Multi-scale detection method and system for aluminum product quality
By combining spectral imaging, Raman spectrometer and laser ultrasound, constructing correlation maps and reinforcement learning, the accuracy problem of multi-scale feature detection on the surface and inside of aluminum products was solved, thereby improving the quality of aluminum products and production efficiency.
Patent Information
- Application Number
- CN202510992558.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-18
- Publication Date
- 2025-09-19
- Estimated Expiration
- 2045-07-18
AI Technical Summary
Existing technologies make it difficult to accurately detect multi-scale features on the surface and interior of aluminum products, resulting in inaccurate adjustment of process parameters and affecting the quality of aluminum products.
Spectral imaging and Raman spectrometers are used to obtain surface defects and corrosion activity of aluminum products, combined with laser ultrasonic scanning of the internal structure. By constructing correlation maps and strengthening learning to adjust detection parameters, comprehensive detection and analysis of multi-scale features can be achieved.
It realizes comprehensive detection and analysis of the multi-scale characteristics of aluminum products, accurately identifies defect patterns and locates the causes of defects, improves the quality of aluminum products and production efficiency, and reduces production costs.
Smart Images

Figure CN120489987B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of quality inspection technology, and in particular to a multi-scale inspection method and system for aluminum product quality. Background Art
[0002] In modern industry, aluminum products are widely used in a wide range of fields, including aerospace, automotive manufacturing, construction, and electronics, due to their excellent physical and mechanical properties, such as light weight, high strength, and good corrosion resistance. For example, in the aerospace field, the quality requirements for aluminum products are extremely high, and any minor defects can cause accidents. In automotive manufacturing, the quality of aluminum products affects the safety and service life of the vehicle. Therefore, ensuring the high quality of aluminum products is crucial, making aluminum product quality inspection a key link in the production process. However, most methods have not yet solved the problem of how to accurately detect the multi-scale characteristics of the surface and internal structure of aluminum products to accurately adjust process parameters and improve aluminum product quality. Summary of the Invention
[0003] In response to the shortcomings of the existing technology, this application provides a multi-scale detection method and system for aluminum product quality.
[0004] In a first aspect, the present application provides a multi-scale quality detection system for aluminum products. The system comprises: deploying a spectral imager and a Raman spectrometer to photograph the surface of the aluminum product to obtain surface defects and corrosion activity of the aluminum product, respectively; scanning the internal structure of the aluminum product using laser ultrasound in a sparse array layout; and generating the internal stress of the aluminum product in real time using a compressed sensing reconstruction algorithm;
[0005] The surface defects, corrosion activity, and internal stress of aluminum products are used as nodes in a graph neural network to construct a correlation graph. The weights of the edges in the correlation graph are dynamically adjusted based on the attention mechanism. The node features of the correlation graph represent the texture characteristics of surface defects, the frequency domain energy ratio of corrosion activity, and the gradient direction of internal stress. The edge weights in the correlation graph represent the similarity of cross-scale features.
[0006] The node features of the association graph are standardized and embedded, and the defect patterns of aluminum products are detected through a convolutional neural network. At the same time, the weights of the edges in the association graph are used as the state space of reinforcement learning to dynamically adjust the shooting frequencies of the spectral imager and Raman spectrum, as well as the scanning path of the laser ultrasound. The cause of the defect of the aluminum product is located through a causal reasoning algorithm based on the association graph, and the defect location of the aluminum product is output. The process parameters are simultaneously adjusted according to the defect pattern, defect location and defect cause of the aluminum product.
[0007] As an optional implementation, the construction logic of the association map includes:
[0008] Mapping the surface defects, corrosion activity and internal stress of aluminum products into nodes of the association graph, and assigning a unique identifier to each node;
[0009] Encode data for each node of the association graph to obtain node features;
[0010] Based on the physical structure of aluminum products, the node features of surface defects are connected with the node features of internal stress, and the node features of internal stress are connected with the node features of corrosion activity according to spatial proximity to form the initial edge connection of the association graph;
[0011] Add cross-layer causal edges of the association graph through the process knowledge base, and verify and adjust the node features and initial edge connections of the association graph.
[0012] As an optional implementation, the logic for adjusting the weights of edges in the association graph includes:
[0013] Determine the query vector and key vector of the attention mechanism based on the spatial attenuation factor between node features and the causal strength in the process knowledge base;
[0014] Construct reinforcement learning. The state space of reinforcement learning represents the weight of the edges in the association graph, the action space represents the weight ratio of the adjustment space attenuation factor and causal strength, and the reward function represents the efficiency and accuracy of defect detection.
[0015] The weights of the edges in the association graph are modified according to the process knowledge base, and the initial edge connections of the association graph are directionally pruned.
[0016] As an optional implementation manner, the detection logic of the defect mode includes:
[0017] The node features after normalized embedding are input into the convolutional neural network, which outputs spatial coordinates, time stamps, and modal confidence scores.
[0018] Dynamically allocate the receptive field of the convolution kernel according to the modal confidence to output a spatiotemporally separated feature map;
[0019] Extract the defect distribution map from the spatial channel of the feature map, and generate the corrosion trend line from the time channel of the feature map;
[0020] The defect distribution map and corrosion trend line are synchronously aligned with the process knowledge base to output the defect pattern.
[0021] As an optional implementation, the adjustment logic of the laser ultrasound scanning path includes:
[0022] The weights of the edges in the association graph are used as the state space of reinforcement learning, and the detection accuracy and energy consumption reduction rate are used as the reward function of reinforcement learning to explore the scanning path of laser ultrasound.
[0023] During the laser ultrasonic scanning process, the temperature and vibration data of the aluminum product are obtained in real time;
[0024] According to the changes in temperature data and vibration data, the state space of reinforcement learning is updated to dynamically adjust the scanning path of laser ultrasound.
[0025] As an optional implementation manner, the defect cause location logic includes:
[0026] Combine the process knowledge base to convert the association map into a causal association map;
[0027] Use counterfactual reasoning to simulate and analyze defect patterns and observe changes in defect patterns;
[0028] The causal association map is integrated with the changes in defect patterns to provide evidence and output the mapping relationship between process parameters and defects to locate the cause of the defect.
[0029] As an optional implementation manner, the output sub-logic of the mapping relationship between process parameters and defects includes:
[0030] The causal relationship map and the change of defect patterns are integrated into the evidence through Bayesian network;
[0031] Cross-validate the results of evidence fusion to evaluate the mapping relationship between process parameters and defects;
[0032] Visualize the mapping relationship between process parameters and defects.
[0033] As an optional implementation, the process parameter adjustment logic includes:
[0034] Determine the adjustment direction of process parameters based on defect distribution map and corrosion trend line;
[0035] Optimize the adjustment direction of process parameters based on the located defect causes and the mapping relationship between process parameters and defects;
[0036] Output the adjustment priority of process parameters for defect locations, and determine the adjustment plan of process parameters based on the adjustment priority of process parameters;
[0037] Adjustments to aluminum products on the production line are implemented according to the process parameter adjustment plan, and surface defects, internal stress and corrosion activity of aluminum products are monitored in real time to determine whether the process parameter adjustment plan should be corrected.
[0038] As an optional implementation, the logic for generating the internal stress includes:
[0039] Dense detection points are set in stress-sensitive areas, and sparse detection points are set in non-stress-sensitive areas to obtain a sparse array layout;
[0040] The internal structure of the aluminum product is scanned by laser ultrasound in a sparse array layout, and the ultrasonic signal is converted into internal stress;
[0041] Based on the historical internal stress, the internal stress of the aluminum product is generated in real time through the compressed sensing algorithm, the stress concentration area and defect location are simultaneously identified, and the sparse array layout is updated according to the error of the internal stress.
[0042] In a second aspect, the present application provides a multi-scale detection method for aluminum product quality, the method comprising: deploying a spectral imager and a Raman spectrometer to photograph the surface of the aluminum product, respectively obtaining surface defects and corrosion activity of the aluminum product;
[0043] Laser ultrasound is used to scan the internal structure of aluminum products in a sparse array layout, while the internal stress of aluminum products is generated in real time using a compressed sensing reconstruction algorithm.
[0044] The surface defects, corrosion activity, and internal stress of aluminum products are used as nodes in a graph neural network to construct a correlation map. The node features of the correlation map represent the texture characteristics of surface defects, the frequency domain energy ratio of corrosion activity, and the gradient direction of internal stress.
[0045] The weights of edges in the association graph are dynamically adjusted based on the attention mechanism. The weights of edges in the association graph represent the similarity of cross-scale features.
[0046] The node features of the association graph are standardized and embedded, and the defect patterns of aluminum products are detected through convolutional neural networks.
[0047] The weights of the edges in the association graph are used as the state space of reinforcement learning to dynamically adjust the shooting frequency of the spectral imager and Raman spectrum, as well as the scanning path of the laser ultrasound.
[0048] Based on the association map, the cause of the defect of the aluminum product is located through the causal reasoning algorithm, and the defect location of the aluminum product is output. At the same time, the process parameters are adjusted according to the defect mode, defect location and defect cause of the aluminum product.
[0049] Compared with the existing technology, the beneficial effects of this application are: by constructing a closed-loop detection system from data acquisition, feature association to feedback adjustment, comprehensive detection and analysis of multi-scale features of aluminum products from the surface to the inside can be achieved, and the defect patterns of aluminum products can be accurately identified and the causes of defects can be located, and based on this, the process parameters can be effectively adjusted, thereby significantly improving the quality of aluminum products, improving production efficiency, reducing production costs, and enhancing the competitiveness of products in the market, providing an innovative and efficient solution for quality control in the production process of aluminum products.
[0050] The surface defects and corrosion activity of aluminum products are obtained by spectral imaging and Raman spectrometer respectively, which can evaluate the surface quality of aluminum products from different dimensions. The spectral imaging instrument obtains reflective spectrum images with high resolution and wide spectral range, which can clearly present the geometric features of the surface and help to detect surface defects such as scratches and pits; the Raman spectrometer obtains Raman scattering spectra through laser excitation of specific wavelengths, which can accurately detect the chemical properties of the surface and effectively identify chemical defects such as corrosion activity, which provides rich and accurate surface information for subsequent quality assessment and process adjustment; laser ultrasound is used to scan the internal structure of aluminum products according to a sparse array layout, and the internal stress is generated in real time through the compressed sensing reconstruction algorithm. Laser ultrasonic testing has the advantages of non-contact, fast, high precision and no damage to aluminum products. It can promptly detect stress concentration areas and defect locations inside aluminum products, providing key data for in-depth understanding of the internal quality status of aluminum products.
[0051] The surface defects, corrosion activity and internal stress of aluminum products are used as nodes of the graph neural network to construct an associated graph, which provides an effective framework for comprehensively analyzing the relationship between the multi-scale characteristics of aluminum products. It more comprehensively and accurately displays the complex relationship between the multi-scale characteristics of aluminum products, and improves the practicality and reliability of the graph. Based on the attention mechanism, it can more comprehensively and accurately measure the relationship between nodes, provide a reasonable basis for the dynamic adjustment of edge weights, improve its computational efficiency and the pertinence of aluminum product quality inspection, and provide a more accurate graph basis for subsequent inspection feedback.
[0052] Standardizing the node features of the association graph and embedding them into the convolutional neural network can effectively improve the training efficiency and accuracy of the convolutional neural network. Taking the weights of the edges in the association graph as the state space of reinforcement learning, and using detection accuracy and energy consumption reduction rate as reward functions to explore the scanning path of laser ultrasound, can make full use of the information contained in the association graph and dynamically adjust the scanning path according to the specific characteristics of the aluminum product. This not only improves the detection accuracy and reduces the detection blind spots, but also avoids unnecessary scanning, reduces energy consumption, and saves detection costs; combining the located defect causes and defect locations to optimize the adjustment direction of process parameters can avoid blind adjustments, reduce the negative impact on production, and improve the success rate of product quality improvement. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be derived from these drawings without inventive work. Among them:
[0054] Figure 1This is a system flow chart of the multi-scale detection system for aluminum product quality provided in an embodiment of the present application;
[0055] Figure 2 A logic diagram for adjusting the weights of edges in a correlation graph of a multi-scale detection system for aluminum product quality provided in an embodiment of the present application;
[0056] Figure 3 A logic diagram for adjusting the scanning path of the laser ultrasonic system for the multi-scale detection of aluminum product quality provided in an embodiment of the present application;
[0057] Figure 4 This is a flow chart of the method for multi-scale detection of aluminum product quality provided in an embodiment of the present application. DETAILED DESCRIPTION
[0058] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described below in conjunction with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.
[0059] Example 1
[0060] like Figure 1 As shown, a system flow chart of a multi-scale detection system for aluminum product quality is provided for an embodiment of the present application. The multi-scale detection system for aluminum product quality includes a scale acquisition module, a scale association module and a detection feedback module.
[0061] The scale acquisition module is used to deploy spectral imagers and Raman spectroscopy to capture the surface of aluminum products, respectively obtaining surface defects and corrosion activity of aluminum products. It also uses laser ultrasound to scan the internal structure of aluminum products according to a sparse array layout, and uses a compressed sensing reconstruction algorithm to generate the internal stress of aluminum products in real time.
[0062] Before deploying the spectral imager and Raman spectrometer, the spectral calibration of the two devices is performed using standard aluminum samples based on the material characteristics of different types of aluminum products. By analyzing the reflection and scattering spectral characteristics of the standard samples at specific wavelengths, the spectral response of the two devices is determined, and the device parameters are adjusted to ensure that the spectral imager and Raman spectrometer can accurately capture the spectral information of the aluminum product surface, thereby improving the accuracy of the spectral imager and Raman spectrometer in capturing the spectral information of the aluminum product surface and reducing the problem of inaccurate detection due to equipment errors.
[0063] When acquiring surface data of aluminum products, a synchronous triggering mechanism is used to enable the spectral imager and Raman spectrometer to simultaneously photograph the same area on the surface of the aluminum product. The spectral imager obtains the reflected spectrum image of the aluminum product surface with high resolution and a wide spectral range, while the Raman spectrometer excites the aluminum product surface with a laser of a specific wavelength to obtain the Raman scattering spectrum, thereby capturing different dimensional information at the same position on the aluminum product surface, providing more comprehensive data support for subsequent analysis, and helping to more accurately identify surface defects and corrosion activity.
[0064] The acquired spectral data contains a large amount of noise and redundant information, and needs to be preprocessed. The reflectance spectrum image obtained by the spectral imager is denoised, and the wavelet transform is used to extract features from the Raman spectral data to remove high-frequency noise while retaining key spectral features. The minimum and maximum normalization method is used to map the two spectral data to the same scale, thereby improving the quality of the spectral data, reducing the interference of noise and redundant information, and making subsequent feature extraction and recognition more accurate and efficient for subsequent analysis.
[0065] The preprocessed reflectance spectrum image is input into the convolutional neural network. The spatial features in the reflectance spectrum image are extracted through the convolutional layer, pooling layer and fully connected layer of the convolutional neural network to identify geometric defects on the surface of aluminum products, such as scratches and pits. The Raman spectrum data is input into the long short-term memory network to analyze the change characteristics of the spectrum over time and identify chemical defects on the surface of aluminum products, such as corrosion activity. This can achieve accurate identification of surface defects and corrosion activity of aluminum products, providing a basis for subsequent quality assessment and process adjustment.
[0066] Specifically, the internal stress generation logic includes:
[0067] Dense detection points are set in stress-sensitive areas, and sparse detection points are set in non-stress-sensitive areas to obtain a sparse array layout;
[0068] The internal structure of the aluminum product is scanned by laser ultrasound in a sparse array layout, and the ultrasonic signal is converted into internal stress;
[0069] Based on the historical internal stress, the internal stress of the aluminum product is generated in real time through the compressed sensing algorithm, the stress concentration area and defect location are simultaneously identified, and the sparse array layout is updated according to the error of the internal stress.
[0070] Aluminum products have different stress distributions in different areas. Stress concentration and defects are more likely to occur in stress-sensitive areas. Setting dense detection points in stress-sensitive areas and sparse detection points in non-stress-sensitive areas can not only ensure the detection accuracy of key areas, but also reduce detection costs and time. According to the structural characteristics, stress conditions and historical data of aluminum products, stress-sensitive areas and non-stress-sensitive areas are determined. Detection points are set at smaller intervals in stress-sensitive areas, and detection points are set at larger intervals in non-stress-sensitive areas to form a sparse array layout. This reduces the number of detection points, reduces detection costs and time, and improves detection efficiency while ensuring detection accuracy.
[0071] Laser ultrasound can penetrate the interior of aluminum products. By detecting ultrasonic signals, internal structural information can be obtained and converted into internal stress. Laser ultrasonic equipment is used to scan aluminum products in a sparse array layout, emitting laser pulses to excite ultrasonic signals. By receiving ultrasonic signals and analyzing parameters such as their propagation time and amplitude, internal structural information can be obtained. This allows for non-contact and rapid acquisition of the internal stress of aluminum products with high detection accuracy and without damaging the aluminum products.
[0072] Based on historical internal stress and current ultrasonic signals, the internal stress of aluminum products is generated in real time through a compressed sensing algorithm, which can promptly detect stress concentration areas and defect locations. The sparse array layout is updated according to the internal stress error, which can improve the accuracy of subsequent detection. The compressed sensing reconstruction algorithm is used to combine historical internal stress and current ultrasonic signals to generate the internal stress of aluminum products in real time and calculate the errors of local and global internal stress. Specifically, when the error of local internal stress is greater than 10%, dense detection points are added. When the error of global internal stress is less than 5%, redundant dense detection points are merged and the sparse array layout is updated. In this way, the internal stress of aluminum products can be generated in real time and accurately, and stress concentration areas and defect locations can be promptly detected. By dynamically updating the sparse array layout, the accuracy and efficiency of detection are improved. The generated internal stress will be used as input to the scale association module to construct a correlation map to further analyze the correlation between internal stress, surface defects and corrosion activity. The updated sparse array layout will be used for the next laser ultrasonic scan to improve the accuracy of subsequent detection.
[0073] The scale association module is used to use the surface defects, corrosion activity and internal stress of aluminum products as nodes of the graph neural network to construct an association graph, and dynamically adjust the weights of the edges in the association graph based on the attention mechanism. The node features of the association graph represent the texture features of surface defects, the frequency domain energy ratio of corrosion activity, and the gradient direction of internal stress. The weights of the edges in the association graph represent the similarity of cross-scale features.
[0074] Specifically, the construction logic of the association graph includes:
[0075] Mapping the surface defects, corrosion activity and internal stress of aluminum products into nodes of the association graph, and assigning a unique identifier to each node;
[0076] Encode data for each node of the association graph to obtain node features;
[0077] Based on the physical structure of aluminum products, the node features of surface defects are connected with the node features of internal stress, and the node features of internal stress are connected with the node features of corrosion activity according to spatial proximity to form the initial edge connection of the association graph;
[0078] Add cross-layer causal edges of the association graph through the process knowledge base, and verify and adjust the node features and initial edge connections of the association graph.
[0079] In order to clearly represent the key features of aluminum products such as surface defects, corrosion activity and internal stress in the association map, it is necessary to map these features into nodes in the map and assign each node a unique identifier to facilitate subsequent operation and management of the nodes. For surface defects, a unique identifier is generated by encoding information such as their type and location, where types include scratches and pits. For corrosion activity, an identifier is generated based on features such as the degree of corrosion and the area of occurrence, while an identifier is generated based on the stress magnitude and location. That is, a hash algorithm is used to combine these data to generate a unique hash value as the node identifier. This ensures that each node in the association map has clear corresponding features and a unique identity, facilitating the accurate identification and processing of each feature in the association map construction and subsequent analysis.
[0080] Different types of nodes have different feature representations. Data encoding converts these node features into numerical forms suitable for graph neural network processing, which helps to better explore the relationship between nodes. For the texture features of surface defects, the image of the defect area is converted into a feature vector through the local binary pattern. For the frequency domain energy ratio of corrosion activity, the Raman spectral data is first Fourier transformed, and then the energy ratio of different frequency bands is calculated, which is used as the encoding feature of the corrosion active node. For the gradient direction of internal stress, the consistency index of the gradient direction of each point in the stress field is calculated, and the structural tensor analysis method is used to use the calculation result as the encoding value of the internal stress node. In this way, the complex features of the node are converted into a unified numerical encoding form, which enhances the comparability and computability between different types of node features and improves the learning efficiency of the graph neural network for node features.
[0081] Based on the physical structure of aluminum products, there is a certain spatial and physical connection between surface defects, internal stress and corrosion activity. By establishing initial edge connections, these related nodes are connected, and the structure of the association map is preliminarily constructed, laying the foundation for the subsequent mining of cross-scale feature relationships; through the finite element analysis method, the physical process of aluminum products during production and use is simulated to determine the potential connection area between surface defects and internal stress. If the simulation finds a stress concentration area under a surface scratch, an edge connection is established between the corresponding surface defect node and the internal stress node.
[0082] For the node characteristics of internal stress and the node characteristics of corrosion activity, based on the material properties and physical and chemical principles of aluminum products, the areas that spatially influence each other are determined. In high-stress areas, aluminum products are more prone to corrosion. Edge connections are established between these corresponding node characteristics of internal stress and node characteristics of corrosion activity according to spatial proximity; thus, the structure of the association map is preliminarily constructed, reflecting the potential connection between different characteristics of aluminum products based on physical structure, and providing a basic framework for further exploring cross-scale feature relationships.
[0083] The process knowledge base contains a large amount of knowledge about the causal relationship between the production process and product characteristics of aluminum products. By adding cross-layer causal edges and integrating this knowledge into the association graph, it can more comprehensively reflect the causal relationship between the multi-scale characteristics of aluminum products. At the same time, the association graph is verified and adjusted to ensure the accuracy and rationality of the association graph; causal rules related to surface defects, internal stress and corrosion activity are extracted from the process knowledge base. If the process knowledge base indicates that excessively high casting temperature will lead to increased internal stress, which in turn causes increased surface corrosion activity, a cross-layer causal edge representing this causal relationship is added between the corresponding node feature of internal stress and the node feature of corrosion activity.
[0084] Through machine learning verification methods, the association graph after adding cross-layer causal edges is verified to check whether there are contradictory causal relationships or unreasonable connections in the association graph. For any problems found, the node features and initial edge connections are corrected through manual intervention or automatic adjustment algorithms; so that the association graph not only contains connections based on physical structure, but also incorporates causal relationships in process knowledge, more comprehensively and accurately reflecting the complex relationship between the multi-scale characteristics of aluminum products, thereby improving the practicality and reliability of the association graph.
[0085] Specifically, if Figure 2 As shown in the figure, the logic for adjusting the weights of edges in the association graph includes:
[0086] Determine the query vector and key vector of the attention mechanism based on the spatial attenuation factor between node features and the causal strength in the process knowledge base;
[0087] Construct reinforcement learning. The state space of reinforcement learning represents the weight of the edges in the association graph, the action space represents the weight ratio of the adjustment space attenuation factor and causal strength, and the reward function represents the efficiency and accuracy of defect detection.
[0088] The weights of the edges in the association graph are modified according to the process knowledge base, and the initial edge connections of the association graph are directionally pruned.
[0089] The attention mechanism can dynamically adjust the weight of the edge according to the characteristic relationship between nodes. The spatial attenuation factor and the causal strength in the process knowledge base are important factors affecting the relationship between nodes. By determining the query vector and key vector corresponding to these factors, a calculation basis is provided for the attention mechanism, thereby achieving reasonable adjustment of the edge weight; for the spatial attenuation factor, according to the physical size of the aluminum product and the spatial distance between nodes, an exponential attenuation function is used to obtain that the longer the distance, the smaller the spatial attenuation factor, and the spatial attenuation factors between different node pairs are calculated simultaneously.
[0090] Extract causal strength information from the process knowledge base and determine the causal strength value for each cross-layer causal edge. For the causal relationship, that is, the high casting temperature leads to increased internal stress, the causal strength coefficient is determined based on historical data and process analysis. Based on the spatial attenuation factor and the causal strength value, the formula of the attention mechanism is combined. , calculate the query vector and key vector , Represents the query vector The transposed matrix of Represents the weight of the edge between nodes in the association graph, which is used to measure the strength of the relationship between nodes. After calculation by the softmax activation function, the value range is ,all The sum is 1, Represents the vector dimension, used for Normalize the calculation results to prevent the gradient of the softmax activation function from disappearing due to the dot product being too large. Represents the spatial attenuation factor, which reflects the influence of the relationship between the spatial position of the node on the weight of the edge. Represents causal strength, which is used to measure the strength of causal relationships between nodes. and They are hyperparameters for balancing the influence of spatial attenuation factor and causal strength, which are optimized and determined through experiments or machine learning methods. By considering spatial attenuation and causal strength, the relationship between nodes can be measured more comprehensively and accurately, providing a more reasonable basis for the dynamic adjustment of edge weights, so that the association map can better reflect the real connection between the multi-scale characteristics of aluminum products.
[0091] In order to further optimize the edge weights so that they can more accurately reflect the relationship between defect detection efficiency and accuracy and node features, a reinforcement learning model is constructed. Through continuous exploration and learning, the optimal edge weight adjustment strategy is found to improve the guiding role of the association graph on aluminum product quality inspection. The state space of reinforcement learning is defined as the edge weights in the association graph, and the action space is defined as the operations to adjust the spatial attenuation factor and the proportion of causal strength weights, including actions to increase or decrease the weight coefficient of the spatial attenuation factor when calculating the edge weight.
[0092] The defect detection efficiency and accuracy are used as the reward function. The detection efficiency is measured by the detection time or the amount of data processed during the detection process, and the detection accuracy is determined by comparing with the actual known defect situation. When the edge weight is adjusted, the detection accuracy is improved and the detection time is shortened, and a higher reward value is given. Deep Q network training reinforcement learning is used to learn the optimal edge weight adjustment. During the training process, reinforcement learning selects the action space according to the current state space, observes the new state and the reward obtained after executing the action space, and continuously updates the strategy to maximize the long-term reward; thus, through reinforcement learning, it can automatically explore and optimize the edge weight adjustment, so that the edge weight in the association graph can be better associated with the defect detection efficiency and detection accuracy, thereby improving the practicality and effectiveness of the association graph in aluminum product quality inspection.
[0093] The process knowledge base contains process knowledge that has been verified in practice. Correcting the edge weights based on this process knowledge can further improve the accuracy of the edge weights. At the same time, the initial edge connections of the association graph are directed pruned to remove those edge connections that have little or unreasonable contributions to defect detection efficiency and accuracy, and optimize the structure of the association graph. Correction rules related to edge weights are extracted from the process knowledge base. If the process knowledge base shows that under certain specific process conditions, the causal relationship between two node features is weak, the weight of the edge between them is reduced accordingly. According to the best strategy obtained by reinforcement learning, the edge weight is adjusted. If reinforcement learning finds that increasing the weight of the edge between node features in a certain area can improve detection accuracy, the reinforcement learning is used to adjust the edge weight. The suggestions of learning are adjusted. For the initial edge connections of the association graph, the importance score of each edge to the defect detection efficiency and accuracy is calculated. The edges with scores less than the score threshold are deleted to achieve directional pruning. By combining the process knowledge base and the results of reinforcement learning, the edge weights are corrected to improve the accuracy and rationality of the edge weights. Directed pruning optimizes the structure of the association graph, reduces redundant connections, and improves the computational efficiency of the association graph and the pertinence of aluminum product quality inspection. The corrected edge weights and the optimized association graph structure are used as the input of the detection feedback module to dynamically adjust the shooting frequency of the spectral imager and Raman spectroscopy and the scanning path of the laser ultrasound, while also providing a more accurate graph basis for locating the cause of the defect.
[0094] The detection feedback module is used to standardize and embed the node features of the association graph, and detect the defect patterns of aluminum products through convolutional neural networks. At the same time, the weights of the edges in the association graph are used as the state space of reinforcement learning to dynamically adjust the shooting frequency of the spectral imager and Raman spectrum, as well as the scanning path of the laser ultrasound. According to the association graph, the cause of the defect of the aluminum product is located through the causal reasoning algorithm, and the defect position of the aluminum product is output. At the same time, the process parameters are adjusted according to the defect pattern, defect location and defect cause of the aluminum product.
[0095] Specifically, the detection logic of the defect mode includes:
[0096] The node features after normalized embedding are input into the convolutional neural network, which outputs spatial coordinates, time stamps, and modal confidence scores.
[0097] Dynamically allocate the receptive field of the convolution kernel according to the modal confidence to output a spatiotemporally separated feature map;
[0098] Extract the defect distribution map from the spatial channel of the feature map, and generate the corrosion trend line from the time channel of the feature map;
[0099] The defect distribution map and corrosion trend line are synchronously aligned with the process knowledge base to output the defect pattern.
[0100] The format and range of node features in the association graph are inconsistent, and it is difficult to effectively learn by directly inputting them into the neural network. Standardized embedding processing can transform these node features into a unified form suitable for network processing, improve the training efficiency and accuracy of the convolutional neural network, and thus accurately detect defect patterns; the node features are reduced in dimension through principal component analysis to remove redundant information while retaining the main features, and then the maximum and minimum normalization is used to map the eigenvalues to the [0,1] interval. The processed node features are input into the pre-trained convolutional neural network, which has multiple convolutional layers and pooling layers and can effectively extract image features; thereby improving the quality and comparability of node features, enabling the convolutional neural network to better learn the relationship between node features and defect patterns, and improving the accuracy and stability of defect pattern detection.
[0101] Different defect patterns have different characteristics in spatial and temporal dimensions. The fixed convolution kernel receptive field is difficult to fully capture these features. Dynamically adjusting the receptive field according to the modal confidence can adaptively focus on defects of different scales and feature distributions, improving the comprehensiveness of feature extraction. The modal confidence output by the convolutional neural network represents the credibility of the prediction of different defect patterns. When the modal confidence of a certain area is low, it indicates that the current receptive field has not fully captured the key features. At this time, the receptive field of the convolution kernel is increased to obtain broader contextual information. Conversely, when the modal confidence is high, the receptive field is appropriately reduced to focus on detail feature extraction. This enhances the adaptability of the convolutional neural network to complex defect patterns, improves the accuracy and efficiency of feature extraction, and can more accurately output spatiotemporally separated feature maps, providing more valuable data for subsequent analysis.
[0102] Extracting defect distribution maps and corrosion trend lines from the time-space separation feature maps can intuitively display the spatial distribution of defects and the trend of corrosion changes over time, providing a key basis for judging defect patterns and subsequent process adjustments; for the feature maps of the spatial channel, the defect area is separated from the background based on the threshold segmentation method to generate a defect distribution map; for the feature maps of the time channel, curve fitting is performed on the characteristic values at different time points, including using the least squares method to fit polynomial curves to generate corrosion trend lines; thereby, complex feature maps are converted into intuitive, easy-to-understand and analyze visual results, which helps engineers quickly grasp the defect situation of aluminum products and provide clear data support for subsequent decision-making.
[0103] The process knowledge base contains a large number of known relationships between defect patterns, process parameters and product characteristics. By aligning the extracted defect distribution map and corrosion trend line with the process knowledge base, the existing knowledge can be used to accurately judge the defect pattern of the current aluminum product, providing direction for subsequent in-depth analysis and problem solving; a matching algorithm is established between defect characteristics and defect pattern descriptions in the process knowledge base, and the defect shape, size and distribution density in the defect distribution map are compared with the standard defect pattern in the process knowledge base to calculate the similarity. For the corrosion trend line, its slope and change period are matched with the trend characteristics corresponding to different corrosion types in the process knowledge base. By comprehensively evaluating these matching results, the most suitable defect pattern is determined; thus, the defect pattern of the aluminum product can be determined quickly and accurately, and with the help of existing knowledge and experience, the subjectivity and uncertainty of defect judgment can be reduced, providing a clear goal for subsequent process adjustments.
[0104] Specifically, if Figure 3 As shown, the adjustment logic of the laser ultrasound scanning path includes:
[0105] The weights of the edges in the association graph are used as the state space of reinforcement learning, and the detection accuracy and energy consumption reduction rate are used as the reward function of reinforcement learning to explore the scanning path of laser ultrasound.
[0106] During the laser ultrasonic scanning process, the temperature and vibration data of the aluminum product are obtained in real time;
[0107] According to the changes in temperature data and vibration data, the state space of reinforcement learning is updated to dynamically adjust the scanning path of laser ultrasound.
[0108] The weights of the edges in the association graph reflect the strength of the association between different features of aluminum products. Using it as the state space of reinforcement learning can make full use of the information contained in the graph, explore the optimal laser ultrasonic scanning path, improve detection accuracy and reduce energy consumption; use the policy gradient algorithm to construct a reinforcement learning model, and use the weight vectors of the edges in the association graph as the state space input. The action space is defined as the adjustment operations on the laser ultrasonic scanning path, including changing the scanning direction, increasing or decreasing the scanning point density, etc. The reward function combines the detection accuracy and energy consumption reduction rate. The detection accuracy is determined by comparing with known defect samples, and the energy consumption reduction rate is calculated based on the power consumption of the laser ultrasonic equipment; through automatic exploration of the scanning path through reinforcement learning, the scanning strategy can be dynamically adjusted according to the specific characteristics of the aluminum product, improving detection efficiency and accuracy, while reducing energy consumption and saving detection costs.
[0109] During the laser ultrasonic scanning process, the temperature and vibration data of aluminum products can reflect changes in the internal structure and defects. Real-time acquisition of this data can promptly detect anomalies, provide a real-time basis for dynamic adjustment of the scanning path, and further improve the accuracy of detection. Temperature sensors and vibration sensors are installed on the laser ultrasonic scanning equipment to obtain temperature and vibration data in real time. This provides real-time and accurate feedback information for the dynamic adjustment of the scanning path, which can promptly capture subtle changes in the internal structure of aluminum products, avoid missing potential defects, and improve the reliability of detection.
[0110] When the temperature and vibration data of aluminum products change, it indicates that there are abnormalities or defects in their internal structure. The scanning path needs to be adjusted in time to more accurately detect these areas and improve detection accuracy. A mapping relationship between temperature and vibration data and scanning path adjustment is established. When the temperature suddenly rises or the vibration amplitude is greater than the set threshold, it is judged that there are defects in the area. According to the strategy of the reinforcement learning model, the scanning point density of the area is increased or the scanning direction is changed. This realizes the dynamic optimization of the scanning path, which can more accurately detect defects inside aluminum products, improve detection accuracy, reduce detection blind spots, avoid unnecessary scanning, and reduce energy consumption.
[0111] Specifically, the logic for locating the cause of the defect includes:
[0112] Combine the process knowledge base to convert the association map into a causal association map;
[0113] Use counterfactual reasoning to simulate and analyze defect patterns and observe changes in defect patterns;
[0114] The causal association map is integrated with the changes in defect patterns to provide evidence and output the mapping relationship between process parameters and defects to locate the cause of the defect.
[0115] The association map mainly shows the association relationship between different characteristics of aluminum products, while the causal association map can clarify the causal direction and strength of these relationships. Converting the association map into a causal association map helps to deeply analyze the causes of defects and provide a more powerful tool for subsequent location of defect causes. The association map is converted through a causal inference method based on Bayesian networks, combined with the causal knowledge in the process knowledge base. In the process knowledge base, the causal relationship is clearly recorded, that is, the high casting temperature leads to increased internal stress. This knowledge is integrated into the edge connection of the association map, the causal direction and strength of the edge are determined, and a causal association map is constructed. This can clearly show the causal relationship between defects and process parameters and product characteristics, providing an intuitive and accurate framework for in-depth analysis of defect causes, reducing the complexity and uncertainty of the analysis.
[0116] Through counterfactual reasoning, assuming that certain process parameters or product characteristics have changed and observing the changes in the defect pattern, we can gain a deep understanding of the causal mechanism of defect generation and accurately identify the cause of the current defect pattern. Using the potential result model, we can perform counterfactual reasoning on the causal association map. For example, by lowering the casting temperature to a certain value, we can simulate and calculate the changes in characteristics such as internal stress and surface defects of aluminum products in this case to observe whether the defect pattern changes. This allows us to deeply analyze the causal relationship of defects in a virtual environment, providing an effective exploration method for locating the cause of defects, avoiding a large number of trial and error experiments in actual production, and saving time and costs.
[0117] Furthermore, the output sub-logic of the mapping relationship between process parameters and defects includes:
[0118] The causal relationship map and the change of defect patterns are integrated into the evidence through Bayesian network;
[0119] Cross-validate the results of evidence fusion to evaluate the mapping relationship between process parameters and defects;
[0120] Visualize the mapping relationship between process parameters and defects.
[0121] Bayesian networks can effectively process uncertain information, and fuse the causal association map with the changes in defect patterns through Bayesian networks. They can fully utilize the information of both and accurately establish the mapping relationship between process parameters and defects. A Bayesian network model is constructed, and the nodes in the causal association map are used as variables of the Bayesian network, the causal strength of the edges is used as the conditional probability between the variables, and the changes in defect patterns obtained by counterfactual reasoning are input into the Bayesian network as new evidence. The Bayesian inference algorithm is used to update the probability distribution of variables in the network to obtain the probabilistic mapping relationship between process parameters and defects. Through the reasoning ability of the Bayesian network, the causal relationship and the actual changes in defect patterns are comprehensively considered, and the resulting mapping relationship is more credible and accurate, providing a solid foundation for subsequent evaluation and application.
[0122] In order to ensure that the mapping relationship between process parameters and defects is accurate and reliable, cross-validation is required to avoid overfitting and misjudgment and improve the generalization ability of the mapping relationship; the historical production data is divided into multiple subsets, and the k-fold cross-validation method is adopted. One of the subsets is used as the test set each time, and the remaining subsets are used as the training set. A Bayesian network is constructed on the training set to perform evidence fusion to obtain the mapping relationship, which is then verified on the test set. The performance of the mapping relationship is evaluated by calculating indicators such as accuracy and recall rate, and the reliability of the mapping relationship is comprehensively evaluated based on the results of multiple cross-validations; thereby improving the accuracy and reliability of the mapping relationship, avoiding incorrect mapping caused by data bias or model overfitting, and providing more reliable results for subsequent applications.
[0123] Visualizing the mapping relationship between process parameters and defects enables engineers to understand and analyze these relationships more intuitively, making decisions more quickly and facilitating process adjustments. Graphical tools are used, including drawing cause-and-effect diagrams or scatter plots. In cause-and-effect diagrams, nodes represent process parameters and defects, edges represent cause-and-effect relationships, and color or line thickness represent the strength of cause-and-effect. For scatter plots, data points are plotted and curves are fitted with process parameters as the horizontal axis and defect-related indicators as the vertical axis to demonstrate the quantitative relationship between the two. By developing a visual interface, the mapping relationship is presented in an intuitive and interactive manner. This converts complex mapping relationships into intuitive and easy-to-understand graphics, reducing the difficulty for engineers to understand and analyze, improving decision-making efficiency, and helping to formulate process adjustment strategies more quickly and accurately.
[0124] Specifically, the adjustment logic of process parameters includes:
[0125] Determine the adjustment direction of process parameters based on defect distribution map and corrosion trend line;
[0126] Optimize the adjustment direction of process parameters based on the located defect causes and the mapping relationship between process parameters and defects;
[0127] Output the adjustment priority of process parameters for defect locations, and determine the adjustment plan of process parameters based on the adjustment priority of process parameters;
[0128] Adjustments to aluminum products on the production line are implemented according to the process parameter adjustment plan, and surface defects, internal stress and corrosion activity of aluminum products are monitored in real time to determine whether the process parameter adjustment plan should be corrected.
[0129] The defect distribution map and corrosion trend line intuitively reflect the defects of aluminum products. Based on this information, we can preliminarily judge which process links have problems, and thus determine the direction of adjustment of process parameters; analyze the concentrated areas and types of defects in the defect distribution map, among which large-area scratches may be related to processing tools or processes, and observe the slope and change trend of the corrosion trend line. Among them, the rapidly rising corrosion trend may be related to insufficient surface treatment process. Based on these analysis results, determine the process parameters that need to be adjusted, such as processing parameters or surface treatment parameters; thus providing preliminary direction guidance for process parameter adjustment, and being able to quickly focus on the process links with problems, and improve the pertinence and efficiency of adjustments.
[0130] The adjustment direction determined solely based on the defect distribution map and corrosion trend line is not accurate enough. Combining the located defect causes and the mapping relationship between process parameters and defects can deeply analyze the root causes of the problems, further optimize the adjustment direction, and ensure the effectiveness of the adjustment. According to the results of defect cause location, for example, it is determined that the internal stress problem is caused by excessively high casting temperature. Combined with the quantitative relationship between casting temperature and internal stress in the mapping relationship between process parameters and defects, a more accurate casting temperature adjustment range is determined. By comprehensively considering factors such as the causal strength and impact range in the defect cause and mapping relationship, the initially determined adjustment direction is refined and optimized. This improves the accuracy and effectiveness of the process parameter adjustment direction, avoids blind adjustments, reduces the negative impact on production, and improves the success rate of product quality improvement.
[0131] Different defect locations have different degrees of impact on product quality. Determining the adjustment priority of process parameters based on the defect location can reasonably allocate resources and give priority to solving the problems that have the greatest impact on product quality. At the same time, specific adjustment plans can be formulated based on the adjustment priority to ensure that the adjustment process is carried out in an orderly manner; the impact weight is determined according to factors such as the criticality of the defect location and the size of the defect, and the defect locations are ranked according to the evaluation results to determine the adjustment priority of the process parameters. For each defect with an adjustment priority, a detailed adjustment plan is formulated in combination with the optimized adjustment direction, including the adjusted parameter values, adjustment sequence and adjustment amplitude, etc.; thereby reasonably arranging the sequence and focus of process parameter adjustments, improving resource utilization efficiency, and enabling process adjustments to more effectively solve product quality problems and improve the overall product quality.
[0132] Implement the process parameter adjustment plan on the production line, observe the changes in surface defects, internal stress and corrosion activity of aluminum products, and judge whether the adjustment plan is effective. If the effect is not ideal, revise the adjustment plan in time to ensure continuous improvement of product quality; adjust the process parameters on the production line according to the adjustment plan, and use the scale acquisition module to continuously monitor indicators such as surface defects, internal stress and corrosion activity of aluminum products after the adjustment. Compare the monitoring data with the data before adjustment and the quality standards, and use statistical analysis methods to judge whether the adjustment is effective. If it is found that the adjustment effect is not good, re-evaluate the defect mode and locate the cause of the defect, and revise the adjustment plan; thereby achieving closed-loop control of process parameter adjustment, being able to timely optimize the adjustment plan according to actual production conditions, continuously improve product quality, ensure the stability of the production process and the reliability of product quality, and continuously accumulate data and experience in the process of continuous monitoring and revision of the adjustment plan, which provides a practical basis for further optimization of process parameter adjustment strategies in the future, and promotes continuous improvement of aluminum product quality inspection and process optimization.
[0133] Example 2
[0134] like Figure 4 As shown, a method flow chart of a multi-scale detection method for aluminum product quality is provided for an embodiment of the present application. The multi-scale detection method for aluminum product quality includes:
[0135] Deploy spectral imaging and Raman spectroscopy to capture the surface of aluminum products, respectively obtaining surface defects and corrosion activity of aluminum products;
[0136] Laser ultrasound is used to scan the internal structure of aluminum products in a sparse array layout, while the internal stress of aluminum products is generated in real time using a compressed sensing reconstruction algorithm.
[0137] The surface defects, corrosion activity, and internal stress of aluminum products are used as nodes in a graph neural network to construct a correlation map. The node features of the correlation map represent the texture characteristics of surface defects, the frequency domain energy ratio of corrosion activity, and the gradient direction of internal stress.
[0138] The weights of edges in the association graph are dynamically adjusted based on the attention mechanism. The weights of edges in the association graph represent the similarity of cross-scale features.
[0139] The node features of the association graph are standardized and embedded, and the defect patterns of aluminum products are detected through convolutional neural networks.
[0140] The weights of the edges in the association graph are used as the state space of reinforcement learning to dynamically adjust the shooting frequency of the spectral imager and Raman spectrum, as well as the scanning path of the laser ultrasound.
[0141] Based on the association map, the cause of the defect of the aluminum product is located through the causal reasoning algorithm, and the defect location of the aluminum product is output. At the same time, the process parameters are adjusted according to the defect mode, defect location and defect cause of the aluminum product.
[0142] Since the principle of solving the problem by the method in the embodiment of the present application is similar to that of the system described above in the embodiment of the present application, the implementation of the method refers to the implementation of the system, and the repeated parts will not be repeated.
Claims
1. Aluminum product quality multi-scale detection system, characterized by: include: Deploy spectral imaging and Raman spectroscopy to capture the surface of aluminum products, identifying surface defects and corrosion activity. Laser ultrasound scans the internal structure of aluminum products using a sparse array layout, while a compressed sensing reconstruction algorithm generates internal stress in real time. The surface defects, corrosion activity, and internal stress of aluminum products are used as nodes in a graph neural network to construct a correlation graph. The weights of the edges in the correlation graph are dynamically adjusted based on the attention mechanism. The node features of the correlation graph represent the texture characteristics of surface defects, the frequency domain energy ratio of corrosion activity, and the gradient direction of internal stress. The edge weights in the correlation graph represent the similarity of cross-scale features. The logic for adjusting the weights of edges in the association graph includes: Determine the query vector and key vector of the attention mechanism based on the spatial attenuation factor between node features and the causal strength in the process knowledge base; Construct reinforcement learning. The state space of reinforcement learning represents the weight of the edges in the association graph, the action space represents the weight ratio of the adjustment space attenuation factor and causal strength, and the reward function represents the efficiency and accuracy of defect detection. Correct the edge weights in the association graph based on the process knowledge base, and perform targeted pruning on the initial edge connections of the association graph; The node features of the association graph are standardized and embedded, and the defect patterns of aluminum products are detected through a convolutional neural network. At the same time, the weights of the edges in the association graph are used as the state space of reinforcement learning to dynamically adjust the shooting frequencies of the spectral imager and Raman spectrum, as well as the scanning path of the laser ultrasound. The cause of the defect of the aluminum product is located through a causal reasoning algorithm based on the association graph, and the defect location of the aluminum product is output. The process parameters are simultaneously adjusted according to the defect pattern, defect location and defect cause of the aluminum product.
2. The aluminum product quality multi-scale detection system according to claim 1, characterized in that: The construction logic of the association graph includes: Mapping the surface defects, corrosion activity and internal stress of aluminum products into nodes of the association graph, and assigning a unique identifier to each node; Encode data for each node of the association graph to obtain node features; Based on the physical structure of aluminum products, the node features of surface defects are connected with the node features of internal stress, and the node features of internal stress are connected with the node features of corrosion activity according to spatial proximity to form the initial edge connection of the association graph; Add cross-layer causal edges of the association graph through the process knowledge base, and verify and adjust the node features and initial edge connections of the association graph.
3. The aluminum product quality multi-scale detection system according to claim 2, characterized in that: The detection logic of the defect mode includes: The node features after normalized embedding are input into the convolutional neural network, which outputs spatial coordinates, time stamps, and modal confidence scores. Dynamically allocate the receptive field of the convolution kernel according to the modal confidence to output a spatiotemporally separated feature map; Extract the defect distribution map from the spatial channel of the feature map, and generate the corrosion trend line from the time channel of the feature map; The defect distribution map and corrosion trend line are synchronously aligned with the process knowledge base to output the defect pattern.
4. The aluminum product quality multi-scale detection system according to claim 3, characterized in that: The adjustment logic of the laser ultrasound scanning path includes: The weights of the edges in the association graph are used as the state space of reinforcement learning, and the detection accuracy and energy consumption reduction rate are used as the reward function of reinforcement learning to explore the scanning path of laser ultrasound. During the laser ultrasonic scanning process, the temperature and vibration data of the aluminum product are obtained in real time; According to the changes in temperature data and vibration data, the state space of reinforcement learning is updated to dynamically adjust the scanning path of laser ultrasound.
5. The aluminum product quality multi-scale detection system according to claim 4, characterized in that: The logic for locating the cause of the defect includes: Combine the process knowledge base to convert the association map into a causal association map; Use counterfactual reasoning to simulate and analyze defect patterns and observe changes in defect patterns; The causal association map is integrated with the changes in defect patterns to provide evidence and output the mapping relationship between process parameters and defects to locate the cause of the defect.
6. The aluminum product quality multi-scale detection system according to claim 5, characterized in that: The output sub-logic of the mapping relationship between process parameters and defects includes: The causal relationship map and the change of defect pattern are integrated into the evidence through Bayesian network; Cross-validate the results of evidence fusion to evaluate the mapping relationship between process parameters and defects; Visualize the mapping relationship between process parameters and defects.
7. The aluminum product quality multi-scale detection system according to claim 6, characterized in that: The adjustment logic of the process parameters includes: Determine the adjustment direction of process parameters based on defect distribution map and corrosion trend line; Optimize the adjustment direction of process parameters based on the located defect causes and the mapping relationship between process parameters and defects; Output the adjustment priority of process parameters for defect locations, and determine the adjustment plan of process parameters based on the adjustment priority of process parameters; Adjustments to aluminum products on the production line are implemented according to the process parameter adjustment plan, and surface defects, internal stress and corrosion activity of aluminum products are monitored in real time to determine whether the process parameter adjustment plan should be corrected.
8. The aluminum product quality multi-scale detection system according to claim 7, characterized in that: The internal stress generation logic includes: setting dense detection points in stress-sensitive areas and setting sparse detection points in non-stress-sensitive areas to obtain a sparse array layout; The internal structure of the aluminum product is scanned by laser ultrasound in a sparse array layout, and the ultrasonic signal is converted into internal stress; Based on the historical internal stress, the internal stress of the aluminum product is generated in real time through the compressed sensing algorithm, the stress concentration area and defect location are simultaneously identified, and the sparse array layout is updated according to the error of the internal stress.
9. A multi-scale detection method for aluminum product quality, implemented based on the multi-scale detection system for aluminum product quality according to any one of claims 1 to 8, characterized in that: include: Deploy spectral imaging and Raman spectroscopy to capture the surface of aluminum products, respectively obtaining surface defects and corrosion activity of aluminum products; Laser ultrasound is used to scan the internal structure of aluminum products in a sparse array layout, while the internal stress of aluminum products is generated in real time using a compressed sensing reconstruction algorithm. The surface defects, corrosion activity, and internal stress of aluminum products are used as nodes in a graph neural network to construct a correlation map. The node features of the correlation map represent the texture characteristics of surface defects, the frequency domain energy ratio of corrosion activity, and the gradient direction of internal stress. The weights of edges in the association graph are dynamically adjusted based on the attention mechanism. The weights of edges in the association graph represent the similarity of cross-scale features. The node features of the association graph are standardized and embedded, and the defect patterns of aluminum products are detected through convolutional neural networks. The weights of the edges in the association graph are used as the state space of reinforcement learning to dynamically adjust the shooting frequency of the spectral imager and Raman spectrum, as well as the scanning path of the laser ultrasound. Based on the association map, the cause of the defect of the aluminum product is located through the causal reasoning algorithm, and the defect location of the aluminum product is output. At the same time, the process parameters are adjusted according to the defect mode, defect location and defect cause of the aluminum product.
Citation Information
Patent Citations
Non-contact laser ultrasonic damage detection method and system based on sparse array excitation
CN113916793A
Hydropower station risk real-time prediction method based on causal attention mechanism and graph learning
CN119130127A