Industrial quality prediction method and device based on graph attention long short-term memory network

Abstract

The application relates to an industrial quality prediction method and equipment based on a graph attention long short-term memory network, which comprises the following steps: acquiring a data set to be predicted in an industrial manufacturing process, performing a standardization operation on the data set to obtain a standardized data set; acquiring a low-order feature variable matrix of the standardized data set through a multilayer perception machine, and constructing an adjacency matrix; inputting the standardized data set, the adjacency matrix and the low-order feature variable matrix into a trained graph attention long short-term memory network model, splicing time feature information and space feature information output by the model through a quality estimator, and obtaining a quality prediction result; the graph attention long short-term memory network model comprises a graph attention layer and a long short-term memory layer arranged in parallel, and is used for extracting the time feature information and the space feature information, respectively. Compared with the prior art, the application has the advantages of fusing time-space features, automatically constructing an adjacency matrix and improving prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of product quality prediction technology, and in particular to an industrial quality prediction method and device based on graph attention long short-term memory networks. Background Technology

[0002] With the expansion of production scale and rapid technological development, quality control in modern industry has become increasingly complex and crucial. Effective process quality control models can predict product manufacturing quality early in the production process to improve processes. Quality control is a vital means of ensuring product quality during manufacturing, and intelligent manufacturing places higher demands on the real-time nature and accuracy of product quality information. Advanced information technologies such as industrial big data and artificial intelligence provide new means for real-time quality prediction of products produced in industrial processes.

[0003] Traditional product quality forecasting is mostly based on statistical process control (SPC) studies, which predict the final product quality by studying the changing trends of quality parameters in the production process. These SPC-based quality forecasting methods largely rely on quality information from the key quality parameters of the product manufacturing process itself, and they depend heavily on the experience and knowledge of quality control experts, thus lacking universality.

[0004] With the rapid development of internet technology and the Internet of Things (IoT), the amount of data collected is unprecedented. Due to limitations in the number of model parameters and operating speed, traditional SPC-based quality prediction methods are unsuitable for industrial big data scenarios. In recent years, quality prediction methods based on artificial intelligence algorithms have been proposed. These methods model the characteristics of quality data generated in the manufacturing process and predict product quality by analyzing the model results. These AI networks automatically represent abstract features from process data using deep hierarchical structures and then directly establish the relationship between learned features and target output, effectively solving the data explosion problem. Compared with traditional methods, these methods have certain optimizations in feature extraction and quality prediction capabilities, and have shown significant improvements in quality prediction errors. However, most deep learning methods ignore the correlation between data and fail to extract high-level information from the data. Furthermore, deep learning models are typical "black box models," meaning that model training adopts an "end-to-end" decision-making model. It is difficult to know what knowledge the deep model has learned from the data, and it is hard to understand the impact of network layers, number of neurons, and activation functions on the results. This makes the effectiveness of deep learning more dependent on past engineering experience and parameter tuning skills.

[0005] In recent years, graph neural networks have become a research hotspot due to their superior performance and interpretability, and have been applied in many fields, such as image recognition, point cloud segmentation, recommender systems, traffic prediction, and chemical structure research. Because data structures possess non-Euclidean space characteristics, deep learning models cannot extract these features. However, in the graph domain, the geometric structure of data can provide additional information, including not only the values ​​of nodes but also the relationships between them. Therefore, compared to general data domains, graph domains can provide more information about the relationships between data. However, the application of graph neural networks in quality prediction for large-scale industrial processes is still relatively limited.

[0006] A search revealed Chinese invention patent application publication number CN117421580A, which discloses a method for predicting process quality time series based on graph neural networks and transfer learning. This method involves collecting process parameters and quality indicators from a production line using a Manufacturing Execution System (MES) to form a dataset. Next, all variables are mapped to nodes, with the corresponding production data serving as node features. The control relationships between process parameters and quality indicators are mapped to edges between nodes, transforming the process relationships into graph data. A graph neural network is then used to calculate the spatial features within the graph data. Simultaneously, a bidirectional long short-term memory network is used to calculate the temporal features of the quality indicators themselves. Finally, the spatial information and temporal features are concatenated to obtain the prediction result. A transfer learning mechanism is then introduced to update the model and output the prediction result. However, this existing patent application involves graph construction, directly constructing graphs from the acquired data, and the process parameters are derived from expert experience. Therefore, it suffers from limitations in input parameters and inaccurate predictions.

[0007] How to achieve accurate prediction of industrial process quality based on graph neural networks has become a technical problem that needs to be solved. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an industrial quality prediction method and device based on graph attention long short-term memory network.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] According to one aspect of the present invention, an industrial quality prediction method based on graph attention long short-term memory networks is provided, the method comprising:

[0011] Step S1: Obtain the dataset of the industrial manufacturing process to be predicted, and perform a standardization operation on the dataset to obtain a standardized dataset;

[0012] Step S2: Obtain the low-order feature variable matrix of the standardized dataset through a multilayer perceptron and construct the adjacency matrix;

[0013] Step S3: Input the standardized dataset, adjacency matrix, and low-order feature variable matrix into the trained graph attention long short-term memory network model. The output temporal and spatial feature information is concatenated and passed through the quality predictor to obtain the quality prediction result.

[0014] The graph attention long short-term memory network model includes parallel graph attention layers and long short-term memory layers, which are used to extract temporal feature information and spatial feature information, respectively.

[0015] Preferably, after obtaining the low-order feature variable matrix of the standardized dataset using a multilayer perceptron in step S2, the K-adjacency algorithm is used to calculate the Euclidean distance or cosine distance. For each variable feature, the K nearest features are selected as adjacent samples to construct an adjacency matrix, thereby converting the one-dimensional signal data into graph data.

[0016] Preferably, different variable parameters are selected to obtain adjacency matrices with different receptive fields, and the spatial features of different receptive fields are fused through feature splicing operations.

[0017] More preferably, the variable parameters are 1, 2, and 3.

[0018] Preferably, in step S3, the adjacency matrix and the low-order feature variable matrix are input into the graph attention long short-term memory network model to extract spatial feature information; and the standardized dataset is input into the graph attention long short-term memory network model to extract temporal feature information.

[0019] Preferably, the calculation process of the long short-term memory layer is as follows:

[0020] i t =σ(W ii d t +b ii +W hi h (t-1) +b hi )

[0021] f t =σ(W if d t +b if +W hf h (t-1) +b hf )

[0022] g t =tanh(W ig d t +b ig +W hg h (t-1) +b hg )

[0023] ot =σ(W io d t +b io +W ho h (t-1) +b ho )

[0024] c t =f t *c (t-1) +i t *g t

[0025] h t =o t *tanh(c t )

[0026] Among them, i t For the input gate, f t For the Gate of Oblivion, g t The current input unit state, o t For output gate, c t h represents the current cell state. t The hidden states of the LSTM constitute the temporal features H extracted by the LSTM. t W ii W if W ig W io These represent the input gate, forget gate, current input unit, and output gate at the current time, respectively. t The weight parameter, W hi W hf W hg W ho The h values ​​in the current input gate, forget gate, current input unit, and output gate are respectively... (t-1) The weight parameter, b ii b if b ig b io These represent the input gate, forget gate, current input unit, and output gate at the current time, respectively. t The bias value, b hi b hf b hg b ho These are the h values ​​for the current input gate, forget gate, current input unit, and output gate, respectively. (t-1) The bias value, h (t-1) Let d be the hidden state of the LSTM at time t-1. t This refers to the data entered at the current moment.

[0027] Preferably, the calculation process of the quality predictor is as follows:

[0028]

[0029] in, Indicates the predicted quality value; FC i This represents the i-th fully connected layer, where i = 1, 2; Flat is the expansion function, i.e., the feature tiling operation; O represents the spliced ​​temporal and spatial feature information.

[0030] Preferably, the loss function of the quality predictor is the mean squared error function, and the n features with large weights on the prediction result are obtained by backpropagation through the loss function, so as to update the training parameters of the graph attention long short-term memory network model.

[0031] Preferably, the dataset includes manufacturing process data such as temperature, pressure, temperature × pressure, material fusion index, and material conversion index.

[0032] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1) This invention uses a multilayer perceptron to obtain low-order features of the original variables, and then uses the K-adjacency algorithm to calculate and construct an adjacency matrix. The one-dimensional process signal is transformed into graph data through graph construction technology. A graph attention network is used to extract spatial features, thereby making fuller use of the correlation between process variables. A long short-term memory network is used to extract temporal features, providing more correlation information for the model's quality prediction. The spatiotemporal features are concatenated to enhance the richness of feature information. The prediction results are output by the quality predictor, thereby improving the prediction accuracy.

[0035] 2) This invention uses a multilayer perceptron for feature pre-selection, which is equivalent to noise reduction, to select variables containing more feature information. Then, in the part of the graph attention network that extracts spatial features, a network with multiple receptive fields is used to extract feature information with different numbers of neighbors, thereby enriching the feature information in space and enhancing the quality prediction effect.

[0036] 3) This invention uses the K-adjacency algorithm to calculate distances. For each variable feature, the K nearest features are selected as adjacent samples according to the distance sorting and an adjacency matrix is ​​constructed to automatically form spatial correlation features, rather than relying on expert experience to determine spatial correlation. This avoids the limitations of expert experience and enhances the adaptability and prediction accuracy of the model.

[0037] 4) By reverse-engineering the attention coefficient, this invention can identify several manufacturing parameters that have a high weight in quality prediction, thus providing interpretability analysis for the model's prediction process. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the industrial quality prediction method based on graph attention long short-term memory network in this invention.

[0039] Figure 2 This is a schematic diagram of the architecture of the attention long short-term memory network model in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0041] This embodiment relates to an industrial quality prediction method based on graph attention long short-term memory network. This method is based on graph neural network and long short-term memory network, abandons the model of manual parameter tuning, performs in-depth mining of industrial data, effectively expresses the influence relationship between data, integrates the spatial and temporal characteristics of manufacturing parameters, improves the accuracy of industrial process quality prediction, and thus performs quality control of manufacturing process.

[0042] This paper extracts spatial features of data using graph neural networks and temporal features using long short-term memory networks. Finally, a two-layer fully connected layer is used to construct a product quality prediction estimator to predict the manufacturing quality of products, thereby improving the accuracy of product prediction in industrial manufacturing processes and reducing quality prediction errors. Furthermore, based on the attention coefficient mechanism of graph neural networks, several manufacturing parameters most important for quality prediction are derived, providing interpretability analysis for the model's prediction results.

[0043] To improve the interpretability of the model, several production parameters with the greatest weight on the final product quality are derived from the attention coefficient of the graph neural network, providing interpretability analysis for the model's prediction results.

[0044] An industrial quality prediction method based on graph attention long short-term memory network, the method includes the following steps:

[0045] Step 1: Obtain the sample dataset D to be processed, with a quality score of Y. Perform standardization on the dataset to obtain the standardized dataset S(D).

[0046] Step 2: Obtain the low-order feature variable matrix X of the sample through multilayer perceptron (MLP), calculate the distance between the feature vectors of each variable in X, and select the K nearest variable features as the adjacency variables for each variable feature to construct the adjacency matrix A.

[0047] Step 3: Input the adjacency matrix A and the feature matrix X into the graph attention network model to extract spatial feature information Z. S ;

[0048] Step 4: Input the standardized raw sample data features S(D) into the Long Short-Term Memory network to extract the temporal feature information H of the sample features. T ;

[0049] Step 5: Extract the spatial feature information Z from the graph attention network. S Temporal feature information H extracted from Long Short-Term Memory Network T The stitching is used as an input feature for the final quality predictor;

[0050] Step 6: Combine the quality prediction scores obtained from the two fully connected layer predictors. The true quality score Y is input into the mean squared error (MSE) formula to obtain the final loss function.

[0051] Step 7: Update the model training parameters through backpropagation of the loss function to improve the model's quality prediction ability.

[0052] This embodiment also relates to an industrial quality prediction method based on a graph attention long short-term memory network. This method was validated in the following experimental environment: the hardware configuration used was an Intel(R) Core(TM) i7-12700H processor and a GTX 3050 graphics card; the software environment was CUDA 11.3 and cuDNN 8.0; and the development environment was Windows 11. The graph attention long short-term memory network model was implemented using PyCharm and the open-source deep learning framework PyTorch 1.12.1.

[0053] The method specifically includes the following steps:

[0054] Step 1: Use the Kaggle publicly available quality prediction dataset. This dataset mainly contains 5 process control parameters and 1 quality rating score. The control parameters include temperature, pressure, temperature × pressure, material fusion index, and material conversion index. The quality rating score is a percentage of expert scores. The dataset contains a total of 3958 manufacturing process data points. A window size W=5, sliding step S=1, and a 7:3 data split are used, resulting in 2770 training data points and 1188 test data points. m*nThere are a total of 3958 data points, each with 5 process control variables, so m = 3958 and n = 5. The dataset is standardized to obtain S(D), and the transformation function is:

[0055]

[0056] Where d∈D, is each data point in D, u is the mean of all variables in that data point, and σ is the standard deviation of all variables in that data point.

[0057] Step 2: Obtain the low-order feature variable matrix X of the sample using a multilayer perceptron (MLP). The expression for X is:

[0058] X = MLP(D) (2)

[0059] D is the original variable dataset, and X is the feature dataset output by the MLP. After standardization, the distance (Euclidean distance) between the feature vectors of each variable in X is calculated:

[0060]

[0061] Where, x :,i Let dist(x) be the i-th variable vector of the feature matrix X. :,i ,x :,j Let be the distance between the i-th variable feature data and the j-th variable feature data in X. For each variable feature, select the K nearest sample features as adjacent samples and construct an adjacency matrix A, where K is a variable parameter, A∈R. m*m It is a sparse matrix of 0-1.

[0062] The distance formula in step two can be replaced with cosine distance:

[0063]

[0064] Cosine distance is not affected by vector dimension and can better handle directional information between vectors, thus extracting more spatial feature information.

[0065] Step 3: Input the adjacency matrix A and the feature variable matrix X into the graph attention network (GAT) to extract the spatial feature information Z of the data. S The formula is as follows:

[0066] Z S =GAT(X,A) (5)

[0067] In step three, the adjacency matrix A is determined by the variable parameter K. To extract spatial feature information more fully, K = 1, 2, 3 is used to extract feature information with different receptive fields. Finally, a feature concatenation operation is used to obtain the final spatial features, as shown in the formula:

[0068] Z S,F =⊙{Z S,1 Z S,2 Z S,3} (6)

[0069] Z S,1 Z S,2 Z S,3 These represent the spatial features extracted when K = 1, 2, and 3, respectively.

[0070] Step 4: Input the standardized data S(D) into a Long Short-Term Memory (LSTM) network to extract temporal features H. t The LSTM calculation process is as follows:

[0071] i t =σ(W ii d t +b ii +W hi h (t-1) +b hi (7)

[0072] f t =σ(W if d t +b if +W hf h (t-1) +b hf (8)

[0073] g t =tanh(W ig d t +b ig +W hg h (t-1) +b hg (9)

[0074] o t =σ(W io d t +b io +W ho h (t-1) +b ho (10)

[0075] c t =f t *c (t-1) +i t *g t (11)

[0076] h t =o t *tanh (c t(12)

[0077] Among them, i t For the input gate, f t For the Gate of Oblivion, g t The current input unit state, o t For output gate, c t h represents the current cell state. t The hidden states of the LSTM constitute the temporal features H extracted by the LSTM. t W ii W if W ig W io d in each gate at the current time t The weight parameter, W hi W hf W hg W ho h in each gate at the current time (t-1) The weight parameter, b ii b if b ig b io d in each gate at the current time t The bias value, b hi b hf b hg b ho Each of the gates at the current time is represented by h. (t-1) The bias value, h (t-1) Let d be the hidden state of the LSTM at time t-1. t This refers to the data entered at the current moment.

[0078] Step 5: Extract the spatial feature information Z from the graph attention network. S,F Temporal feature information H extracted from Long Short-Term Memory Network T The concatenation is used as the input feature O for the final quality predictor, as shown in the following formula:

[0079] O=⊙{Z S,F H t} (13)

[0080]

[0081] in, Indicates the predicted quality value; FC i This represents the i-th fully connected layer; Flat is the expansion function, i.e., the feature tiling operation.

[0082] Step 6: Combine the quality prediction scores obtained from the two fully connected layer predictors. The true quality score Y is input into the mean squared error (MSE) to obtain the final loss function:

[0083]

[0084] in, and y i represents the predicted and true values ​​of the i-th sample, respectively; N is the number of samples, i.e., the number of 2770 training datasets.

[0085] Step 7: Update the model training parameters through backpropagation using the MSE loss function to improve the model's predictive quality. By backpropagating the trained model parameters, the features with the highest weights on the prediction results are derived, providing interpretability analysis for the model's prediction process.

[0086] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0087] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0088] The processing unit performs the various methods and processes described above. For example, in some embodiments, the methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the methods by any other suitable means (e.g., by means of firmware).

[0089] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0090] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0091] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An industrial quality prediction method based on graph attention long short-term memory networks, characterized in that, The method includes: Step S1: Obtain the dataset of the industrial manufacturing process to be predicted, and perform a standardization operation on the dataset to obtain a standardized dataset; Step S2: Obtain the low-order feature variable matrix of the standardized dataset using a multilayer perceptron and construct an adjacency matrix. After obtaining the low-order feature variable matrix of the standardized dataset using a multilayer perceptron in Step S2, calculate the Euclidean distance or cosine distance using the K-adjacency algorithm. For each variable feature, select the K nearest features as adjacent samples and construct an adjacency matrix, thereby converting the one-dimensional signal data into graph data. Select different variable parameters to obtain adjacency matrices with different receptive fields, and fuse the spatial features of different receptive fields through feature concatenation operations. Step S3: Input the standardized dataset, adjacency matrix, and low-order feature variable matrix into the trained graph attention long short-term memory network model. The output temporal and spatial feature information is concatenated and passed through the quality predictor to obtain the quality prediction result. At the same time, based on the attention coefficient mechanism of the graph neural network, several manufacturing parameters most important for quality prediction are deduced. In step S3, the adjacency matrix and low-order feature variable matrix are input into the graph attention long short-term memory network model to extract spatial feature information; the standardized dataset is input into the graph attention long short-term memory network model to extract temporal feature information. The graph attention long short-term memory network model includes parallel graph attention layers and long short-term memory layers, which are used to extract temporal feature information and spatial feature information, respectively.

2. The industrial quality prediction method based on graph attention long short-term memory network as described in claim 1, characterized in that, The variable parameters are 1, 2, and 3.

3. The industrial quality prediction method based on graph attention long short-term memory network as described in claim 1, characterized in that, The calculation process for the Long Short-Term Memory layer is as follows: in, For input gate, For the Gate of Oblivion This represents the current state of the input unit. For output gate, This represents the current state of the cell. The hidden states of the LSTM constitute the temporal features extracted by the LSTM. ; , , , These are the current input gate, forget gate, current input unit, and output gate, respectively. The weight parameters, , , , These are the input gate, forget gate, current input unit, and output gate at the current time. The weight parameters, , , , These are the current input gate, forget gate, current input unit, and output gate, respectively. The bias value, , , , These are the current input gate, forget gate, current input unit, and output gate, respectively. The bias value, Let t-1 be the hidden state of the LSTM. This refers to the data entered at the current moment.

4. The industrial quality prediction method based on graph attention long short-term memory network as described in claim 1, characterized in that, The calculation process of the quality predictor is as follows: in, Indicates the predicted quality value; Indicates the first There are 1 fully connected layer, i=1,2; Flat is the expansion function, i.e., feature tiling operation; O This refers to the spliced ​​temporal and spatial feature information.

5. The industrial quality prediction method based on graph attention long short-term memory network as described in claim 1, characterized in that, The loss function of the quality predictor is the mean squared error function. By backpropagating through the loss function, the n features with large weights on the prediction result are obtained and used to update the training parameters of the graph attention long short-term memory network model.

6. The industrial quality prediction method based on graph attention long short-term memory network as described in claim 1, characterized in that, The dataset includes manufacturing process data for temperature, pressure, temperature × pressure, material fusion index, and material conversion index.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

Images