Automatic test method and system for heterogeneous process verification and system software debugging

Abstract

The invention provides an automatic test method and system for heterogeneous process verification and system software debugging, and relates to the technical field of software debugging, and the method comprises the steps: collecting and producing holographic monitoring data, and carrying out semantic annotation through a preset ontology model; a multi-head attention mechanism and a graph neural network are adopted to train an ontology model, and associated features of process parameters, equipment states and defect types are extracted; constructing an initial knowledge graph based on the features, and calculating a node weight relationship by using causal reasoning to generate a dynamic weight matrix; and in combination with real-time process data, features are extracted and fused through a deep residual network, and defect risk assessment and early warning are realized. Through dynamic updating and feature fusion of the knowledge graph, the accuracy and automation degree of heterogeneous process verification are improved, and intelligent early warning of process quality is realized.

Description

Technical Field

[0001] The present invention relates to software debugging technology, and in particular to an automated testing method and system for heterogeneous process verification and system software debugging. Background Art

[0002] With the deepening digital transformation of the manufacturing industry, heterogeneous processes are gaining widespread application in intelligent manufacturing. Heterogeneous processes involve the coordinated coordination of multiple processing techniques, generating massive amounts of monitoring data during the production process. Currently, holographic monitoring technology is widely used in industrial sites to collect real-time production data, including multi-dimensional information such as process parameters, equipment status, and environmental parameters. This data contains rich process knowledge and laws, which is of great significance for ensuring product quality and improving production efficiency. Traditional process verification and software debugging methods rely primarily on manual judgment and experience, which are difficult to meet the development needs of intelligent manufacturing. With the development of artificial intelligence technology, automated testing methods based on machine learning have gradually become a research hotspot.

[0003] In practical applications, existing heterogeneous process verification and system software debugging methods struggle to effectively handle multi-source heterogeneous data. Due to a lack of semantic understanding of production data, they are unable to accurately extract deep correlations between process parameters, equipment status, and defect types, resulting in low reliability of verification results.

[0004] Existing methods often use static knowledge representation when processing dynamic time-series data, ignoring the time-evolving nature of process knowledge. This makes it difficult for the system to adapt to dynamic changes in process parameters and fail to detect potential quality risks in a timely manner.

[0005] Current automated testing methods lack effective feature fusion mechanisms and fail to fully leverage historical process knowledge to guide real-time monitoring. This results in low early warning accuracy when the system handles complex working conditions, making it prone to missed or false alarms, impacting production efficiency and product quality. Summary of the Invention

[0006] The embodiments of the present invention provide an automated testing method and system for heterogeneous process verification and system software debugging, which can solve the problems in the prior art.

[0007] A first aspect of an embodiment of the present invention provides an automated testing method for heterogeneous process verification and system software debugging, comprising: Collect production holographic monitoring datasets from heterogeneous production processes; use a preset ontology model to semantically annotate the production holographic monitoring dataset to generate an annotated dataset; based on the annotated dataset, use a multi-head attention mechanism and graph neural network to train the preset ontology model to extract correlation features between process parameters, equipment status, and defect types; Call the trained preset ontology model to perform feature extraction and association analysis on the production holographic monitoring data set; construct an initial knowledge graph based on the extracted features; use the causal reasoning mechanism to calculate the weight relationship between nodes and generate a dynamic weight matrix; dynamically update the initial knowledge graph based on the dynamic weight matrix and output a process knowledge vector containing time-series evolution characteristics; Receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network; fuse the real-time features with the process knowledge vectors and input them into the prediction model; output defect risk assessment results based on the prediction model; when the defect risk assessment results exceed the dynamic threshold, generate early warning information.

[0008] The production holographic monitoring dataset is semantically annotated using a preset ontology model to generate an annotated dataset. Based on the annotated dataset, a multi-head attention mechanism and graph neural network are used to train the preset ontology model to extract the correlation features between process parameters, equipment status, and defect types, including: The data in the production holographic monitoring dataset is semantically annotated using a preset ontology model to generate an annotated dataset. A multi-head attention mechanism is trained based on the annotated dataset, and the association weights of the annotated dataset are calculated using the multi-head attention mechanism to generate an association weight matrix. The association weight matrix is ​​input into a graph neural network, and the graph neural network is used to aggregate information on the annotated dataset, extract local structural features of the annotated dataset, and generate node feature vectors. The node feature vector is input into the improved ontology model training module, and the node representation in the preset ontology model is updated based on the node feature vector. The preset ontology model is optimized and trained using a loss function to obtain a trained ontology model. The trained ontology model is used to extract the correlation features between process parameters, equipment status and defect types.

[0009] The node feature vector is input into the improved ontology model training module. The node representation in the preset ontology model is updated based on the node feature vector. The preset ontology model is optimized and trained using a loss function to obtain a trained ontology model. The trained ontology model is used to extract the correlation features between process parameters, equipment status, and defect types, including: The node feature vectors are input into the improved ontology model training module, which uses a message passing mechanism to iteratively update the node feature vectors over multiple rounds. In each round, feature aggregation is performed based on attention weights. A bidirectional gated recurrent unit is used to capture the temporal dependencies of the node feature vectors, and residual connections are used to maintain the integrity of the node feature vectors. The node representation in the preset ontology model is updated based on the updated node feature vector; a multi-task loss function is constructed, including node feature reconstruction loss and structure preservation loss, and the preset ontology model is optimized and trained using the multi-task loss function to obtain the trained ontology model; the trained ontology model is used to extract the correlation features between process parameters, equipment status and defect types.

[0010] The trained preset ontology model is called to perform feature extraction and association analysis on the production holographic monitoring data set; an initial knowledge graph is constructed based on the extracted features; the weight relationship between nodes is calculated using the causal reasoning mechanism to generate a dynamic weight matrix; the initial knowledge graph is dynamically updated based on the dynamic weight matrix, and the output process knowledge vector containing time-series evolution features includes: Call the trained preset ontology model to perform feature extraction and correlation analysis on the production holographic monitoring data set to obtain multi-dimensional feature vectors and correlation matrices; construct the initial knowledge graph based on the multi-dimensional feature vectors and correlation matrix; The causal reasoning mechanism is used to calculate the weight relationship between nodes in the initial knowledge graph. The direct causal links between nodes are identified through the conditional temporal causal analysis method. The mutual information entropy is used to evaluate the degree of association between nodes. A dynamic weight matrix is ​​generated based on the direct causal links and the degree of association. The initial knowledge graph is dynamically updated based on the dynamic weight matrix, the connection relationship between nodes is adjusted according to the weight size, the connections with weights lower than the preset weight threshold are deleted, and the connections with significant causal relationships are retained; the node attribute features, topological structure features and temporal evolution features are extracted from the updated knowledge graph, and the node attribute features, topological result features and temporal evolution features are fused to generate a process knowledge vector.

[0011] The initial knowledge graph is dynamically updated based on the dynamic weight matrix. The connection relationship between nodes is adjusted according to the weight size, and the connection with weight lower than the preset weight threshold is deleted, while the connection with significant causal relationship is retained. The node attribute features, topological structure features and temporal evolution features are extracted from the updated knowledge graph, and the node attribute features, topological result features and temporal evolution features are fused to generate a process knowledge vector including: Based on the distribution law of weight values ​​in the dynamic weight matrix, the weight mean and standard deviation are weighted and combined to obtain the preset weight threshold; the weight values ​​in the dynamic weight matrix are compared with the preset weight threshold, and the node connections with weight values ​​higher than the preset weight threshold and significant causal relationships are retained, and the node connections with weight values ​​lower than the preset weight threshold are deleted to obtain the optimized knowledge graph; For the optimized knowledge graph, the parameter type, value range, and adjustment method of the node are extracted in combination with the process specifications to form the node attribute characteristics; based on the connection relationship of the node, the degree distribution, path characteristics, and hierarchical structure are extracted to form the topological structure characteristics; and the historical data of the node are used to extract the change trend and periodic law to form the time series evolution characteristics. The node attribute features, topological structure features and temporal evolution features are weighted and combined into a dynamic graph holographic feature spectrum, which is then input into the feature importance calculation module to obtain the influence weight of each feature on the process. The three types of features are adaptively fused according to the influence weight to generate a feature vector reflecting process knowledge.

[0012] Receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network. Fusion the real-time features with the process knowledge vectors and input them into the prediction model. Output defect risk assessment results based on the prediction model. When the defect risk assessment results exceed the dynamic threshold, generate warning information including: Receive process knowledge vectors and collect heterogeneous process data in real time, perform time series alignment and normalization on the heterogeneous process data, and generate a standardized process data matrix. The standardized process data matrix is ​​input into a preset deep residual network. The deep residual network contains multiple cascaded feature extraction modules and residual connection modules. The feature extraction module extracts multi-level process features. The residual connection module is used to maintain the original information, fuse the original information with the multi-level process features, and output a real-time feature matrix. A feature fusion network is used to fuse the real-time feature matrix and the process knowledge vector. The feature fusion network adaptively calculates the correlation weights of the two types of features in the spatial dimension and channel dimension through the attention calculation unit, and uses the feature aggregation unit to perform a weighted combination of the two types of features based on the correlation weights to generate a fused feature vector. The fused feature vector is input into the prediction model for risk assessment. The time series feature extraction unit of the prediction model extracts the time series change features. The time series change features are input into the risk analysis unit to calculate the probability of occurrence of different types of defects and generate defect risk assessment results. Based on historical risk distribution and current operating parameters, the risk warning threshold is updated in real time through the threshold calculation unit of the dynamic threshold warning module; when the defect risk assessment result exceeds the risk warning threshold, the threshold calculation unit is used to generate warning information containing defect type, risk level and disposal suggestions.

[0013] A second aspect of an embodiment of the present invention provides an automated testing system for heterogeneous process verification and system software debugging, including: The first unit is used to collect a production holographic monitoring dataset from a heterogeneous production process; semantically annotate the production holographic monitoring dataset using a preset ontology model to generate an annotated dataset; based on the annotated dataset, a multi-head attention mechanism and a graph neural network are used to train the preset ontology model to extract correlation features between process parameters, equipment status, and defect types; The second unit is used to call the trained preset ontology model to perform feature extraction and association analysis on the production holographic monitoring data set; construct an initial knowledge graph based on the extracted features; use the causal reasoning mechanism to calculate the weight relationship between nodes and generate a dynamic weight matrix; dynamically update the initial knowledge graph based on the dynamic weight matrix and output a process knowledge vector containing time-series evolution characteristics; The third unit is used to receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network; fuse the real-time features with the process knowledge vectors and input them into the prediction model; output the defect risk assessment results based on the prediction model; and generate early warning information when the defect risk assessment results exceed the dynamic threshold.

[0014] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0016] The beneficial effects of this application are as follows: 1. By establishing a semantically annotated ontology model and combining it with a multi-head attention mechanism and graph neural network, we can achieve in-depth mining and feature extraction of the complex correlation between process parameters, equipment status and defect types, greatly improving the analysis accuracy and efficiency of process data.

[0017] 2. Based on the dynamically updated knowledge graph technology, the causal reasoning mechanism is introduced to calculate the weight relationship, which can accurately grasp the temporal evolution law of process parameters, realize the dynamic representation and precise modeling of process knowledge, and significantly enhance the system's adaptability to process changes.

[0018] 3. By using a deep residual network to extract real-time features and integrating them with process knowledge vectors, real-time monitoring and risk warning of the production process are achieved, which can timely detect and prevent potential process defects and effectively ensure the stability and consistency of product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the process flow of the automated testing method for heterogeneous process verification and system software debugging according to an embodiment of the present invention; Figure 2 A bar chart comparing the performance of the improved ontology model in different scenarios according to the embodiment of the present invention; Figure 3 This is a logic block diagram of knowledge graph update and feature fusion based on dynamic weight matrix in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0021] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0022] Figure 1 FIG. 1 is a flow chart of an automated testing method for heterogeneous process verification and system software debugging according to an embodiment of the present invention. Figure 1 As shown, the method includes: Collect production holographic monitoring datasets from heterogeneous production processes; use a preset ontology model to semantically annotate the production holographic monitoring dataset to generate an annotated dataset; based on the annotated dataset, use a multi-head attention mechanism and graph neural network to train the preset ontology model to extract correlation features between process parameters, equipment status, and defect types; Call the trained preset ontology model to perform feature extraction and association analysis on the production holographic monitoring data set; construct an initial knowledge graph based on the extracted features; use the causal reasoning mechanism to calculate the weight relationship between nodes and generate a dynamic weight matrix; dynamically update the initial knowledge graph based on the dynamic weight matrix and output a process knowledge vector containing time-series evolution characteristics; Receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network; fuse the real-time features with the process knowledge vectors and input them into the prediction model; output defect risk assessment results based on the prediction model; when the defect risk assessment results exceed the dynamic threshold, generate early warning information.

[0023] In an optional embodiment, a preset ontology model is used to semantically annotate the production holographic monitoring dataset to generate an annotated dataset. Based on the annotated dataset, a multi-head attention mechanism and a graph neural network are used to train the preset ontology model to extract correlation features between process parameters, equipment status, and defect types, including: The data in the production holographic monitoring dataset is semantically annotated using a preset ontology model to generate an annotated dataset. A multi-head attention mechanism is trained based on the annotated dataset, and the association weights of the annotated dataset are calculated using the multi-head attention mechanism to generate an association weight matrix. The association weight matrix is ​​input into a graph neural network, and the graph neural network is used to aggregate information on the annotated dataset, extract local structural features of the annotated dataset, and generate node feature vectors. The node feature vector is input into the improved ontology model training module, and the node representation in the preset ontology model is updated based on the node feature vector. The preset ontology model is optimized and trained using a loss function to obtain a trained ontology model. The trained ontology model is used to extract the correlation features between process parameters, equipment status and defect types.

[0024] A pre-defined ontology model is constructed, consisting of process parameter nodes, equipment status nodes, and defect type nodes. The process parameter nodes store process parameter information during the production process, such as temperature, pressure, and speed; the equipment status nodes store equipment operating status information, such as vibration, noise, and temperature rise; and the defect type nodes store product defect information, such as cracks, deformation, and wrinkles. These nodes are connected by relationship edges, representing the association between them.

[0025] When semantically annotating a production holographic monitoring dataset, the first step is to read the raw data from the dataset. Taking the steel production process as an example, the raw data includes process parameters such as a rolling temperature of 950°C, a rolling pressure of 80 MPa, and a rolling speed of 5 m / s; equipment status information such as a vibration amplitude of 0.5 mm, a noise level of 85 dB, and a temperature rise of 15°C; and defect information such as cracks and deformations. Based on a pre-set ontology model, this raw data is mapped to corresponding nodes to generate annotated data. For example, a rolling temperature of 950°C is annotated to the temperature node, a vibration amplitude of 0.5 mm is annotated to the vibration node, and cracks are annotated to the defect type node.

[0026] A multi-head attention mechanism is trained based on labeled data. This mechanism consists of multiple attention heads, each of which independently calculates the degree of association between nodes. For temperature and crack nodes, an attention head calculates the similarity of the node representation vectors to obtain association weights, with larger weights indicating a higher degree of association. The calculation results of all attention heads are combined to obtain a complete association weight matrix.

[0027] The association weight matrix is ​​fed into a graph neural network, which aggregates features based on the node's neighbor information. For example, for a temperature node, information from its connected neighbors, such as pressure and velocity nodes, is aggregated to extract local structural features. By stacking multiple layers of graph neural networks, a node feature vector containing rich semantic information is ultimately generated.

[0028] The node feature vectors are fed into the improved ontology model training module. This module updates the node representations in the ontology model based on the node feature vectors. A distance metric is used as the loss function to calculate the difference between the updated node representations and the annotated data. Backpropagation is used to optimize the ontology model parameters, enabling the model to more accurately represent the relationships between nodes.

[0029] The trained ontology model is used to extract relevant features. New production data is first mapped to nodes in the ontology model. The model then extracts relevant features between nodes related to process parameters, equipment status, and defect types, enabling analysis of anomalies in the production process. For example, if excessive temperature causes increased equipment vibration, which in turn causes product cracks, the model can extract this relevant feature.

[0030] The solution of this application can: By semantically annotating data using a pre-set ontology model, unstructured production data is transformed into structured knowledge representation, improving data comprehensibility and usability. A multi-head attention mechanism is used to calculate the association weights between nodes, enabling the capture of complex relationships between nodes from multiple perspectives. A graph neural network is used to aggregate information from the annotated data, fully considering the local structural characteristics of the nodes and improving the accuracy of feature extraction. The model is optimized using an improved ontology model training module, making it more adaptable to actual production scenarios. Based on the trained ontology model, correlation features between process parameters, equipment status, and defect types can be automatically extracted, facilitating the timely detection of production anomalies and preventing equipment failures and product defects. This method has strong generalization capabilities and can be applied to different production scenarios.

[0031] In an optional embodiment, the node feature vector is input into the improved ontology model training module, the node representation in the preset ontology model is updated based on the node feature vector, and the preset ontology model is optimized and trained using a loss function to obtain a trained ontology model; and the correlation features between process parameters, equipment status, and defect types are extracted using the trained ontology model, including: The node feature vectors are input into the improved ontology model training module, which uses a message passing mechanism to iteratively update the node feature vectors over multiple rounds. In each round, feature aggregation is performed based on attention weights. A bidirectional gated recurrent unit is used to capture the temporal dependencies of the node feature vectors, and residual connections are used to maintain the integrity of the node feature vectors. The node representation in the preset ontology model is updated based on the updated node feature vector; a multi-task loss function is constructed, including node feature reconstruction loss and structure preservation loss, and the preset ontology model is optimized and trained using the multi-task loss function to obtain the trained ontology model; the trained ontology model is used to extract the correlation features between process parameters, equipment status and defect types.

[0032] The improved ontology model is trained based on node feature vectors. Before model training, the input node feature vectors must be preprocessed. Each node's feature vector contains multidimensional feature information, including process parameters, equipment status, and defect type. Feature normalization is used to normalize the eigenvalues ​​of each dimension in the feature vector to the same numerical range, facilitating subsequent model training.

[0033] In the message passing mechanism, each iteration consists of two phases: feature aggregation and feature update. During the feature aggregation phase, feature information of neighboring nodes is collected for each node. An attention mechanism is used to calculate the importance weights of neighboring nodes. The larger the weight, the greater the influence of the neighboring node on the current node. For example, a process parameter node is connected to multiple device status nodes. By calculating the attention weights, the critical device status with the greatest impact on the process parameter can be identified. Specifically, the temperature of the process parameter node can be set to 500 degrees, and the device status nodes connected to it include device A at 490 degrees and device B at 510 degrees. After calculating the attention weights, it is found that the influence weight of device B's status on the process parameter is 0.7, which is higher than the 0.3 of device A, indicating that the status of device B has a greater impact on the process parameter.

[0034] During the feature update phase, aggregated neighbor features are fused with the node's own features to update the node's feature representation. Through multiple rounds of iteration, node features can gradually incorporate higher-order neighbor information. For example, if the number of iterations is set to three, the first round incorporates directly connected neighbor information, the second round incorporates second-order neighbor information, and the third round incorporates third-order neighbor information, allowing the node feature representation to fully capture the structural information of the ontology network.

[0035] To capture the temporal dependencies of node features, a bidirectional gated recurrent unit (BGRU) is used to process node feature sequences. This BGRU consists of two gates: an update gate and a reset gate, which control the retention and forgetting of historical information. By transferring features in both forward and reverse directions, we can simultaneously consider feature dependencies between past and future moments. For example, if the temperature of a device at different times is 490°C, 500°C, and 510°C, bidirectional feature transfer can reveal an upward trend in the device's temperature.

[0036] Incorporating residual connections preserves the original feature information of a node, avoiding information loss caused by multiple iterations. Residual connections sum the original features with the updated features, allowing the model to learn new feature representations while preserving the original attribute information of the node. For example, the original feature of a process parameter node is an eight-dimensional vector. After feature updating, a new eight-dimensional feature vector is generated. Residual connections sum these two feature vectors to obtain the final node representation.

[0037] After obtaining the updated node feature vector, the node representation in the pre-set ontology model needs to be updated. First, a correspondence is established between the nodes in the pre-set ontology model and the updated feature vector, and then the original node representation is replaced with the updated feature vector. This allows the learned feature information to be injected into the pre-set ontology model.

[0038] A multi-task loss function is constructed to optimize model training. The node feature reconstruction loss measures the difference between updated node features and the original features, while the structure preservation loss is used to preserve the relationship structure between nodes in the ontology model. The combined effect of these two loss functions allows for learning new feature representations while maintaining the semantic structure of the ontology model. During training, optimization algorithms such as stochastic gradient descent are used to iteratively update model parameters until the loss function converges.

[0039] The trained ontology model is used to extract correlation features between process parameters, equipment status, and defect types. By analyzing the connectivity and feature similarity between nodes, correlation patterns between different node types can be discovered. For example, a certain defect type has a strong correlation with a specific range of process parameter values ​​and equipment status combinations. This correlation feature can provide guidance for defect prevention and process optimization.

[0040] The solution of this application can: Through improved message passing and attention mechanisms, local and global dependencies between nodes in the ontology network can be effectively captured, improving the accuracy and completeness of feature representation. The attention mechanism can adaptively assign importance weights to different neighboring nodes, highlighting the influence of key nodes. The combined design of bidirectional gated recurrent units and residual connections ensures the modeling of temporal dependencies between node features while avoiding feature loss during deep network training. Bidirectional feature transfer can fully utilize sequence information, while residual connections can maintain the integrity of original features. A multi-task loss function is constructed for model optimization, unifying the two goals of feature reconstruction and structure preservation into a single training framework. This multi-task learning approach can learn more expressive feature representations while maintaining the semantic connections between nodes in the ontology model, improving the model's generalization and robustness.

[0041] Figure 2 The bar chart shows the performance comparison and analysis of the improved ontology model in different scenarios according to the embodiment of the present invention: This figure compares the performance of three different technical solutions in five industrial analysis scenarios. The traditional ontology model (white bars) uses conventional machine learning methods such as decision trees and random forests, relying primarily on manually defined features. The basic improved model (diagonal bars) utilizes a basic deep learning architecture such as a simple CNN or RNN network. The attention-enhanced improved model (dark bars) employs a Transformer-based attention mechanism architecture. In terms of performance, the three models achieved accuracies of 72.5%, 81.2%, and 89.7%, respectively, in process parameter correlation analysis. They also achieved 68.9%, 77.5%, and 86.2%, respectively, in equipment status prediction. They performed particularly well in defect type identification and analysis, reaching 75.3%, 83.6%, and 91.5%, respectively. When processing complex multivariate data fusion tasks, the accuracies were 66.7%, 75.9%, and 87.3%, respectively. In capturing time series features, the three models achieved 62.1%, 76.3%, and 88.6%, respectively. The data clearly shows that the improved model using the attention mechanism significantly outperforms the other two solutions in all application scenarios. In particular, its performance advantage is more obvious when processing complex industrial data, fully demonstrating the important value of the attention mechanism in the field of industrial intelligent analysis.

[0042] In an optional embodiment, a pre-trained ontology model is called to perform feature extraction and association analysis on the production holographic monitoring data set; an initial knowledge graph is constructed based on the extracted features; a causal reasoning mechanism is used to calculate the weight relationship between nodes to generate a dynamic weight matrix; the initial knowledge graph is dynamically updated based on the dynamic weight matrix, and the output process knowledge vector containing time-series evolution features includes: Call the trained preset ontology model to perform feature extraction and correlation analysis on the production holographic monitoring data set to obtain multi-dimensional feature vectors and correlation matrices; construct the initial knowledge graph based on the multi-dimensional feature vectors and correlation matrix; The causal reasoning mechanism is used to calculate the weight relationship between nodes in the initial knowledge graph. The direct causal links between nodes are identified through the conditional temporal causal analysis method. The mutual information entropy is used to evaluate the degree of association between nodes. A dynamic weight matrix is ​​generated based on the direct causal links and the degree of association. The initial knowledge graph is dynamically updated based on the dynamic weight matrix, the connection relationship between nodes is adjusted according to the weight size, the connections with weights lower than the preset weight threshold are deleted, and the connections with significant causal relationships are retained; the node attribute features, topological structure features and temporal evolution features are extracted from the updated knowledge graph, and the node attribute features, topological result features and temporal evolution features are fused to generate a process knowledge vector.

[0043] A pre-trained ontology model is used for feature extraction. This ontology model utilizes a multi-layer neural network structure based on deep learning, consisting of a feature extraction layer, a correlation analysis layer, and a semantic mapping layer. The feature extraction layer uses a convolutional neural network to extract multi-dimensional features from production data, extracting feature vectors covering multiple dimensions, such as equipment parameters, process parameters, and environmental parameters. The correlation analysis layer uses an attention mechanism to calculate the correlation weights between different features and generate a correlation matrix. The semantic mapping layer maps the extracted feature vectors to a predefined ontology concept space. In implementation, the raw production data is first segmented into time windows, each with a length of 1 hour and a sliding step of 10 minutes. For the data within each time window, the ontology model is used to extract feature vectors and calculate the correlation matrix.

[0044] An initial knowledge graph is constructed based on the extracted eigenvectors and the correlation matrix. During construction, eigenvectors are treated as nodes in the graph, and correlations as edges between nodes. Node attributes include the values ​​of each dimension of the eigenvector, and edge attributes include the correlation coefficient. By setting a correlation threshold, only connections with correlations above the threshold are retained. For example, if the correlation threshold is set to 0.7, a connection is established between the corresponding nodes when the correlation coefficient between two eigenvectors is greater than 0.7.

[0045] Causal inference mechanisms are used to calculate weighted relationships between nodes. First, direct causal links are identified based on conditional temporal causal analysis. For each pair of connected nodes, their temporal data is analyzed and conditional mutual information is calculated to determine whether a direct causal relationship exists. Mutual information entropy between nodes is then calculated to assess the degree of association. For each identified direct causal link, weights are calculated based on the degree of association to generate a dynamic weight matrix. Larger weights indicate more significant causal relationships.

[0046] The knowledge graph is updated based on a dynamic weight matrix. A weight threshold of 0.5 is set, and connections with weights below the threshold are deleted. For retained connections, the connection strength is adjusted based on the weight value. Three types of features are extracted from the updated graph: node attribute features, including the original feature vectors of nodes; topological structure features, including node degrees and centrality; and temporal evolution features, including the changing trends of node attributes over time. Finally, these three types of features are fused to produce the final process knowledge vector.

[0047] The solution of this application can: This technical solution combines deep learning with knowledge graphs to achieve intelligent analysis and knowledge extraction of production data, significantly improving the efficiency and accuracy of data analysis. By introducing a causal reasoning mechanism, the causal relationship between process parameters is accurately identified, avoiding false associations caused by relying solely on correlation analysis. The knowledge graph constructed by this solution has the ability to dynamically update, can promptly reflect the changing patterns in the production process, and provide a more accurate basis for process optimization. It integrates multi-dimensional features such as node attributes, topological structure, and temporal evolution, making the extracted process knowledge more comprehensive and rich. The process knowledge vector formed based on this solution can be directly applied to subsequent process optimization and quality control, providing reliable data support for intelligent decision-making in the production process, and effectively improving production efficiency and product quality.

[0048] In an optional embodiment, the initial knowledge graph is dynamically updated based on a dynamic weight matrix, the connection relationship between nodes is adjusted according to the weight, connections with weights below a preset weight threshold are deleted, and connections with significant causal relationships are retained; node attribute features, topological structure features, and temporal evolution features are extracted from the updated knowledge graph, and the node attribute features, topological result features, and temporal evolution features are fused to generate a process knowledge vector, including: Based on the distribution law of weight values ​​in the dynamic weight matrix, the weight mean and standard deviation are weighted and combined to obtain the preset weight threshold; the weight values ​​in the dynamic weight matrix are compared with the preset weight threshold, and the node connections with weight values ​​higher than the preset weight threshold and significant causal relationships are retained, and the node connections with weight values ​​lower than the preset weight threshold are deleted to obtain the optimized knowledge graph; For the optimized knowledge graph, the parameter type, value range, and adjustment method of the node are extracted in combination with the process specifications to form the node attribute characteristics; based on the connection relationship of the node, the degree distribution, path characteristics, and hierarchical structure are extracted to form the topological structure characteristics; and the historical data of the node are used to extract the change trend and periodic law to form the time series evolution characteristics. The node attribute features, topological structure features and temporal evolution features are weighted and combined into a dynamic graph holographic feature spectrum, which is then input into the feature importance calculation module to obtain the influence weight of each feature on the process. The three types of features are adaptively fused according to the influence weight to generate a feature vector reflecting process knowledge.

[0049] like Figure 3 As shown, the method further includes: Obtain an initial process knowledge graph, which contains process parameter nodes and their connections. Construct a dynamic weight matrix, where each element represents the connection weight between corresponding nodes. Based on a large amount of historical data, a sliding time window is used to calculate the correlation coefficients between nodes. These correlation coefficients are normalized and used as dynamic weights. For example, in a certain manufacturing process, the correlation coefficient between the temperature and pressure nodes within a month is 0.85, which is normalized to a weight of 0.82.

[0050] A statistical analysis of the dynamic weight matrix was performed to determine the distribution of all weight values. The calculated mean of the weights was 0.65, with a standard deviation of 0.15. Based on the experience of process experts, the mean and standard deviation were weighted to obtain a preset weight threshold of 0.60. Each weight in the dynamic weight matrix was compared with the preset threshold. Connections with weights greater than 0.60 and a clear causal relationship between nodes were retained; otherwise, connections were deleted. For example, if the weight between the temperature and pressure nodes was 0.82, which was greater than the threshold and had a causal relationship, the connection was retained. However, if the weight between the temperature and vibration nodes was 0.45, which was less than the threshold, the connection was deleted.

[0051] Node features are extracted from the optimized knowledge graph. First, node attribute features are extracted, including parameter type, value range, and adjustment method. For example, the parameter type of the temperature node is continuous, with a value range of 20-100°C, and PID closed-loop adjustment is adopted. Secondly, topological structure features are extracted, including the degree distribution, shortest path, and hierarchical relationship of the node. For example, the out-degree of the temperature node is 3, the in-degree is 2, and the shortest path length with the pressure node is 1, which is located in the control layer. Then, the time series evolution features are extracted, and the change trend and periodic pattern of the node parameters are analyzed based on historical data. For example, the temperature shows an upward trend from 8 to 18 o'clock every day and a downward trend from 18 to 8 o'clock, with a clear diurnal change cycle.

[0052] The three extracted features are fused, and the dynamic graph holographic feature spectrum is first calculated. Using a sliding time window, the attribute changes, topological structure changes, and temporal evolution characteristics of the nodes are statistically analyzed within each window. For example, within a certain time window, the attributes of the temperature node remain stable, a new connection is added to the topological structure, and the temporal characteristics fluctuate. Based on feature importance analysis, the degree of influence of each feature on the process is calculated. The attribute feature influence weight is 0.4, the topological feature weight is 0.35, and the temporal feature weight is 0.25. The three features are adaptively fused according to the influence weights to generate a process knowledge vector. For example, the knowledge vector of the temperature node reflects its parameter characteristics, network position, and dynamic change patterns.

[0053] The solution of this application can: The connection relationship of the knowledge graph is optimized through a dynamic weight matrix and a preset threshold, redundant connections are deleted, key connections are highlighted, and the accuracy and interpretability of the knowledge graph are improved. Based on the optimized knowledge graph, the causal relationship between process parameters can be analyzed more accurately, providing a reliable basis for process optimization. Node features are extracted from multiple dimensions, and the attribute information, network position and time series changes of the nodes are comprehensively considered. A sliding time window is used to track the dynamic evolution process of the features, realizing a comprehensive characterization of process knowledge. The fused feature vector contains rich process knowledge and can be used for process parameter optimization and anomaly diagnosis. Based on feature importance analysis, the adaptive fusion of features is realized, and the fusion weight is dynamically adjusted according to the degree of influence of different features on the process. The adaptive fusion strategy improves the accuracy and robustness of knowledge representation, so that the fused feature vector better reflects the essential laws of the process. This method has strong versatility and can be extended to different types of manufacturing processes.

[0054] In an optional embodiment, a process knowledge vector is received and combined with heterogeneous process data collected in real time, and real-time features are extracted through a deep residual network; the real-time features are fused with the process knowledge vector and input into a prediction model; a defect risk assessment result is output based on the prediction model; when the defect risk assessment result exceeds a dynamic threshold, an early warning message is generated, including: Receive process knowledge vectors and collect heterogeneous process data in real time, perform time series alignment and normalization on the heterogeneous process data, and generate a standardized process data matrix. The standardized process data matrix is ​​input into a preset deep residual network. The deep residual network contains multiple cascaded feature extraction modules and residual connection modules. The feature extraction module extracts multi-level process features. The residual connection module is used to maintain the original information, fuse the original information with the multi-level process features, and output a real-time feature matrix. A feature fusion network is used to fuse the real-time feature matrix and the process knowledge vector. The feature fusion network adaptively calculates the correlation weights of the two types of features in the spatial dimension and channel dimension through the attention calculation unit, and uses the feature aggregation unit to perform a weighted combination of the two types of features based on the correlation weights to generate a fused feature vector. The fused feature vector is input into the prediction model for risk assessment. The time series feature extraction unit of the prediction model extracts the time series change features. The time series change features are input into the risk analysis unit to calculate the probability of occurrence of different types of defects and generate defect risk assessment results. Based on historical risk distribution and current operating parameters, the risk warning threshold is updated in real time through the threshold calculation unit of the dynamic threshold warning module; when the defect risk assessment result exceeds the risk warning threshold, the threshold calculation unit is used to generate warning information containing defect type, risk level and disposal suggestions.

[0055] The core of the process parameter-based defect warning method lies in the fusion of multi-source data features and dynamic risk assessment. It receives a process knowledge vector containing information such as process knowledge rules and process parameter ranges. This vector has 1024 dimensions, with each dimension corresponding to a knowledge feature. It also collects heterogeneous process data, including temperature, pressure, and speed, at a sampling frequency of 100 Hz.

[0056] Time alignment is performed on the collected heterogeneous process data, unifying data of different frequencies into the same timestamp sequence. Data normalization is performed using the maximum and minimum normalization method to generate a standardized process data matrix. This matrix contains multiple time windows, each containing data at 512 time points.

[0057] The standardized process data is fed into a deep residual network for feature extraction. This network consists of eight cascaded feature extraction modules, each consisting of two convolutional layers and one activation layer. A residual connection module adds the input data to the feature extraction results, effectively preserving the original information. The network outputs a 256-dimensional real-time feature matrix.

[0058] The feature fusion network employs a dual-branch structure, processing the real-time feature matrix and the process knowledge vector separately. In the spatial dimension, weight coefficients are calculated by calculating the similarity matrix between the two feature types. In the channel dimension, attention weights are derived based on feature importance. The feature aggregation unit concatenates and maps the weighted features, outputting a 512-dimensional fused feature vector.

[0059] The prediction model consists of a time series feature extraction unit and a risk analysis unit. The time series feature extraction unit uses a long short-term memory network structure and takes time series data with a window length of 100 as input. The risk analysis unit consists of three fully connected layers, with the output layer dimension equal to the number of defect types. This mapping yields the risk probability of each defect type.

[0060] The dynamic threshold warning module updates thresholds based on the risk distribution over the past seven days and current operating parameters. Specifically, it calculates the mean and standard deviation of the risk probability. When operating parameters are within the normal range, the threshold is set to the mean plus twice the standard deviation; when operating parameters approach the boundary, the threshold is appropriately lowered. When the risk probability of any defect type exceeds the corresponding threshold, an alert is generated.

[0061] During the steel rolling process, process parameters such as rolling temperature, pressure, and speed are collected in real time. At a certain moment, the rolling speed was 5 meters per second, the temperature was 850 degrees Celsius, and the pressure was 800 MPa. After feature extraction and fusion, the predicted probability of edge cracking defects was 0.82, exceeding the current threshold of 0.75. The system promptly generated a warning message indicating a high risk of edge cracking and recommended reducing the rolling speed. Based on this, the operator reduced the speed to 4.5 meters per second, reducing the defect risk to a safe level.

[0062] The solution of this application can: The combined application of a deep residual network and a feature fusion network achieves a deep fusion of process knowledge and real-time data, enhancing feature representation and making defect prediction more accurate and reliable. A dynamic threshold mechanism adjusts warning thresholds in real time based on historical risk distribution and operating parameters, effectively balancing the timeliness and accuracy of warnings and reducing false alarm rates. Predictive model design based on time series feature extraction and risk analysis captures the dynamic changes in process parameters, enabling quantitative assessment of different defect risk types and providing effective guidance for process optimization and quality control.

[0063] A second aspect of an embodiment of the present invention provides an automated testing system for heterogeneous process verification and system software debugging, including: The first unit is used to collect a production holographic monitoring dataset from a heterogeneous production process; semantically annotate the production holographic monitoring dataset using a preset ontology model to generate an annotated dataset; based on the annotated dataset, a multi-head attention mechanism and a graph neural network are used to train the preset ontology model to extract correlation features between process parameters, equipment status, and defect types; The second unit is used to call the trained preset ontology model to perform feature extraction and association analysis on the production holographic monitoring data set; construct an initial knowledge graph based on the extracted features; use the causal reasoning mechanism to calculate the weight relationship between nodes and generate a dynamic weight matrix; dynamically update the initial knowledge graph based on the dynamic weight matrix and output a process knowledge vector containing time-series evolution characteristics; The third unit is used to receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network; fuse the real-time features with the process knowledge vectors and input them into the prediction model; output the defect risk assessment results based on the prediction model; and generate early warning information when the defect risk assessment results exceed the dynamic threshold.

[0064] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call instructions stored in the memory to execute the aforementioned method.

[0065] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0066] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automated testing method for heterogeneous process verification and system software debugging, characterized in that: include: Collect production holographic monitoring data sets during heterogeneous process production; Use the preset ontology model to semantically annotate the production holographic monitoring dataset to generate an annotated dataset; Based on the labeled dataset, a multi-head attention mechanism and graph neural network are used to train a preset ontology model to extract the correlation features between process parameters, equipment status and defect types; Call the trained preset ontology model to perform feature extraction and association analysis on the production holographic monitoring data set; build the initial knowledge graph based on the extracted features; Use causal reasoning mechanism to calculate the weight relationship between nodes and generate a dynamic weight matrix; Dynamically update the initial knowledge graph based on the dynamic weight matrix and output the process knowledge vector containing temporal evolution characteristics; Receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network; Fuse the real-time features with the process knowledge vectors and input them into the prediction model; output the defect risk assessment results based on the prediction model; When the defect risk assessment result exceeds the dynamic threshold, an early warning message is generated.

2. The method according to claim 1, characterized in that Use the preset ontology model to semantically annotate the production holographic monitoring dataset to generate an annotated dataset; Based on the labeled dataset, a multi-head attention mechanism and graph neural network are used to train a preset ontology model to extract the correlation features between process parameters, equipment status, and defect types, including: Use the preset ontology model to semantically annotate the data in the production holographic monitoring dataset to generate an annotated dataset; A multi-head attention mechanism is trained based on the labeled dataset. The association weights of the labeled dataset are calculated through the multi-head attention mechanism to generate an association weight matrix. The association weight matrix is ​​input into the graph neural network, and the graph neural network is used to aggregate information on the labeled dataset, extract local structural features of the labeled dataset, and generate node feature vectors. The node feature vector is input into the improved ontology model training module, and the node representation in the preset ontology model is updated based on the node feature vector. The preset ontology model is optimized and trained using a loss function to obtain a trained ontology model. The trained ontology model is used to extract the correlation features between process parameters, equipment status and defect types.

3. The method according to claim 2, characterized in that The node feature vector is input into the improved ontology model training module. The node representation in the preset ontology model is updated based on the node feature vector. The preset ontology model is optimized and trained using a loss function to obtain a trained ontology model. The trained ontology model is used to extract the correlation features between process parameters, equipment status, and defect types, including: The node feature vectors are input into the improved ontology model training module, which uses a message passing mechanism to iteratively update the node feature vectors over multiple rounds. In each round, feature aggregation is performed based on attention weights. A bidirectional gated recurrent unit is used to capture the temporal dependencies of the node feature vectors, and residual connections are used to maintain the integrity of the node feature vectors. The node representation in the preset ontology model is updated based on the updated node feature vector; a multi-task loss function is constructed, including node feature reconstruction loss and structure preservation loss, and the preset ontology model is optimized and trained using the multi-task loss function to obtain the trained ontology model; the trained ontology model is used to extract the correlation features between process parameters, equipment status and defect types.

4. The method according to claim 1, characterized in that Call the trained preset ontology model to perform feature extraction and association analysis on the production holographic monitoring data set; build the initial knowledge graph based on the extracted features; Use causal reasoning mechanism to calculate the weight relationship between nodes and generate a dynamic weight matrix; The initial knowledge graph is dynamically updated based on the dynamic weight matrix, and the output process knowledge vector containing temporal evolution characteristics includes: Call the trained preset ontology model to perform feature extraction and correlation analysis on the production holographic monitoring data set to obtain multi-dimensional feature vectors and correlation matrices; construct the initial knowledge graph based on the multi-dimensional feature vectors and correlation matrix; The causal reasoning mechanism is used to calculate the weight relationship between nodes in the initial knowledge graph. The direct causal links between nodes are identified through the conditional temporal causal analysis method. The mutual information entropy is used to evaluate the degree of association between nodes. A dynamic weight matrix is ​​generated based on the direct causal links and the degree of association. The initial knowledge graph is dynamically updated based on the dynamic weight matrix, the connection relationship between nodes is adjusted according to the weight size, the connections with weights lower than the preset weight threshold are deleted, and the connections with significant causal relationships are retained; the node attribute features, topological structure features and temporal evolution features are extracted from the updated knowledge graph, and the node attribute features, topological result features and temporal evolution features are fused to generate a process knowledge vector.

5. The method according to claim 4, characterized in that The initial knowledge graph is dynamically updated based on the dynamic weight matrix. The connection relationship between nodes is adjusted according to the weight size, and the connection with weight lower than the preset weight threshold is deleted, while the connection with significant causal relationship is retained. The node attribute features, topological structure features and temporal evolution features are extracted from the updated knowledge graph, and the node attribute features, topological result features and temporal evolution features are fused to generate a process knowledge vector including: Based on the distribution law of weight values ​​in the dynamic weight matrix, the weight mean and standard deviation are weighted and combined to obtain the preset weight threshold; the weight values ​​in the dynamic weight matrix are compared with the preset weight threshold, and the node connections with weight values ​​higher than the preset weight threshold and significant causal relationships are retained, and the node connections with weight values ​​lower than the preset weight threshold are deleted to obtain the optimized knowledge graph; For the optimized knowledge graph, the parameter type, value range, and adjustment method of the node are extracted in combination with the process specifications to form the node attribute characteristics; based on the connection relationship of the node, the degree distribution, path characteristics, and hierarchical structure are extracted to form the topological structure characteristics; and the historical data of the node are used to extract the change trend and periodic law to form the time series evolution characteristics. The node attribute features, topological structure features and temporal evolution features are weighted and combined into a dynamic graph holographic feature spectrum, which is then input into the feature importance calculation module to obtain the influence weight of each feature on the process. The three types of features are adaptively fused according to the influence weight to generate a feature vector reflecting process knowledge.

6. The method according to claim 1, characterized in that Receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network. Fuse the real-time features with the process knowledge vectors and input them into the prediction model. Output defect risk assessment results based on the prediction model; When the defect risk assessment result exceeds the dynamic threshold, the generated warning information includes: Receive process knowledge vectors and collect heterogeneous process data in real time, perform time series alignment and normalization on the heterogeneous process data, and generate a standardized process data matrix. The standardized process data matrix is ​​input into a preset deep residual network. The deep residual network contains multiple cascaded feature extraction modules and residual connection modules. The feature extraction module extracts multi-level process features. The residual connection module is used to maintain the original information, fuse the original information with the multi-level process features, and output a real-time feature matrix. A feature fusion network is used to fuse the real-time feature matrix and the process knowledge vector. The feature fusion network adaptively calculates the correlation weights of the two types of features in the spatial dimension and channel dimension through the attention calculation unit, and uses the feature aggregation unit to perform a weighted combination of the two types of features based on the correlation weights to generate a fused feature vector. The fused feature vector is input into the prediction model for risk assessment. The time series feature extraction unit of the prediction model extracts the time series change features. The time series change features are input into the risk analysis unit to calculate the probability of occurrence of different types of defects and generate defect risk assessment results. Based on historical risk distribution and current operating parameters, the risk warning threshold is updated in real time through the threshold calculation unit of the dynamic threshold warning module; when the defect risk assessment result exceeds the risk warning threshold, the threshold calculation unit is used to generate warning information containing defect type, risk level and disposal suggestions.

7. An automated testing system for heterogeneous process verification and system software debugging, used to implement the method of any one of claims 1 to 6, characterized in that: include: The first unit is used to collect the production holographic monitoring data set in the heterogeneous process production process; use the preset ontology model to semantically annotate the production holographic monitoring data set to generate an annotated data set; Based on the labeled dataset, a multi-head attention mechanism and graph neural network are used to train a preset ontology model to extract the correlation features between process parameters, equipment status and defect types; The second unit is used to call the trained preset ontology model to perform feature extraction and association analysis on the production holographic monitoring data set; and to construct the initial knowledge graph based on the extracted features; Use causal reasoning mechanism to calculate the weight relationship between nodes and generate a dynamic weight matrix; Dynamically update the initial knowledge graph based on the dynamic weight matrix and output the process knowledge vector containing temporal evolution characteristics; The third unit is used to receive process knowledge vectors, combine them with heterogeneous process data collected in real time, and extract real-time features through a deep residual network; Fuse the real-time features with the process knowledge vectors and input them into the prediction model; output the defect risk assessment results based on the prediction model; When the defect risk assessment result exceeds the dynamic threshold, an early warning message is generated.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: The computer program instructions, when executed by a processor, implement the method of any one of claims 1 to 6.

Images