a process of establishing a quality prediction model and a quality prediction process, wherein the process of establishing a quality prediction module includes the steps of: S1. Provide training data, the training data including one-to-one corresponding process variables and quality variables, and obtain initial graph structure data after preprocessing the process variables; S2, after inputting the initial graph structure data into a multi-layer graph neural network composed of a multi-layer convolutional network (Graph Convolution Networks, GCN) and a graph attention network (Graph Neural Network, GNN), the graph feature data is extracted; S3. Use the graph feature data as the input of the Long Short-Term Memory (LSTM) network to capture the timing information of the graph feature data:

[0019]

[0020]

[0021]

[0022]

[0023]

[0024] The characteristic of LSTM is that it uses some gate structures as basic units. It's the Gate of Forgetfulness. is the input gate, is the output gate; h represents the hidden layer state, and the result of each gate function is determined by the hidden layer state at the previous moment. And the input features at the current moment Jointly determined, so the time series information features can be extracted. 𝜎 represents a nonlinear activation function, represents the intermediate state variable of LSTM, ⊙ represents the multiplication of corresponding elements one by one, represents the hidden state variable, , , , , , , Represents the weights and bias parameters that need to be learned.

[0025] S4. Output the time series information of the graph feature data through the fully connected layer, perform corresponding learning with the corresponding quality variable results, complete the training process, and establish a quality prediction model.

[0026] Specifically, in the process of establishing the quality prediction model, firstly, the training data consisting of one-to-one corresponding process variables and quality variables are collected, and the training data is preprocessed to obtain the initial graph structure data. After that, the initial graph structure data is input into a multi-layer graph neural network consisting of a multi-layer convolutional network and a graph attention network, and the graph feature data is extracted, that is, the graph structure data corrected under the guidance of knowledge is obtained. After that, the graph feature data is input into the long short-term memory network for training to capture the time series information of the graph feature data. Finally, the time series information of the graph feature data is output through the fully connected layer, and the corresponding quality variable results are learned to complete the training process. Adding process variables to the training process improves the reliability of the quality prediction model. It is worth pointing out that the above-mentioned process variables refer to data that can be measured and obtained by sensors in the industrial production process, such as temperature, pressure, etc.

[0027] Further, in step S1, the preprocessing process of the input data includes correcting the initial adjacency matrix, and the corrected adjacency matrix is ​​combined with the process variables for preprocessing to obtain the initial graph structure data. It is worth pointing out that the correction process of the initial adjacency matrix is ​​adjusted through the relevant field knowledge between the process variables, that is, the technicians in this field analyze the correlation between the process variables through theoretical knowledge to correct the initial adjacency matrix, thereby better preprocessing the process variable data and improving the accuracy of the subsequent training prediction process.

[0028] Specifically, the correction process of the initial adjacency matrix includes: providing a knowledge domain correction table to correct the initial adjacency matrix. It is worth pointing out that the knowledge domain correction table refers to a data table composed of correlation data between different process variables in different industrial production processes.

[0029] Specifically, MIC uses the concept of MI (mutual information) to standardize the calculation and ensure that the result is in the range of [0,1], that is: Formula (7) Where 𝑝(i) and 𝑝(𝑗) are the marginal probability density functions of random variables i and 𝑗, respectively. Therefore, the MIC calculation result between variables i and 𝑗 can be obtained as: Formula (8) in , Indicates the dimension of the dataset.

[0030] So, It represents the initial adjacency matrix obtained after MIC calculation. The initial adjacency matrix A can be calculated by the following formula: Formula (9) Formula (10) However, the elements in the matrix represents the correlation between the i-th variable and the j-th variable, The value of can only be 1 or 0. The former indicates that there is a strong correlation between the two variables, and the latter indicates that the correlation between the two variables is weak. Since the initial adjacency matrix A may contain incorrect variable correlation information. For example, in a certain industrial process, in the presence of prior knowledge, that is, when the influence between two nodes (i.e., two process variables) is relatively small, the corresponding value of these two nodes in the knowledge domain correction table is 0. However, after MIC calculation, it is found that there is a strong correlation between them. For example, variable 1 represents the top temperature and variable 2 represents the top pressure. According to the MIC calculation results, the initial adjacency matrix Elements in and The values ​​of are all 0. However, according to expert experience and mechanism knowledge, the temperature and pressure at the same location are highly correlated, that is, in the knowledge domain correction table, the node data corresponding to the temperature and pressure at the same location is 1. Therefore, if Figure 2 As shown, the domain knowledge correction table can be used to correct this relationship, that is, under the guidance of the domain knowledge correction table, and The value of is corrected to 1 to form the corrected final adjacency matrix as part of the input of the hybrid prediction model. This graph structure correction process enhances the interpretability of the constructed adjacency matrix and improves the accuracy of graph model training.

[0031] Further, the initial graph structure is defined as , where V is the set of nodes, E is the set of edges, A∈{0,1}𝑛×𝑛 is the adjacency matrix of graph 𝐺, and 𝑛 is the number of nodes. The original time series data is represented as 𝐷={𝑋𝑡,𝑌𝑡}, 𝑡= 1,2,⋯𝑚, 𝑋𝑡∈𝑅𝑛,𝑌𝑡∈𝑅, 𝑅𝑛 and 𝑅 refer to the n-dimensional real number domain and real number domain. Here, 𝑋𝑡 represents the process variable at time 𝑡, and 𝑌𝑡 represents the key quality variable at time 𝑡. 𝑚 is the sample size of the data set, 𝑛 is the dimension of the process variable 𝑋𝑡, and is also the number of nodes in the graph structure. Therefore, the node feature 𝑋𝐺 can be defined as follows:

[0032] In order to make full use of quality variables for supervised learning, we introduced 𝑌𝑡 as one of the important nodes in the graph structure. In the process of graph construction, 𝑌𝑡 and 𝑋𝑡 are used together for correlation analysis and correction of the knowledge domain correction table obtained under knowledge guidance. The following example analyzes the refinement process of the initial structure graph, as shown in Table 1: Table 1 List of process variables Process variable name Variable Description <![CDATA[u 1 ]]> Top temperature <![CDATA[u 2 ]]> Top pressure <![CDATA[u 3 ]]> Return flow <![CDATA[u 4 ]]> Flow to the next process <![CDATA[u 5 ]]> Temperature of plate 6 <![CDATA[u 6 ]]> Bottom temperature A <![CDATA[u 7 ]]> Bottom temperature B y C4 content In the graph structure, node 1 represents the highest temperature, and node 2 represents the highest pressure. The original value of the adjacency matrix between node 1 and node 2 is 0. According to physical principles, the temperature and pressure at the same position are highly correlated, that is, the temperature and pressure nodes at the same position in the knowledge domain correction table are related, so the unconnected nodes are modified according to the knowledge domain correction table to connect the nodes. Since the threshold of these two nodes in the knowledge domain correction table is a value greater than the threshold of 0.2, the values ​​of the corresponding nodes in the adjacency matrix are modified accordingly. According to the actual deployment, the correlation between the top pressure and temperature and the bottom temperature is low, that is, the values ​​between these nodes in the knowledge domain correction table are all 0. Then, according to the knowledge domain correction table, the connected nodes in the original graph structure are disconnected, that is, nodes 1 and 2 are no longer associated with nodes 6 and 7, and the corresponding values ​​of the adjacency matrix are changed to 0. According to expert experience, the key quality variables are closely related to the process variables, so a value greater than 0.2 is assigned between these two nodes in the adjacency matrix. Therefore, the corrected node 8 is now connected to all other nodes, and a value greater than 0.2 is assigned to each node. For those nodes whose relevance is difficult to determine through domain knowledge, we do not make further adjustments and retain the adjacency matrix and graph structure obtained through MIC.

[0033] Furthermore, in step S2, GCN is used to obtain the embedding features of different nodes, and GAT is used to highlight the contribution of different nodes in information aggregation. A two-layer structure is adopted, and each layer consists of a GCN and a GAT.

[0034] Furthermore, in step S3, the graph feature data is processed by two long short-term memory networks in succession to obtain the time series information of the graph feature data. The double-layer LSTM can extract time series features better than the single-layer. Furthermore, in the dimensionality reduction process of step S4, the residual connection data of the original graph structure data is introduced, and it passes through two layers of fully connected layers together with the time series information to obtain the final result. The introduction of residual connection data reduces the complexity of the model and alleviates the problem of gradient disappearance.

[0035] Further, in the training process, first, the initial adjacency matrix is ​​corrected using the correction table of relevant domain knowledge, that is, after exploring the relationship between variables, the process variable data can be better preprocessed to improve the accuracy of the subsequent training prediction process. Secondly, the Adam algorithm is used as the optimizer, which has the advantage of adaptive learning rate adjustment, which helps the model converge to the optimal solution faster. That is, in the training process of S1-S4, after the model structure is determined, the back propagation algorithm needs to be used for model training and parameter solution, and the optimization algorithm used is the Adam algorithm. Because in the training process of the neural network, the selection of hyperparameters is particularly important because it directly affects the learning ability of the neural network. In this case, the trial and error method is used to determine the model hyperparameters. In the training stage, the Adam algorithm is used as the optimizer, which has the advantage of adaptive learning rate adjustment, which helps the model converge to the optimal solution faster.

[0036] Further, if Figure 3 As shown, the quality prediction process includes the following steps: S5. Provide data to be detected, analyze process variables in the data to be detected, obtain an initial adjacency matrix to be detected and an initial graph structure to be detected, correct the initial adjacency matrix to be detected, and refine and adjust the initial graph structure with the corrected adjacency matrix to be detected; S6. Input the initial graph structure into the quality prediction model, and use the multi-layer graph neural network (GNN) structure composed of a multi-layer graph convolutional network (GCN) and a graph attention network (GAT) to extract graph features. The outputs of the two layers are connected and input into the LSTM layer to capture the temporal dependency in the sequential data. S7. Pass the results captured by the LSTM layer through the fully connected layer to obtain the prediction results of the quality variables.

[0037] Furthermore, in step S5, the maximum information coefficient method is used to refine and adjust the initial structure graph.

[0038] Furthermore, MIC adopts the concept of MI (mutual information) to standardize the calculation and ensure that the result is in the range of [0,1].

[0039] Furthermore, the knowledge domain correction table is used to correct the initial adjacency matrix to be detected.

[0040] Further, in step S6, the initial graph structure data to be detected is sent to the GCN layer. As shown in the following formula, a degree matrix is ​​calculated using the corrected input adjacency matrix: .

[0041] Formula (11) in is Ai,j after the self-loop, i and j represent the serial numbers of two different nodes, and Ai,j is the adjacency matrix. Next, use the following formula to transfer features: Formula (12) in express A collection of express , h (𝐿+1) represents the feature variable of the next node, 𝑊(𝐿) is the parameter matrix that can be learned through model training, and σ represents the sigmoid activation function.

[0042] Then, continue to propagate the node features to the GAT layer, then: Formula (13) in, Represents the node feature variable of the current layer, represents the node feature variable of the previous layer, 𝑊𝐿 is a learnable parameter matrix, and represents the normalized attention coefficient between node i and node j.

[0043] In the above method, process variables are added to the training process to construct an industrial process prediction model, which improves the reliability of the quality prediction model. In addition, domain knowledge (i.e., expert experience and process mechanism knowledge to determine process variables) and quality information are fully utilized, i.e., a knowledge-guided supervised hybrid graph prediction model. After correlation analysis of the original process data and exploration of the relationship between variables, the knowledge domain correction table is used to optimize the graph structure to achieve high-precision industrial process quality prediction. In the training process, the Adam algorithm is used as the optimizer, which has the advantage of adaptive learning rate adjustment and helps the model converge to the optimal solution faster. In the prediction process, the maximum information coefficient method is used to refine and adjust the initial structure graph to be detected, that is, in the case of prior knowledge, the influence between two nodes may be relatively small, but after MIC calculation, it is found that there is a strong correlation between them. Therefore, the domain knowledge correction table can be used to correct this relationship. This correction process of the graph structure enhances the interpretability of the constructed adjacency matrix and improves the accuracy of graph model training.

[0044] For those skilled in the art, various changes and modifications will undoubtedly be obvious after reading the above description. Therefore, the attached claims should be regarded as covering all changes and modifications of the true intent and scope of the present invention. Any and all equivalent ranges and contents within the scope of the claims should be considered to still be within the intent and scope of the present invention.

Claims

1. A method for predicting engineering quality based on process variables, comprising a quality prediction model establishment process and a quality prediction process, characterized in that: The process of building a quality prediction module includes the following steps: S1. Provide training data, which includes one-to-one corresponding process variables and quality variables, and obtain initial graph structure data after preprocessing the input process variables; S2, after inputting the initial graph structure data into a multi-layer graph neural network composed of a multi-layer convolutional network and a graph attention network, extracting graph feature data; S3, using the graph feature data as the input of the long short-term memory network to capture the temporal information of the graph feature data; S4. Output the time series information of the graph feature data through the fully connected layer, perform corresponding learning with the corresponding quality variable results, complete the training process, and establish a quality prediction model.

2. The engineering quality prediction method based on process variables according to claim 1 is characterized in that: In step S1, a knowledge domain correction table is provided to correct the initial adjacency matrix, and the corrected adjacency matrix is ​​combined with process variables for preprocessing to obtain initial graph structure data.

3. The engineering quality prediction method based on process variables according to claim 2 is characterized in that: In step S2, the multi-layer graph neural network adopts a two-layer structure, each layer of which is composed of a multi-layer convolutional network and a graph attention network.

4. The engineering quality prediction method based on process variables according to claim 3 is characterized in that: In step S3, the graph feature data is processed continuously by two long short-term memory networks to obtain the time series information of the graph feature data.

5. The engineering quality prediction method based on process variables according to claim 4 is characterized in that: In the dimensionality reduction process of step S4, the residual connection data of the original graph structure data is introduced and passed through two fully connected layers together with the time series information to obtain the final result.

6. The engineering quality prediction method based on process variables according to claim 5 is characterized in that: During the training process, the Adam algorithm is used as the optimizer.

7. The engineering quality prediction method based on process variables according to claim 2 is characterized in that: The quality prediction process includes the following steps: S5. Provide data to be detected, analyze process variables in the data to be detected, obtain an initial adjacency matrix to be detected and an initial graph structure to be detected, correct the initial adjacency matrix to be detected, and refine and adjust the initial graph structure to be detected with the corrected adjacency matrix to be detected; S6. Input the initial graph structure to be detected into the quality prediction model, extract graph features using a multi-layer graph neural network structure consisting of a multi-layer graph convolutional network and a graph attention network, connect the outputs of the two layers and input them into a long short-term memory network to capture the time dependency in the sequential data; S7. Pass the results captured by the long short-term memory network layer through the fully connected layer to obtain the prediction results of the quality variables.

8. The method for predicting engineering quality based on process variables according to claim 7 is characterized in that: In step S5, the maximum information coefficient method is used to refine and adjust the initial graph structure to be detected.

9. The engineering quality prediction method based on process variables according to claim 8 is characterized in that: The domain knowledge correction table is used to correct the initial adjacency matrix to be detected.

Images