Quality detection method based on combination of time sequence feature extraction module and graph neural network

Abstract

The invention discloses a quality detection method by combining a time sequence feature extraction module with a graph neural network, and relates to the field of data acquisition and analysis, and the method comprises the steps: 1, remodeling a one-dimensional time sequence into a group of two-dimensional tensors based on a plurality of cycles, processing the two-dimensional tensors through an Inception module, converting the learned two-dimensional representation into a one-dimensional space, and carrying out the processing of the two-dimensional tensors through an Inception module; and finally, aggregating different one-dimensional representations to obtain one-dimensional time sequence data. Step 2, inputting the one-dimensional time sequence data into the graph structure, regarding each node of the graph structure as a sensor, introducing a randomly initialized sensor code for each sensor, and then performing graph structure learning to obtain the weight of the edge of the graph structure; step 3, fusing one-dimensional time sequence data and sensor codes by introducing a feature extractor, calculating an attention coefficient between adjacent nodes, performing normalization processing on the attention coefficient, and diffusing and updating node states; and step 4, classifying the quality grades of the products.

Description

Technical Field

[0001] The present invention relates to the field of data collection and analysis, and in particular to a quality detection method combining a time series feature extraction module with a graph neural network. Background Art

[0002] In terms of quality inspection in the field of industrial production, the main methods currently used are to rely on manual inspection of products and measurement by measuring equipment. Traditional manual inspection and equipment measurement have some problems in the quality inspection process. First of all, they require a lot of manpower, material resources and time, especially for products that need to check the quality inside a closed space, manual inspection means destroying the product. Moreover, with the large-scale and complex industrial production systems, quality inspection by manually extracting features is becoming more and more difficult. In addition, the existing quality inspection methods are often carried out after the product is produced, which is essentially picking out waste products from the finished products. This method not only wastes human resources, but also increases production costs. In order to overcome these problems, statistical and artificial intelligence methods have been introduced into quality control. Through these methods, unqualified products that are about to be formed or have just begun to be formed can be discovered in time. By using automatic monitoring of sensor data and taking timely measures, early warning can be given to avoid the production of unqualified products and effectively save costs.

[0003] However, the real industrial production process has a clear production process and continuity. If the data collected by sensors is used to detect the quality of the products produced, the difficulty lies in the large amount of time series data generated by a large number of interconnected sensors. For example, in a water treatment plant, there may be many sensors measuring water level, flow, water quality, valve status, etc. There are complex relationships between these components, so the correlation between sensors must be considered.

[0004] In the Chinese invention patent "A product quality detection method based on deep neural network" (publication number CN 111753876 A), Li Shangyong et al. designed a deep neural network model to perform real-time non-destructive detection of industrial products based on equipment sensor data, such as temperature and humidity, vibration, noise, current, voltage, etc., and reflected the operating status of the machine to provide a reference for quality inspectors. However, the deep neural network structure used does not have the ability to explicitly capture the spatiotemporal relationship between sensors, especially the spatial relationship. In the Chinese invention patent "A method for detecting anomalies in industrial processes based on graph neural networks" (publication number CN 117724452 A), Liu Zehao et al. improved the modeling ability of the industrial system state and improved the anomaly detection effect of the industrial system. This method can capture the spatial relationship between sensors, but it does not have the ability to explore the intra-cycle and inter-cycle features of time series data. Moreover, this method is used for anomaly detection and has not been applied to the field of quality inspection.

[0005] In actual production, a large number of interconnected sensors generate a large amount of time series data. These time series data have complex intra-cycle and inter-cycle changes. The data from these sensors can be linked together in a complex and nonlinear way, resulting in difficulties in prediction and fault location. Therefore, the spatiotemporal correlation between sensors must be considered.

[0006] Therefore, technicians in this field are committed to developing a new quality inspection method that can better apply a large amount of time series data generated by a large number of interconnected sensors in actual production to solve the above-mentioned defects in the prior art. Summary of the invention

[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to overcome the problem that the deep neural network model cannot explicitly model the complex spatiotemporal relationship between sensors, thereby leading to difficulty in prediction, difficulty in fault location, low quality detection accuracy and poor reliability.

[0008] To achieve the above objectives, the present invention provides a quality inspection method combining a time series feature extraction module with a graph neural network, which has the ability to capture complex spatiotemporal correlations between sensors. It is particularly suitable for product quality inspection in complex production processes where a large number of sensors are connected in a complex and nonlinear way and generate a large amount of time series data, and can bring higher detection accuracy and higher reliability.

[0009] The present invention provides a quality detection method for a time series feature extraction module combined with a graph neural network, comprising the following steps: Step 1: Reshape the one-dimensional time series into a set of two-dimensional tensors based on multiple cycles through the time series feature extraction module to capture the intra-cycle and inter-cycle changes in the two-dimensional space, and then use the Inception module to process the two-dimensional tensor, convert the learned two-dimensional representation to one-dimensional space, and finally aggregate different one-dimensional representations to obtain the one-dimensional time series data of the intra-cycle and inter-cycle features that have been captured; Step 2: In a graph structure learning module, the one-dimensional time series data is input into a graph structure, and each node of the graph structure is regarded as a sensor, a randomly initialized sensor code is introduced for each sensor, and then graph structure learning is performed to obtain the weight of the edge of the graph structure, wherein the edge of the graph structure represents the degree of dependence between the sensors; Step 3: In the graph state update module, by introducing a feature extractor, the one-dimensional time series data and the sensor code are fused, the attention coefficient between adjacent nodes is obtained by using nonlinear activation calculation, and the attention coefficient is normalized. Then, according to the obtained attention coefficient, a multi-layer graph neural network is used to diffuse the node state and update the state of the graph structure, so that each node is fused with the information aggregation representation of the node and its neighbors, and the graph structure is explicitly represented in a higher-dimensional Euclidean space; Step 4: In the product quality classification module, the diffused and updated graph structure is input into a fully connected layer, and the fully connected layer outputs a vector for classifying the quality level of the product.

[0010] Furthermore, the step 1 includes the following sub-steps: Step 1.1: The one-dimensional time series is a time series with a length of T and C sensor variables. The original one-dimensional vector form is , the one-dimensional time series is analyzed in the frequency domain by Fast Fourier Transform (FFT) as follows: Among them, FFT(⋅) and Amp(⋅) represent the calculation of FFT and amplitude values; represents the calculated amplitude of each frequency, averaged over C dimensions by Avg(⋅); the jth value Represents the intensity of the periodic basis function of frequency j, and the corresponding period length is ; Select only the first k amplitude values ​​and get the most significant frequencies and its unnormalized amplitude , where k is a hyperparameter and the selected frequency also corresponds to the cycle length ; Due to the conjugation of the frequency domain, only If the frequency is within the range, the one-dimensional time series can be expressed as: Step 1.2: Based on the selected frequency and the corresponding period length , the one-dimensional time series is converted into a plurality of the two-dimensional tensors by the following equation: Among them, Padding (⋅) is to expand the time series with zeros in the time dimension; and Respectively represent the number of rows and columns of the converted two-dimensional tensor; Frequency-based The i-th reshaped time series, whose columns and rows represent the corresponding period lengths Based on the selected frequency and the estimated period, a set of the two-dimensional tensors are obtained. , representing k different time-two-dimensional changes derived from different periods.

[0011] Furthermore, the step 1 further comprises the following sub-steps: Step 1.3: Based on cycle length , convert the one-dimensional time series to two-dimensional space and select a two-dimensional module: in, is the two-dimensional tensor after the j-th transformation; After conversion, the two-dimensional tensor is processed by the Inception module, and then the learned two-dimensional table is converted back to one-dimensional space for aggregation, and Trunc(⋅) is used to truncate the filling sequence, and finally calculated .

[0012] Furthermore, in step 1.3, the Inception module includes a multi-scale two-dimensional convolution kernel, which can simultaneously aggregate multi-scale intra-cycle changes (columns) and inter-cycle changes (rows), covering both adjacent time points and adjacent cycles.

[0013] Furthermore, the step 1 further comprises the following sub-steps: Step 1.4: Fusion of k different one-dimensional representations Since the calculated amplitude A reflects the relative importance of the selected frequency and period, and thus the importance of the two-dimensional tensor corresponding to each transformation, the one-dimensional time series data based on amplitude aggregation is expressed as: .

[0014] Furthermore, the step 2 includes the following sub-steps: Step 2.1: Input the one-dimensional time series data into the graph structure, and regard each node of the graph structure as a sensor, capturing the features of each sensor, i.e., embedding vectors: Wherein, the embedding vector is a representation of the sensor encoding and is randomly initialized; The prior information can be expressed as a set of candidate relationships for each sensor i , that is, each sensor Possible sensors to rely on: ; Step 2.2: Perform graph structure learning and calculate sensor The embedding vector Relationship with Candidate The normalized dot product between Then select the first k normalized dot products as the degree of association between sensor i and sensor j: in, Represents the exponential of the top-k values ​​in the input, i.e., the normalized dot product; the value of K can be chosen by the user based on the desired level of sparsity.

[0015] Furthermore, in step 3, LeakyRelu is used as a nonlinear activation to calculate the attention coefficient between adjacent nodes, and the attention coefficient is normalized using the Softmax function.

[0016] Furthermore, the step 3 includes the following sub-steps: Step 3.1: Define a graph attention-based model, and use the output of the temporal feature extraction module in step 1 , denoted as , introduce a feature extractor based on the graph attention, and the embedding vector Perform fusion, learn the graph structure and fuse the node information and neighbors: in, is a trainable weight matrix that applies a shared linear transformation to each node; Step 3.2: Calculate the attention coefficient between nodes ; is the vector of learning coefficients of the attention mechanism; The attention coefficient Calculate according to the following formula: Step 3.3: Calculate the aggregate representation of node i : And get the aggregate representation of all N nodes, recorded as: .

[0017] Furthermore, step 4 includes the following sub-steps: Step 4.1: With the corresponding embedding vector The time series of is multiplied element by element (denoted as ), and use the results on all nodes as the input of the fully connected layer with output dimension N. The output of the fully connected layer is a vector : Step 4.2: Vector After entering the Softmax layer, we get the vector : vector It is a k-dimensional vector, each component represents a quality level, the values ​​are between 0 and 1, and the sum is 1; the product quality is judged according to the quality level represented by the vector with the largest value.

[0018] Furthermore, using vector With observational data The mean square error between is used as the loss function to minimize the model training: .

[0019] The quality detection method of a time series feature extraction module combined with a graph neural network provided by the present invention has at least the following technical effects: The technical solution provided by the present invention introduces a two-dimensional time series feature extraction module to reshape the one-dimensional time series into a two-dimensional tensor, in which each column contains a time point within a cycle, and each row involves a time point in the same stage between different cycles. Therefore, by converting the one-dimensional time series into a set of two-dimensional tensors, the intra-cycle and inter-cycle changes in the two-dimensional space are unified, and the intra-cycle and inter-cycle changes in the two-dimensional space are further captured by the two-dimensional convolution kernel. The graph structure is naturally similar to the structure of the sensor network. Its vertices can be used to simulate each sensor, and the edges of the graph structure represent the relationship between different sensors. The graph neural network aims to learn the relationship between sensors in the form of a graph, and then identify and explain the deviation from the learned pattern, so as to capture the spatial correlation. The technical solution provided by the present invention has the ability to capture complex spatiotemporal correlations between sensors, is suitable for product quality detection of complex production processes with a large amount of time series data, and can bring higher detection accuracy and higher reliability.

[0020] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is an overall workflow diagram of a quality inspection method of a preferred embodiment of the present invention; Figure 2 yes Figure 1 The working flow chart of the temporal feature extraction module of the illustrated embodiment; Figure 3 yes Figure 1 Workflow diagram of the graph structure learning, graph state updating and product quality classification modules of the illustrated embodiment. DETAILED DESCRIPTION

[0022] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0023] The embodiment of the present invention proposes a product quality detection method combining a time series feature extraction module with a graph neural network, which can simultaneously capture the spatiotemporal correlation between different sensors. Therefore, it is more suitable for characterization learning than the prior art, captures the complex spatiotemporal relationship between different sensors in actual production, and can obtain more accurate quality detection results. Specifically, first, through the time series feature extraction module, the one-dimensional time series is reshaped into a two-dimensional tensor, captured as intra-cycle and inter-cycle changes in two-dimensional space, and then these data are used as inputs of graph structure node features, graph structure learning is performed, the weights of graph structure edges are obtained, and the graph structure constructed based on the above steps is updated based on the attention mechanism. Finally, through a fully connected layer, a vector is output, which can be used to classify the quality level of the product. Among them, the time series feature extraction module enables the model to capture complex intra-cycle and inter-cycle changes, and model the time correlation of the detection data. The attention mechanism of the graph neural network can be used to capture the correlation between different sensors, so it can be used for product quality detection.

[0024] Example 1 like Figure 1 As shown, an embodiment of the present invention provides a quality detection method of a time series feature extraction module combined with a graph neural network, comprising the following steps: Step 1. Each stage of the time series contains intra-cycle changes and inter-cycle changes. The former refers to the short-term time pattern within a period of time. The latter can reflect the long-term trend of different consecutive periods. In order to solve the limitations of the one-dimensional time series in terms of representation ability, the one-dimensional time series is reshaped into a set of two-dimensional tensors based on multiple cycles through the time series feature extraction module to capture the intra-cycle and inter-cycle changes in the two-dimensional space, where each column contains a time point within a cycle, and each row involves the time point of the same stage between different cycles, thereby extending the analysis of time changes to two-dimensional space, making it easy to capture two-dimensional changes with a two-dimensional kernel. The two-dimensional tensor is then processed using the Inception module to convert the learned two-dimensional representation into a one-dimensional space. Finally, different one-dimensional representations are aggregated to obtain the one-dimensional time series data that has captured the intra-cycle and inter-cycle features; Step 2: In the graph structure learning module, one-dimensional time series data is input into the graph structure, and each node of the graph structure is regarded as a sensor. To represent the different features of each sensor, a randomly initialized sensor code is first introduced for each sensor, and then trained together with the rest of the model to perform graph structure learning. The edges of the graph represent the dependencies between sensors. The cosine similarity between these encoding vectors is used to assign different weights to the edges in the graph, that is, the degree of correlation between two different sensors. The graph structure is constructed based on the relationship matrix that expresses the weights of different nodes. The learned graph structure affects the information transmission and aggregation of connected nodes, and is used to construct the system state structure and update the system state. Step 3: In the graph state update module, the one-dimensional time series data obtained by the time series feature extraction module is used as the input of the graph structure node state learned by the graph structure module. By introducing the feature extractor, the one-dimensional time series data and the sensor encoding are fused, and LeakyRelu is used as the nonlinear activation to calculate the attention coefficient between adjacent nodes. The attention coefficient is normalized by the Softmax function. Then, based on the obtained attention coefficient, the multi-layer graph neural network is used to diffuse the node state and update the state of the graph structure, so that each node fuses the information aggregation representation of the node and its neighbors, explicitly represents the graph structure in a higher-dimensional Euclidean space, and captures the complex spatial correlation between sensors. Step 4: In the product quality classification module, the diffused and updated graph structure is input into the fully connected layer, and the fully connected layer outputs a vector to classify the quality level of the product.

[0025] Example 2 Based on Example 1, step 1 includes the following sub-steps: Figure 2 As shown: Step 1.1: A one-dimensional time series is a time series of length T with C sensor variables. The original one-dimensional vector form is , a one-dimensional time series is analyzed in the frequency domain by means of Fast Fourier Transform (FFT) as follows: Among them, FFT(⋅) and Amp(⋅) represent the calculation of FFT and amplitude values; represents the calculated amplitude of each frequency, averaged over C dimensions by Avg(⋅); the jth value Represents the intensity of the periodic basis function of frequency j, and the corresponding period length is ; Select only the first k amplitude values ​​and get the most significant frequencies and its unnormalized amplitude , where k is a hyperparameter and the selected frequency also corresponds to the cycle length ; Due to the conjugation of the frequency domain, only If the frequency is within the range, the one-dimensional time series can be expressed as: Step 1.2: Based on the selected frequency and the corresponding period length , converting a one-dimensional time series into multiple two-dimensional tensors by the following equation: Among them, Padding (⋅) is to expand the time series with zeros in the time dimension; and Respectively represent the number of rows and columns of the converted two-dimensional tensor; Frequency-based The i-th reshaped time series, whose columns and rows represent the corresponding period lengths Based on the selected frequency and estimated period, a set of two-dimensional tensors is obtained. , representing k different time-two-dimensional changes derived from different periods.

[0026] Step 1.3: Based on cycle length , convert a one-dimensional time series to two-dimensional space and select a two-dimensional module: in, is the two-dimensional tensor after the jth transformation; After the conversion, the two-dimensional tensor is processed by the Inception module, and then the learned two-dimensional table is converted back to one-dimensional space for aggregation, and Trunc(⋅) is used to truncate the filling sequence, and finally the calculation is obtained. .

[0027] In particular, the Inception module includes multi-scale two-dimensional convolution kernels that can simultaneously aggregate multi-scale intra-cycle changes (columns) and inter-cycle changes (rows), covering both adjacent time points and adjacent cycles.

[0028] Step 1.4: Fusion of k different one-dimensional representations ; Since the calculated amplitude A reflects the relative importance of the selected frequency and period, and thus corresponds to the importance of each transformed two-dimensional tensor, the one-dimensional time series data based on amplitude aggregation is expressed as: .

[0029] Example 3 Based on Example 2, step 2 includes the following sub-steps: Figure 3 As shown: Step 2.1: Input the one-dimensional time series data into the graph structure, and regard each node of the graph structure as a sensor, capturing the features of each sensor, i.e., the embedding vector: Among them, the embedding vector is the representation of the sensor code and is randomly initialized; The prior information can be expressed as a set of candidate relations for each sensor i , that is, each sensor Possible sensors to rely on: ; Step 2.2: Perform graph structure learning and calculate sensor The embedding vector Relationship with Candidate The normalized dot product between Then select the first k normalized dot products as the degree of association between sensor i and sensor j: in, Represents the exponential of the top-k values ​​in the input, i.e., the normalized dot product; the value of K can be chosen by the user based on the desired level of sparsity.

[0030] Example 4 Based on Example 3, step 3 includes the following sub-steps: Figure 3 As shown: Step 3.1: Define a graph attention-based model and transform the output of the temporal feature extraction module in step 1 into , denoted as , introduce a graph attention-based feature extractor, and embedding vector Perform fusion, learn the graph structure and fuse the information and neighbors of the nodes: in, is a trainable weight matrix that applies a shared linear transformation to each node; Step 3.2: Calculate the attention coefficient between nodes ; is the vector of learning coefficients of the attention mechanism; Attention coefficient Calculate according to the following formula: Step 3.3: Calculate the aggregate representation of node i : And get the aggregate representation of all N nodes, recorded as: .

[0031] Example 5 Based on Example 4, step 4 includes the following sub-steps: Figure 3 As shown: Step 4.1: With the corresponding embedding vector The time series of is multiplied element by element (denoted as ), and use the results on all nodes as the input of the fully connected layer with output dimension N. The output of the fully connected layer is a vector : Step 4.2: Vector After entering the Softmax layer, we get the vector : vector It is a k-dimensional vector, each component represents a quality level, the values ​​are between 0 and 1, and the sum is 1; the product quality is judged according to the quality level represented by the vector with the largest value.

[0032] In particular, using the vector With observational data The mean square error between is used as the loss function to minimize the model training: .

[0033] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A quality detection method combining a time series feature extraction module with a graph neural network, characterized in that: The method comprises the following steps: Step 1: Reshape the one-dimensional time series into a set of two-dimensional tensors based on multiple cycles through the time series feature extraction module to capture the intra-cycle and inter-cycle changes in the two-dimensional space, and then use the Inception module to process the two-dimensional tensor, convert the learned two-dimensional representation to one-dimensional space, and finally aggregate different one-dimensional representations to obtain the one-dimensional time series data of the intra-cycle and inter-cycle features that have been captured; Step 2: In a graph structure learning module, the one-dimensional time series data is input into a graph structure, and each node of the graph structure is regarded as a sensor, a randomly initialized sensor code is introduced for each sensor, and then graph structure learning is performed to obtain the weight of the edge of the graph structure, wherein the edge of the graph structure represents the degree of dependence between the sensors; Step 3: In the graph state update module, by introducing a feature extractor, the one-dimensional time series data and the sensor code are fused, the attention coefficient between adjacent nodes is obtained by using nonlinear activation calculation, and the attention coefficient is normalized. Then, according to the obtained attention coefficient, a multi-layer graph neural network is used to diffuse the node state and update the state of the graph structure, so that each node is fused with the information aggregation representation of the node and its neighbors, and the graph structure is explicitly represented in a higher-dimensional Euclidean space; Step 4: In the product quality classification module, the diffused and updated graph structure is input into a fully connected layer, and the fully connected layer outputs a vector for classifying the quality level of the product.

2. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 1 is characterized in that: The step 1 comprises the following sub-steps: Step 1.1: The one-dimensional time series is a time series with C sensor variables and a length of T. The original one-dimensional vector form is , the one-dimensional time series is analyzed in the frequency domain by Fast Fourier Transform (FFT) as follows: Among them, FFT(⋅) and Amp(⋅) represent the calculation of FFT and amplitude values; represents the calculated amplitude of each frequency, averaged over C dimensions by Avg(⋅); the jth value Represents the intensity of the periodic basis function of frequency j, and the corresponding period length is ; Select only the first k amplitude values ​​and get the most significant frequencies and its unnormalized amplitude , where k is a hyperparameter and the selected frequency also corresponds to the cycle length ; Due to the conjugation of the frequency domain, only If the frequency is within the range, the one-dimensional time series can be expressed as: Step 1.2: Based on the selected frequency and the corresponding period length , the one-dimensional time series is converted into a plurality of the two-dimensional tensors by the following equation: Among them, Padding (⋅) is to expand the time series with zeros in the time dimension; and Respectively represent the number of rows and columns of the converted two-dimensional tensor; Frequency-based The i-th reshaped time series, whose columns and rows represent the corresponding period lengths Based on the selected frequency and the estimated period, a set of the two-dimensional tensors are obtained. , representing k different time-two-dimensional changes derived from different periods.

3. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 2 is characterized in that: The step 1 also includes the following sub-steps: Step 1.3: Based on cycle length , convert the one-dimensional time series to two-dimensional space and select a two-dimensional module: in, is the two-dimensional tensor after the j-th transformation; After conversion, the two-dimensional tensor is processed by the Inception module, and then the learned two-dimensional table is converted back to one-dimensional space for aggregation, and Trunc(⋅) is used to truncate the filling sequence, and finally calculated .

4. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 3 is characterized in that: In step 1.3, the Inception module includes a multi-scale two-dimensional convolution kernel, which can simultaneously aggregate multi-scale intra-cycle changes (columns) and inter-cycle changes (rows), covering both adjacent time points and adjacent cycles.

5. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 3 is characterized in that: The step 1 also includes the following sub-steps: Step 1.4: Fusion of k different one-dimensional representations Since the calculated amplitude A reflects the relative importance of the selected frequency and period, and thus the importance of the two-dimensional tensor corresponding to each transformation, the one-dimensional time series data based on amplitude aggregation is expressed as: 。 6. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 5 is characterized in that: The step 2 includes the following sub-steps: Step 2.1: Input the one-dimensional time series data into the graph structure, and regard each node of the graph structure as a sensor, capturing the features of each sensor, i.e., embedding vectors: Wherein, the embedding vector is a representation of the sensor encoding and is randomly initialized; The prior information can be expressed as a set of candidate relationships for each sensor i , that is, each sensor Possible sensors to rely on: ; Step 2.2: Perform graph structure learning and calculate sensor The embedding vector Relationship with Candidate The normalized dot product between Then select the first k normalized dot products as the degree of association between sensor i and sensor j: in, Represents the exponential of the top-k values ​​in the input, i.e., the normalized dot product; the value of K can be chosen by the user based on the desired level of sparsity.

7. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 6 is characterized in that: In step 3, LeakyRelu is used as a nonlinear activation to calculate the attention coefficient between adjacent nodes, and the attention coefficient is normalized using the Softmax function.

8. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 7 is characterized in that: The step 3 includes the following sub-steps: Step 3.1: Define a graph attention-based model, and use the output of the temporal feature extraction module in step 1 , denoted as , introduce a feature extractor based on the graph attention, and the embedding vector Perform fusion, learn the graph structure and fuse the node information and neighbors: in, is a trainable weight matrix that applies a shared linear transformation to each node; Step 3.2: Calculate the attention coefficient between nodes ; is the vector of learning coefficients of the attention mechanism; The attention coefficient Calculate according to the following formula: Step 3.3: Calculate the aggregate representation of node i : And get the aggregate representation of all N nodes, recorded as: .

9. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 8 is characterized in that: The step 4 includes the following sub-steps: Step 4.1: With the corresponding embedding vector The time series of is multiplied element by element (denoted as ), and use the results on all nodes as the input of the fully connected layer with output dimension N. The output of the fully connected layer is a vector : Step 4.2: Vector After entering the Softmax layer, we get the vector : vector It is a k-dimensional vector, each component represents a quality level, the values ​​are between 0 and 1, and the sum is 1; the product quality is judged according to the quality level represented by the vector with the largest value.

10. The quality detection method of the time series feature extraction module combined with the graph neural network as claimed in claim 9 is characterized in that: Using Vectors With observational data The mean square error between is used as the loss function to minimize the model training: 。

Images