Mold structure health monitoring method and system based on artificial intelligence
By combining and optimizing the artificial intelligence network, extracting mold structure vectors and building semantic relationship spaces, the problem of inefficiency of traditional mold health monitoring methods is solved, and accurate, efficient monitoring and intelligent management of mold health status is achieved.
Patent Information
- Application Number
- CN202411028929.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-07-30
AI Technical Summary
Traditional mold health monitoring methods are inefficient, making it difficult to comprehensively and accurately evaluate fine mold damage and potential failures, lack customized optimization and multi-model collaborative working mechanisms, and inaccurate information loss and relationship characterization.
Three interrelated artificial intelligence networks are optimized by combining the combination, mold structure vectors are extracted through the first artificial intelligence network, semantic relationship space is constructed by the second artificial intelligence network, and health status estimation is carried out by the third artificial intelligence network, and reference training cost mechanism is introduced for Network parameter optimization.
It significantly improves the automation and intelligence of mold structure health monitoring, realizes accurate health status assessment, and provides data support for mold maintenance, repair and replacement.
Smart Images

Figure CN118587206B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a mold structure health monitoring method and system based on artificial intelligence. Background Art
[0002] In the field of mold manufacturing, the mold is a key component in the production process, and its health status directly affects the quality and production efficiency of the product. Traditional mold health monitoring methods often rely on manual regular inspections, which is not only inefficient, but also difficult to fully and accurately evaluate the subtle damage and potential failures of the mold.
[0003] Most existing solutions focus on the application of a single model, lacking customized optimization for mold structural characteristics and a mechanism for multi-model collaboration. In addition, these methods often have problems such as information loss, inaccurate relationship representation, or inaccurate estimation results in key links such as feature extraction, semantic relationship construction, and health status estimation.
[0004] Therefore, there is an urgent need for a more efficient, accurate and adaptable AI-based mold structure health monitoring method to solve the problems existing in the existing technology. Summary of the invention
[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a mold structure health monitoring method based on artificial intelligence, the method comprising:
[0006] Acquire target mold structure image data for health status knowledge extraction;
[0007] Performing mold structure vector representation on the target mold structure image data according to the first artificial intelligence network to generate a mold structure vector sequence;
[0008] Loading the mold structure vector sequence into a second artificial intelligence network to perform image semantic domain transformation to generate a mold structure semantic relationship space;
[0009] The mold structure health state is estimated for the mold structure semantic relationship space according to the third artificial intelligence network, and the health state knowledge data of the target mold structure image data is generated; wherein, the first artificial intelligence network, the second artificial intelligence network and the third artificial intelligence network are generated by optimizing the combined network parameters according to the reference training cost, and the reference training cost includes the first training cost and the second training cost; the first training cost is generated by calculating the feature distance between the first example mold structure semantic relationship space and the second example mold structure semantic relationship space, the first example mold structure semantic relationship space is generated by performing image semantic domain transformation on the example mold structure image after performing mold structure vector representation by the first artificial intelligence network and then performing image semantic domain transformation on the second artificial intelligence network, and the second example mold structure semantic relationship space is generated by performing image semantic domain transformation on the example image semantic features corresponding to the example mold structure image; the second training cost is generated by calculating the feature distance between the estimated health state knowledge data and the labeled health state knowledge data, and the estimated health state knowledge data is generated by performing mold structure health state estimation on the first example mold structure semantic relationship space by the third artificial intelligence network.
[0010] On the other hand, an embodiment of the present application also provides a mold visual inspection system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.
[0011] Based on the above aspects, the embodiment of the present application realizes accurate and efficient monitoring of the health status of the mold structure by combining and optimizing three interrelated artificial intelligence networks. Specifically, the first artificial intelligence network is first used to vectorize the target mold structure image, and the key structural features of the mold are effectively extracted; then, the image semantic domain is transformed through the second artificial intelligence network, and a spatial model reflecting the deep semantic relationship of the mold structure is constructed, which significantly enhances the representation ability of information; finally, the third artificial intelligence network estimates the health status based on the semantic relationship space and outputs detailed and accurate health status knowledge data. In addition, by introducing a reference training cost mechanism, the three networks are jointly optimized, wherein the first training cost focuses on improving the accuracy and consistency of the semantic representation of the mold structure, and the second training cost is directly related to the accuracy of the health status estimation. The two work together to make the network parameter adjustment more scientific and reasonable. As a result, not only the automation level and intelligence level of mold structure health monitoring are significantly improved, but also data support is provided for the maintenance, repair and replacement of the mold through accurate health status assessment. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 It is a schematic diagram of the execution flow of the mold structure health monitoring method based on artificial intelligence provided in an embodiment of the present application.
[0013] Figure 2 It is a schematic diagram of the hardware architecture of the mold visual inspection system provided in an embodiment of the present application. DETAILED DESCRIPTION
[0014] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an artificial intelligence-based mold structure health monitoring method provided in an embodiment of the present application. The artificial intelligence-based mold structure health monitoring method is introduced in detail below.
[0015] Step S110, obtaining target mold structure image data for health status knowledge extraction.
[0016] Specifically, the target mold structure image data refers to the image data of the mold that needs to be evaluated for health status, which is captured by the high-definition camera on the production line and contains various detailed information of the mold, such as shape, size, material properties, and possible defects. For example, suppose that mold A on the production line needs to be checked for health status after a round of production cycle. The high-definition camera on the production line takes photos of mold A from different angles, and these photos are the target mold structure image data. These image data are stored in the form of digital files and sent to the server through the network for processing.
[0017] That is, in this embodiment, for a certain precision mold manufacturing production line, the server, as a data processing center, is responsible for executing the complete process from image to health status knowledge. First, in step S110, the server receives the target mold structure image data captured by the camera on the precision mold manufacturing production line through the network interface. These target mold structure image data usually contain various details of the mold, such as the shape, size, material properties and possible defects (such as cracks, wear, etc.) of the mold.
[0018] In detail, suppose that mold A on the production line needs to be checked for health status after a round of production cycle. The high-definition camera on the production line captures images of mold A from multiple angles and sends these target mold structure image data to the server in the form of digital files. After the data processing system on the server receives these target mold structure image data, it automatically triggers the health status knowledge extraction process.
[0019] Step S120: performing mold structure vector representation on the target mold structure image data according to the first artificial intelligence network to generate a mold structure vector sequence.
[0020] The mold structure vector sequence refers to a vector set generated after the target mold structure image data is processed by the first artificial intelligence network. Each mold structure vector represents a local or global feature in the target mold structure image data, and the entire sequence fully reflects the structural information of the mold.
[0021] In this embodiment, after receiving the target mold structure image data, the server enters step S120 and uses a pre-trained first artificial intelligence network (assuming a convolutional neural network CNN) to perform mold structure vector representation on the target mold structure image data. The CNN network extracts key features such as edges, textures, etc. in the target mold structure image data through multi-layer convolution and pooling operations, and converts these features into high-dimensional vectors to generate a mold structure vector sequence.
[0022] In detail, the server loads the first artificial intelligence network, which has been trained on a large number of mold structure images and can accurately identify the key features of the mold. When the image data of mold A is input into the first artificial intelligence network, the first artificial intelligence network first extracts local features of the image data, such as edges and textures, through the convolution layer. Then, the dimension of the feature map is reduced through the pooling layer to reduce the amount of calculation. Finally, the fully connected layer integrates these features into a high-dimensional vector, which contains the structural information of mold A, such as shape, size, etc., forming a mold structure vector sequence.
[0023] Step S130, loading the mold structure vector sequence into a second artificial intelligence network for image semantic domain transformation to generate a mold structure semantic relationship space.
[0024] The mold structure semantic relationship space refers to a high-dimensional vector space generated after the mold structure vector sequence is processed by the second artificial intelligence network. Each dimension represents a semantic concept related to the health status of the mold structure (such as cracks, deformation, wear, etc.). The structural characteristics of the mold and their mutual relationships are expressed in a semantic form.
[0025] In step S130, the server takes the mold structure vector sequence as input and loads it into the second artificial intelligence network (assuming it is a long short-term memory network LSTM or a graph neural network GNN) for image semantic domain transformation. The function of this second artificial intelligence network is to map the mold structure vector sequence into a high-dimensional semantic relationship space, so that the mold's structural features and their mutual relationships can be expressed in a semantic form.
[0026] In detail, the LSTM or GNN model on the server receives the mold structure vector sequence from the CNN. The LSTM model captures the temporal dependencies in the mold structure vector sequence through its internal memory unit, and understands the overall layout of the mold structure and the relative position relationship between the components. The GNN model is more suitable for processing graphic data in the mold structure, such as the connection relationship between the components of the mold, and constructs a semantic graph of the mold structure through the representation of nodes and edges. Regardless of which model is used, a mold structure semantic relationship space is eventually generated. Each dimension in this mold structure semantic relationship space represents a semantic concept related to the health status of the mold structure, such as cracks, deformation, wear, etc.
[0027] Step S140, the mold structure health state is estimated for the mold structure semantic relationship space according to the third artificial intelligence network, and the health state knowledge data of the target mold structure image data is generated. The first artificial intelligence network, the second artificial intelligence network and the third artificial intelligence network are generated by optimizing the combined network parameters according to the reference training cost, and the reference training cost includes the first training cost and the second training cost. The first training cost is generated by calculating the feature distance between the first example mold structure semantic relationship space and the second example mold structure semantic relationship space. The first example mold structure semantic relationship space is generated by performing image semantic domain transformation on the example mold structure image after performing mold structure vector representation through the first artificial intelligence network and then performing image semantic domain transformation on the second artificial intelligence network. The second example mold structure semantic relationship space is generated by performing image semantic domain transformation on the example image semantic features corresponding to the example mold structure image. The second training cost is generated by calculating the feature distance between the estimated health state knowledge data and the labeled health state knowledge data. The estimated health state knowledge data is generated by the third artificial intelligence network estimating the mold structure health state of the first example mold structure semantic relationship space.
[0028] In step S140, the server takes the mold structure semantic relationship space as input and loads it into the third artificial intelligence network (assuming it is a multi-layer perceptron MLP or deep forest) to estimate the mold structure health status. This third artificial intelligence network outputs mold health status knowledge data, such as health index, defect type, defect location, etc., by learning the mapping relationship between the mold structure semantic relationship space and the health status.
[0029] In detail, the MLP or deep forest model on the server receives the mold structure semantic relationship space from the LSTM / GNN. These models learn the complex relationship between features and health status in the semantic relationship space through multiple layers of nonlinear transformation. For example, if the value of the "crack" dimension in the semantic relationship space is high, it may be judged that the mold has a crack defect. Finally, the health status knowledge data of mold A is output, including detailed information such as health index (such as 85 points, indicating that the overall condition of the mold is good but there are minor defects), defect type (such as crack), defect location (such as mold edge), etc.
[0030] It is worth noting that the first, second and third artificial intelligence networks mentioned in the above steps do not work in isolation, but work together through a combined network parameter optimization process. This optimization process is based on a reference training cost, including a first training cost and a second training cost.
[0031] For example, the server first obtains a large number of example mold structure images and their corresponding annotated health status knowledge data as a training set. Then, the first artificial intelligence network is used to represent the mold structure vector of the example mold structure image to generate an example mold structure vector sequence. Next, the second artificial intelligence network is used to map these vector sequences into the mold structure semantic relationship space to generate a first example mold structure semantic relationship space. At the same time, mold structure image recognition is performed on the example mold structure image (possibly using another pre-trained image recognition network) to generate example image semantic features, and the fourth artificial intelligence network (similar to the second artificial intelligence network but may have a different structure to ensure that the input feature dimension matches) is used to map these semantic features into another semantic relationship space to generate a second example mold structure semantic relationship space. Finally, the feature distance between the two example semantic relationship spaces is calculated to generate a first training cost. This first training cost is used to optimize the parameters of the first and second artificial intelligence networks so that they can more accurately represent the mold structure and its semantic relationship.
[0032] After generating the first example mold structure semantic relationship space, the server uses the third artificial intelligence network to estimate the mold structure health status and generate estimated health status knowledge data. Then, the feature distance between these estimated data and the labeled health status knowledge data is calculated to generate a second training cost. This second training cost is used to optimize the parameters of the third artificial intelligence network so that it can more accurately extract health status knowledge from the semantic relationship space.
[0033] The server weights the first training cost and the second training cost (different weights may be assigned according to their importance) to generate a reference training cost. Then, the first, second, and third artificial intelligence networks are jointly optimized based on this reference training cost. This optimization process is an iterative process that minimizes the reference training cost by continuously adjusting the network parameters until the preset convergence condition or iteration limit is reached.
[0034] In addition to the basic optimization process described above, the server can also adopt a variety of strategies to further improve the performance and generalization ability of the network.
[0035] For example, after generating the second example mold structure semantic relationship space, the server can further identify the semantic features of key components in the example image and determine the semantic influence relationship subspaces corresponding to these features in the semantic relationship space. Then, feature domain mapping is performed on each subspace and the first example semantic relationship space to generate a key component structure semantic relationship space. Finally, these subspaces are replaced with the corresponding key component structure semantic relationship spaces to generate a cross-domain semantic relationship space. The third artificial intelligence network is used to estimate the health state of this cross-domain semantic relationship space, and the feature distance between the annotated health state knowledge data is calculated as an additional training cost (such as the third training cost). This strategy helps the network pay more attention to the key components in the mold and improve the accuracy of health state estimation.
[0036] For another example, the server can also load the first and second example mold structure semantic relationship spaces into another network (such as the fifth artificial intelligence network) to extract semantic relationship links based on changes in the structural health state. Then, another network (such as the sixth artificial intelligence network) is used to perform attention processing on these link vectors to generate weighted semantic relationship link vectors. Additional training costs (such as the sixth training cost) are calculated by comparing the link vectors before and after processing. In addition, the third artificial intelligence network can be used to estimate the health state of the weighted link vectors, and the characteristic distance between the weighted health state knowledge data and the annotated health state knowledge data is calculated as another additional training cost (such as the seventh training cost). These additional training costs help the network better understand the inherent mechanism of changes in the health state of the mold structure.
[0037] For another example, before extracting the health status knowledge, the server can also perform surface quality recognition on the target mold structure image data. A pre-trained surface quality recognition network is used to generate surface quality semantic features, and the image semantic domain is transformed to generate the target semantic relationship space. Then, this target semantic relationship space is fused with the mold structure semantic relationship space to generate the target mold structure semantic relationship space. Finally, the third artificial intelligence network is used to estimate the health status of this fused semantic relationship space. This strategy helps to comprehensively consider the surface quality and structural health status of the mold, and improve the comprehensiveness and accuracy of health status knowledge extraction.
[0038] Based on the above steps, the embodiment of the present application realizes accurate and efficient monitoring of the health status of the mold structure by combining and optimizing three interrelated artificial intelligence networks. Specifically, the first artificial intelligence network is first used to vectorize the target mold structure image, and the key structural features of the mold are effectively extracted; then, the image semantic domain is transformed through the second artificial intelligence network, and a spatial model reflecting the deep semantic relationship of the mold structure is constructed, which significantly enhances the representation ability of information; finally, the third artificial intelligence network estimates the health status based on the semantic relationship space and outputs detailed and accurate health status knowledge data. In addition, by introducing a reference training cost mechanism, the three networks are jointly optimized, wherein the first training cost focuses on improving the accuracy and consistency of the semantic representation of the mold structure, and the second training cost is directly related to the accuracy of the health status estimation. The two work together to make the network parameter adjustment more scientific and reasonable. As a result, not only the automation level and intelligence level of mold structure health monitoring are significantly improved, but also data support is provided for the maintenance, repair and replacement of the mold through accurate health status assessment.
[0039] In a possible implementation, the combined network parameter optimization process of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network includes:
[0040] Step S101 : acquiring sample network learning data, wherein the sample network learning data includes a plurality of sample mold structure images and annotated health status knowledge data corresponding to each sample mold structure image.
[0041] In this embodiment, in order to optimize the first, second and third artificial intelligence networks used in the above process, the server first needs to obtain a large amount of sample network learning data. The sample network learning data consists of multiple example mold structure images and the annotated health status knowledge data corresponding to each image.
[0042] In detail, the server accesses the pre-prepared sample network learning data through the network interface or local storage. This sample network learning data contains hundreds of sample mold structure images of different molds. Each sample mold structure image has been carefully inspected by professionals and annotated with corresponding health status knowledge, such as health index, defect type, defect location, etc. These annotated health status knowledge data are accurate and reliable and are used for subsequent network training and parameter optimization.
[0043] Step S102, performing mold structure vector representation on the example mold structure image according to the first artificial intelligence network to generate an example mold structure vector sequence, and performing image semantic domain transformation on the example mold structure vector sequence according to the second artificial intelligence network to generate the first example mold structure semantic relationship space.
[0044] In this embodiment, for each example mold structure image, the server inputs it into the CNN network. The CNN network first extracts local features of the image, such as edges, textures, etc., through the convolution layer. Then, the dimension of the feature map is reduced through the pooling layer to reduce the amount of calculation. Finally, the fully connected layer integrates these features into a high-dimensional vector to form an example mold structure vector sequence. This example mold structure vector sequence accurately captures key features such as the shape and size of the mold.
[0045] Next, the server takes the generated example mold structure vector sequence as input, loads it into the second artificial intelligence network (LSTM or GNN), performs image semantic domain transformation, and generates the first example mold structure semantic relationship space. For example, for each example mold structure vector sequence, the server inputs it into the LSTM or GNN model. The LSTM model captures the temporal dependencies in the vector sequence through its internal memory units, understands the overall layout of the mold structure and the relative position relationship between the components. The GNN model constructs a semantic graph of the mold structure through the representation of nodes and edges, and processes the connection relationship between mold components. Regardless of which model is used, a high-dimensional first example mold structure semantic relationship space is eventually generated. Each dimension in the first example mold structure semantic relationship space represents a semantic concept related to the health status of the mold structure.
[0046] Step S103, performing mold structure image recognition on the example mold structure image to generate corresponding example image semantic features, and performing image semantic domain transformation on the example image semantic features according to the fourth artificial intelligence network to generate the second example mold structure semantic relationship space, and the image processing result of the fourth artificial intelligence network matches the input characteristic dimension of the third artificial intelligence network.
[0047] In this embodiment, the server performs mold structure image recognition on the example mold structure image, generates corresponding example image semantic features, and uses a fourth artificial intelligence network (similar to the second artificial intelligence network but may have a different structure) to perform image semantic domain transformation on these example image semantic features to generate a second example mold structure semantic relationship space.
[0048] In detail, the server uses another pre-trained image recognition network (such as ResNet) to identify the example mold structure image and extract the key semantic features in the example mold structure image. These semantic features are then input into the fourth artificial intelligence network for further semantic domain transformation. The output of this fourth artificial intelligence network matches the input feature dimension of the third artificial intelligence network to ensure the coherence of subsequent processing. Finally, a second example mold structure semantic relationship space is generated, which provides a semantic representation of the mold structure from another perspective.
[0049] Step S104: calculating a semantic dependency feature distance based on the first example mold structure semantic relationship space and the second example mold structure semantic relationship space to generate a first training cost.
[0050] In this embodiment, the server calculates the semantic dependency feature distance between the first example mold structure semantic relationship space and the second example mold structure semantic relationship space to generate a first training cost.
[0051] Specifically, the server uses a distance measurement method (such as Euclidean distance, cosine similarity, etc.) to calculate the feature distance between the semantic relationship spaces of the two examples. This feature distance reflects the difference in the semantic information of the mold structure under the two different representation methods. The server uses this feature distance as the first training cost for subsequent network parameter optimization.
[0052] Step S105, estimating the mold structure health state of the first example mold structure semantic relationship space according to the third artificial intelligence network, generating first estimated health state knowledge data, and generating a second training cost based on feature distance calculation between the first estimated health state knowledge data and the labeled health state knowledge data.
[0053] In this embodiment, the server takes the first example mold structure semantic relationship space as input, loads it into the third artificial intelligence network (MLP or deep forest), estimates the mold structure health state, and generates first estimated health state knowledge data. Then, based on the feature distance between these first estimated health state knowledge data and the labeled health state knowledge data, the second training cost is calculated.
[0054] In detail, the server inputs the first example semantic relationship space into the MLP or deep forest model, which outputs the health status knowledge data of the mold by learning the mapping relationship between the semantic relationship space and the health status. The server compares these first estimated health status knowledge data with the labeled data, calculates the feature distance between them, and generates a second training cost. This second training cost reflects the performance of the network on the health status estimation task.
[0055] Step S106: Optimize network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost and the second training cost.
[0056] The server weights the first training cost and the second training cost to generate a reference training cost, and jointly optimizes the first, second, and third artificial intelligence networks based on the reference cost. In detail, the server assigns different weights to the first training cost and the second training cost according to their importance, and performs weighted fusion to generate the final reference training cost. Then, the server uses optimization algorithms such as gradient descent to continuously adjust the parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network to minimize the reference training cost. This optimization process is an iterative process until the preset convergence condition or the number of iterations is reached. Through this process, the server can improve the performance and generalization ability of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network in processing mold structure images and health status estimation tasks.
[0057] In a possible implementation, step S106 includes:
[0058] Step S1061: obtaining a first training importance value corresponding to the first training cost, and obtaining a second training importance value corresponding to the second training cost.
[0059] Step S1062: weight-fusing the first training cost and the second training cost according to the first training importance value and the second training importance value to generate a first reference training cost.
[0060] Step S1063: Optimize network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network based on the first reference training cost.
[0061] In this embodiment, the server first needs to evaluate the first training cost to determine its importance in the entire optimization process. This importance value reflects the contribution of the first training cost to improving the overall performance of the network.
[0062] In detail, there is an optimization strategy module inside the server, which is responsible for analyzing the historical data of the first training cost and the current network status. Through analysis, the optimization strategy module can evaluate the importance of the first training cost in the current optimization stage and quantify it into a specific value - the first training importance value. This first training importance value may be calculated based on multiple factors, such as the decline rate of the first training cost, the impact on the generalization ability of the network, etc.
[0063] Similar to the first training cost, the server also needs to evaluate the second training cost to determine its importance value. For example, the optimization strategy module also analyzes the relevant data of the second training cost and considers its direct contribution to the performance of the network in the health state estimation task. Based on these factors, the module calculates the second training importance value of the second training cost. This second training importance value reflects the importance of the second training cost in improving the accuracy of health state estimation.
[0064] After obtaining the importance values of the first and second training costs, the server needs to weight the two training costs according to these importance values to generate a comprehensive training cost indicator - the first reference training cost. For example, the optimization strategy module converts the importance values of the first and second training costs into corresponding weight coefficients according to a pre-set weight allocation strategy (which may be based on experience, heuristics, or learned). Then, the optimization strategy module performs a weighted summation of the two training costs according to these weight coefficients to generate the first reference training cost. This process ensures that the server can simultaneously consider the accuracy of the mold structure representation and the performance of the health status estimation during the optimization process.
[0065] Finally, the server uses the generated first reference training cost as the optimization target to optimize the parameters of the first, second, and third artificial intelligence networks. For example, the server starts an optimization algorithm (such as gradient descent, Adam optimizer, etc.), uses the first reference training cost as the loss function, and iteratively adjusts the network parameters. In each iteration, the server calculates the first reference training cost under the current network parameters and updates the network parameters based on the gradient information. This process is automatic until the preset convergence conditions are reached (such as the training cost no longer decreases significantly, the number of iterations reaches the upper limit, etc.). Through this process, the server can gradually improve the performance of the network, making it more accurate and efficient in processing mold structure images and health status estimation tasks.
[0066] In a possible implementation, after step S103, the method further includes:
[0067] Step A110, determining key component semantic features in the example image semantic features, and determining a semantic influence relationship subspace corresponding to the key component semantic features in the second example mold structure semantic relationship space.
[0068] In this embodiment, the server first uses image recognition technology (such as a pre-trained CNN network) to process the example mold structure image to generate corresponding example image semantic features. In detail, the server receives the example mold structure image data and inputs it into the pre-trained image recognition network. The image recognition network extracts key features in the image, such as edges, textures, shapes, etc., through multi-layer convolution and pooling operations, and integrates these features into a high-dimensional vector representation, namely, the example image semantic features. These example image semantic features capture the overall and local information of the mold structure and provide a basis for subsequent processing.
[0069] In this embodiment, the server then uses a fourth artificial intelligence network (similar to the second artificial intelligence network but may have a different structure) to perform further image semantic domain transformation on the semantic features of the example image to generate a second example mold structure semantic relationship space.
[0070] In detail, the server inputs the semantic features of the example image into the fourth artificial intelligence network. The fourth artificial intelligence network converts the image semantic features into a high-dimensional semantic relationship space through nonlinear transformation and feature mapping operations. In this semantic relationship space, each dimension represents a semantic concept related to the health status of the mold structure, such as cracks, deformation, wear, etc. The second example mold structure semantic relationship space therefore contains rich semantic information of the mold structure.
[0071] The server identifies semantic features of key components in the example image semantic features, and determines semantic influence relationship subspaces corresponding to these features in the second example mold structure semantic relationship space.
[0072] Step A120 , performing feature domain mapping on each semantic influence relationship subspace and the first example mold structure semantic relationship space to generate a key component structure semantic relationship space corresponding to each key component.
[0073] In detail, the server uses an internally integrated feature selection algorithm or expert knowledge base to filter out key component semantic features that are closely related to the health status of the mold from the semantic features of the example image. These key component semantic features may represent components in the mold that are easily damaged or affect the overall performance, such as the mold core, edges, etc. Then, the server locates the areas affected by these key component semantic features in the second example semantic relationship space to form semantic influence relationship subspaces. These semantic influence relationship subspaces reflect the interaction between key components and other components or the overall structure.
[0074] In this embodiment, the server performs feature domain mapping on each semantic impact relationship subspace and the first example mold structure semantic relationship space to generate a key component structure semantic relationship space corresponding to each key component.
[0075] In detail, the server uses feature mapping techniques (such as linear transformation, nonlinear mapping, etc.) to align and fuse each semantic influence relationship subspace in the second example semantic relationship space with the first example semantic relationship space. Through this process, the server generates a specialized semantic relationship space for each key component - the key component structure semantic relationship space. These key component structure semantic relationship spaces not only contain the structural information of the key component itself, but also incorporate the interaction relationship with other components, providing a more refined perspective for health status estimation.
[0076] Step A130: Switch the semantic influence relationship subspace corresponding to each key component in the second example mold structure semantic relationship space to the corresponding key component structure semantic relationship space to generate a cross-domain semantic relationship space.
[0077] The server switches the semantic influence relationship subspace corresponding to each key component in the second example mold structure semantic relationship space to the corresponding key component structure semantic relationship space to generate a cross-domain semantic relationship space.
[0078] In detail, the server reconstructs the second example semantic relationship space, replacing the original semantic influence relationship subspace with the corresponding key component structure semantic relationship space. This process is equivalent to embedding multiple local refined semantic relationship subspaces in the global semantic relationship space to form a cross-domain semantic relationship space. This space not only retains the overall semantic information of the mold structure, but also highlights the detailed structure and interaction relationship of the key components.
[0079] Step A140, estimating the mold structure health state of the cross-domain semantic relationship space according to the third artificial intelligence network to generate second estimated health state knowledge data, and estimating the mold structure health state of the second example mold structure semantic relationship space according to the third artificial intelligence network to generate third estimated health state knowledge data.
[0080] Step A150: Calculate a third training cost based on the second estimated health status knowledge data and the third estimated health status knowledge data.
[0081] The server uses the third artificial intelligence network to estimate the mold structure health state in the cross-domain semantic relationship space to generate the second estimated health state knowledge data; at the same time, it also estimates the second example semantic relationship space to generate the third estimated health state knowledge data. Then, based on the difference between the two estimated results and the labeled health state knowledge data, the third training cost is calculated.
[0082] The server inputs the cross-domain semantic relationship space and the second example semantic relationship space into the third artificial intelligence network respectively. The third artificial intelligence network outputs the health status knowledge data of the mold by learning the mapping relationship between the semantic relationship space and the health status. The server compares these estimated data with the annotated data, calculates the feature distance or error index between them, and generates the second and third training costs respectively. The third training cost pays special attention to the degree to which the cross-domain semantic relationship space improves the accuracy of health status estimation.
[0083] Step S106 includes: optimizing network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost, the second training cost, and the third training cost.
[0084] Finally, the server optimizes the network parameters of the first, second and third artificial intelligence networks according to the combined influence of the first training cost, the second training cost and the third training cost.
[0085] The server fuses these three training costs according to certain weights (the weights may be based on historical data, expert experience, or learning algorithms) to generate the final reference training cost. Then, the server uses gradient descent or other optimization algorithms to adjust the parameters of the network according to the gradient information of the reference training cost. This process is an iterative process until the preset convergence condition or iteration limit is reached. Through this process, the server can simultaneously optimize the network's performance in multiple aspects such as mold structure representation, semantic relationship transformation, and health status estimation.
[0086] In a possible implementation, after step A140, it also includes: a fourth training cost generated by calculating the feature distance between the second estimated health status knowledge data and the labeled health status knowledge data, and a fifth training cost generated by calculating the feature distance between the third estimated health status knowledge data and the labeled health status knowledge data.
[0087] The performing network parameter optimization on the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost, the second training cost, and the third training cost includes: calculating a second reference training cost according to the first training cost, the second training cost, the third training cost, the fourth training cost, and the fifth training cost. The performing network parameter optimization on the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network based on the second reference training cost.
[0088] In this embodiment, the server first uses a third artificial intelligence network (such as a multi-layer perceptron MLP or a deep forest) to estimate the mold structure health state in the cross-domain semantic relationship space to generate second estimated health state knowledge data. At the same time, the second example mold structure semantic relationship space is also estimated to generate third estimated health state knowledge data.
[0089] In detail, the server uses the cross-domain semantic relationship space and the second example semantic relationship space as inputs to the third artificial intelligence network. The network learns the complex mapping relationship between the semantic relationship space and the health status through multiple layers of nonlinear transformation and feature mapping. Finally, the network outputs the health status knowledge data of the mold, including detailed information such as health index, defect type, and defect location. For the cross-domain semantic relationship space, the second estimated health status knowledge data is generated; for the second example semantic relationship space, the third estimated health status knowledge data is generated.
[0090] The server then performs feature distance calculation based on the second estimated health status knowledge data and the labeled health status knowledge data to generate a fourth training cost. At the same time, the server also performs feature distance calculation based on the third estimated health status knowledge data and the labeled health status knowledge data to generate a fifth training cost.
[0091] In detail, the server uses a certain distance measurement method (such as Euclidean distance, cosine similarity, etc.) to calculate the feature distances between the second estimated health status knowledge data and the third estimated health status knowledge data and the labeled data. These distances reflect the degree of difference between the estimated results and the actual annotations. The server uses these distances as training costs, named as the fourth training cost and the fifth training cost respectively. These two costs quantify the accuracy of the network in health status estimation in the cross-domain semantic relationship space and the second example semantic relationship space.
[0092] After obtaining the first training cost, the second training cost, the third training cost, the fourth training cost and the fifth training cost, the server needs to combine them to calculate the second reference training cost.
[0093] For example, there is a training cost fusion module inside the server, which is responsible for assigning different weights to each training cost according to its importance and the needs of network optimization. These weights may be based on historical data, expert experience, or learning algorithms. Then, the module performs a weighted summation of each training cost according to these weights to generate a second reference training cost. This process ensures that the server can fully consider the performance of the network in multiple aspects such as mold structure representation, semantic relationship transformation, and health status estimation during the optimization process.
[0094] Finally, the server optimizes the network parameters of the first, second, and third artificial intelligence networks based on the second reference training cost. In detail, the server starts the optimization algorithm (such as gradient descent, Adam optimizer, etc.), takes the second reference training cost as the loss function, and iteratively adjusts the network parameters. In each iteration, the server calculates the second reference training cost under the current network parameters and updates the network parameters based on the gradient information. This process is automatic until the preset convergence conditions are reached (such as the training cost no longer decreases significantly, the number of iterations reaches the upper limit, etc.). Through this process, the server can gradually improve the performance of the network, making it more accurate and efficient in processing mold structure images, generating semantic relationship spaces, and estimating health status.
[0095] In a possible implementation manner, after step S103, the method further includes:
[0096] Step B110: Loading the first example mold structure semantic relationship space into the fifth artificial intelligence network to extract semantic relationship links based on changes in structural health status to generate a first semantic relationship link vector, and loading the second example mold structure semantic relationship space into the fifth artificial intelligence network to extract semantic relationship links based on changes in structural health status to generate a second semantic relationship link vector.
[0097] Step B120, performing attention processing on the first semantic relationship link vector according to the sixth artificial intelligence network to generate a third semantic relationship link vector, and performing attention processing on the second semantic relationship link vector according to the sixth artificial intelligence network to generate a fourth semantic relationship link vector.
[0098] Step B130: Calculate a sixth training cost based on the third semantic relationship link vector and the fourth semantic relationship link vector.
[0099] The calculating of the second reference training cost based on the first training cost, the second training cost, the third training cost, the fourth training cost and the fifth training cost includes: calculating the second reference training cost based on the first training cost, the second training cost, the third training cost, the fourth training cost, the fifth training cost and the sixth training cost.
[0100] In this embodiment, the server first loads the first example mold structure semantic relationship space and the second example mold structure semantic relationship space generated by processing the example mold structure image into the fifth artificial intelligence network to extract semantic relationship links.
[0101] In detail, the fifth artificial intelligence network is a network specially designed to capture and represent semantic relationship links. It receives the first and second example semantic relationship spaces as input and analyzes the feature vectors in these spaces and their mutual relationships. Through complex network structures and algorithms, the fifth artificial intelligence network is able to identify semantic relationship links that are closely related to the changes in the health status of the mold structure and represent these links in vector form, namely the first semantic relationship link vector and the second semantic relationship link vector. These link vectors capture the dependency and change relationship between the features of the mold structure under different health states.
[0102] The server then loads the first and second semantic relationship link vectors into the sixth artificial intelligence network for attention processing to generate optimized third semantic relationship link vectors and fourth semantic relationship link vectors.
[0103] In detail, the sixth artificial intelligence network is an attention mechanism network that can automatically assign different attention weights according to the importance of the input link vector. During the processing, the sixth artificial intelligence network first analyzes each element in the link vector (i.e., each semantic relationship link) and evaluates their contribution to the estimation of the mold structure health status. Then, the network assigns corresponding attention weights to each element based on the evaluation results, and generates an optimized link vector through operations such as weighted summation. In this way, important semantic relationship links are more emphasized in the optimized link vector, while minor links are relatively weakened. In the end, the server obtained the third and fourth semantic relationship link vectors, which focus more on the key changes in the mold structure health status than the original link vectors.
[0104] The server calculates the sixth training cost based on the difference between the third semantic relationship link vector and the fourth semantic relationship link vector. In detail, the server uses a distance measurement method (such as Euclidean distance, Manhattan distance, etc.) to quantify the difference between the third and fourth semantic relationship link vectors. This difference reflects the degree of inconsistency between the two semantic relationship spaces after attention processing, and also indirectly reflects the performance of the network in extracting and processing semantic relationship links. The server uses this difference as the sixth training cost for the subsequent network parameter optimization process. The introduction of the sixth training cost helps the server to more comprehensively evaluate the network's ability to handle complex semantic relationships and promote the continuous improvement of the network in this regard.
[0105] After obtaining all the training costs including the sixth training cost, the server combines them to calculate the second reference training cost. For example, there is a training cost fusion module inside the server, which is responsible for assigning different weights to each training cost according to its importance and the needs of network optimization. These weights may be based on historical data, expert experience, or learning algorithms. Then, the module performs a weighted sum operation on each training cost according to these weights to generate a second reference training cost. This cost comprehensively considers the performance of the network in multiple aspects such as mold structure representation, semantic relationship transformation, health state estimation, and semantic relationship link processing, and provides comprehensive guidance for the optimization of the server's network parameters. By minimizing the second reference training cost, the server can continuously improve the comprehensive performance of the internal artificial intelligence network and provide more accurate and reliable technical support for the intelligent management of precision mold manufacturing production lines.
[0106] In a possible implementation manner, after step B120, the method further includes:
[0107] Step B121, estimating the mold structure health state of the third semantic relationship link vector based on the third artificial intelligence network to generate fourth estimated health state knowledge data, and estimating the mold structure health state of the fourth semantic relationship link vector based on the third artificial intelligence network to generate fifth estimated health state knowledge data.
[0108] Step B122: Calculate a seventh training cost based on the fourth estimated health status knowledge data and the fifth estimated health status knowledge data.
[0109] The calculating of the second reference training cost based on the first training cost, the second training cost, the third training cost, the fourth training cost, the fifth training cost and the sixth training cost comprises: obtaining an iteration threshold value. Calculating a third training importance value corresponding to each training cost according to the iteration threshold value. Performing weight fusion on the first training cost, the second training cost, the third training cost, the fourth training cost, the fifth training cost, the sixth training cost and the seventh training cost according to the third training importance value corresponding to each training cost to generate a second reference training cost.
[0110] In this embodiment, the server first uses the third artificial intelligence network to estimate the mold structure health state of the third semantic relationship link vector (the first semantic relationship link vector after attention processing) to generate fourth estimated health state knowledge data. At the same time, the server also estimates the fourth semantic relationship link vector (the second semantic relationship link vector after attention processing) to generate fifth estimated health state knowledge data.
[0111] In detail, the server uses the third and fourth semantic relationship link vectors as inputs to the third artificial intelligence network. The third artificial intelligence network outputs the health status knowledge data of the mold by learning the mapping relationship between the semantic relationship link and the health status. For the third semantic relationship link vector, the third artificial intelligence network outputs the fourth estimated health status knowledge data; for the fourth semantic relationship link vector, the third artificial intelligence network outputs the fifth estimated health status knowledge data. These fifth estimated health status knowledge data contain detailed information such as the health index, defect type, and defect location of the mold.
[0112] Next, the server calculates a seventh training cost based on the difference between the fourth and fifth estimated health state knowledge data and the labeled health state knowledge data. The server uses a distance metric method (such as mean square error, cross entropy, etc.) to quantify the difference between the fourth and fifth estimated health state knowledge data and the labeled health state knowledge data, and uses this difference as the seventh training cost. This cost reflects the accuracy of the network when using the optimized semantic relationship link to estimate the health state.
[0113] After obtaining all training costs including the seventh training cost, the server starts to calculate the second reference training cost. This process involves obtaining the iteration threshold value, calculating the third training importance value, and weight fusion of the training cost.
[0114] In detail, the server maintains an iteration counter internally to record the number of iterations of the current optimization process. The server also presets an iteration threshold value, which determines when to adjust the weight of the training cost. When the number of iterations reaches the threshold value, the server triggers the weight adjustment mechanism. According to the number of iterations and the iteration threshold value, the server calculates the importance of each training cost in the current iteration stage. This importance value may be based on some heuristic rules or learning algorithms, aiming to balance the impact of different training costs on the optimization process. For example, in the early iterations, the server may pay more attention to the accuracy of mold structure representation and semantic relationship transformation (i.e., the first and second training costs), while in the later iterations, it gradually increases its attention to the accuracy of health state estimation (i.e., the third to seventh training costs).
[0115] The server performs a weighted summation operation on each training cost according to the third training importance value corresponding to each training cost. This weighted summation process comprehensively considers the contribution of all training costs in the optimization process and generates a second reference training cost. The second reference training cost is a comprehensive indicator that reflects the overall performance of the network in multiple aspects. By minimizing this cost, the server can guide the network parameters to adjust in a more optimal direction. Throughout the process, comprehensive optimization of the internal artificial intelligence network is achieved through precise data processing and intelligent algorithm decision-making. This optimization not only improves the accuracy of mold structure health status estimation, but also enhances the ability of the artificial intelligence network to handle complex semantic relationships and adapt to the needs of different iteration stages.
[0116] In a possible implementation manner, after step S103, the method further includes:
[0117] Step C110, loading the first example mold structure semantic relationship space into the fifth artificial intelligence network to extract semantic relationship links based on changes in the structural health state, and loading the extracted semantic relationship link vector into the sixth artificial intelligence network for attention processing to generate a fifth semantic relationship link vector.
[0118] Step C120, loading the second example mold structure semantic relationship space into the fifth artificial intelligence network to extract semantic relationship links based on changes in the structural health state, and loading the extracted semantic relationship link vector into the sixth artificial intelligence network for attention processing to generate a sixth semantic relationship link vector.
[0119] Step C130 : calculating an eighth training cost based on the fifth semantic relationship link vector and the sixth semantic relationship link vector.
[0120] Step S106 includes: optimizing network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost, the second training cost, and the eighth training cost.
[0121] In this embodiment, the server first loads the first example mold structure semantic relationship space and the second example mold structure semantic relationship space into the fifth artificial intelligence network to extract semantic relationship links based on changes in structural health status.
[0122] In detail, the fifth artificial intelligence network is a network specially designed to capture the changes in semantic relationships of mold structures during changes in health status. It specifically receives a semantic relationship space as input, and analyzes feature vectors and their relationships in the space through a complex network structure and algorithm, identifying semantic relationship links that are closely related to changes in the health status of the structure. These links represent the dependency and evolution relationship between the features of the mold structure in different health states, which is of great significance for understanding the degradation process of the mold and predicting potential defects. The extracted semantic relationship links are represented in vector form, namely semantic relationship link vectors. The server then loads these semantic relationship link vectors into the sixth artificial intelligence network for attention processing.
[0123] The sixth artificial intelligence network is an attention mechanism network that can intelligently analyze the input semantic relationship link vector and automatically assign different attention weights according to the importance of the link. During the processing, the sixth artificial intelligence network will evaluate the contribution of each semantic relationship link to the estimation of the health status of the mold structure and assign higher weights to important links. Through operations such as weighted summation, the sixth artificial intelligence network generates optimized semantic relationship link vectors, namely the fifth semantic relationship link vector and the sixth semantic relationship link vector. These optimized vectors focus more on the key links of the change in the health status of the mold structure, providing more valuable information for subsequent health status estimation.
[0124] The server calculates the eighth training cost based on the difference between the fifth semantic relationship link vector and the sixth semantic relationship link vector. In detail, the server uses a distance measurement method (such as Euclidean distance, cosine similarity, etc.) to quantify the degree of inconsistency between the two optimized semantic relationship link vectors. This difference reflects the degree of difference between the two semantic relationship spaces after link extraction and attention processing, and also indirectly reflects the performance of the network in capturing and processing semantic relationship links. The server uses this difference as the eighth training cost for the subsequent network parameter optimization process. The introduction of the eighth training cost helps the server to more comprehensively evaluate the network's ability to process complex semantic relationship links and promote continuous improvement of the network in this regard.
[0125] After obtaining all relevant training costs including the eighth training cost, the server jointly optimizes the first, second and third artificial intelligence networks based on these costs.
[0126] In detail, there is an optimization algorithm module inside the server, which is responsible for adjusting the parameters of the network based on the first training cost (based on the feature distance between the two example semantic relationship spaces), the second training cost (based on the feature distance between the health state estimation result and the labeled data), and the eighth training cost (based on the difference in the optimized semantic relationship link vector). The optimization algorithm module will comprehensively consider the importance of these costs and their impact on network performance, and gradually adjust the network parameters through gradient descent, back propagation or other optimization techniques to minimize the sum of these costs. This optimization process is an iterative process. The server will continuously repeat the steps of semantic relationship link extraction, attention processing, cost calculation and parameter adjustment until the preset convergence condition or iteration limit is reached. Through this process, the server can continuously improve the performance of the internal artificial intelligence network in processing mold structure images, generating semantic relationship spaces, extracting semantic relationship links, and performing health state estimation, providing strong technical support for the intelligent management of precision mold manufacturing production lines.
[0127] In a possible implementation manner, before step S140, the method further includes:
[0128] Step D110 , performing surface quality recognition on the target mold structure image data to generate surface quality semantic features.
[0129] Step D120, performing image semantic domain transformation on the surface quality semantic features to generate a target semantic relationship space.
[0130] Step D130: Fusing the target semantic relationship space with the mold structure semantic relationship space to generate a target mold structure semantic relationship space.
[0131] Step S140 includes: estimating the mold structure health status of the target mold structure semantic relationship space according to the third artificial intelligence network, and generating health status knowledge data of the target mold structure image data.
[0132] In this embodiment, the server first receives the target mold structure image data, which usually include multiple angle views of the mold for comprehensive evaluation of the state of the mold.
[0133] In detail, the server uses a pre-trained surface quality recognition network (possibly a model based on a convolutional neural network (CNN)) to process the target mold structure image data. This surface quality recognition network has been extensively trained and can accurately identify various defects on the mold surface, such as scratches, pits, stains, etc. Through multi-layer convolution and pooling operations, the surface quality recognition network extracts key surface quality features in the image and represents these features as high-dimensional vectors, namely surface quality semantic features. These surface quality semantic features capture subtle changes and potential problems on the mold surface, providing important information for subsequent health status estimation.
[0134] The server then performs image semantic domain transformation on the generated surface quality semantic features to generate a target semantic relationship space. In detail, the server processes the surface quality semantic features using another pre-trained image semantic domain transformation network (which may be a model based on a graph neural network GNN or a self-attention mechanism). This network transforms the surface quality features into a high-dimensional semantic relationship space through nonlinear transformation and feature mapping operations. In this space, each dimension represents a semantic concept related to the mold surface quality, such as scratch depth, pit size, stain distribution, etc. Through this process, the server generates a target semantic relationship space, which contains rich semantic information about the mold surface quality.
[0135] Thus, the server fuses the generated target semantic relationship space with the previous mold structure semantic relationship space to generate the target mold structure semantic relationship space. In detail, there is a semantic relationship space fusion module inside the server, which is responsible for effectively integrating the semantic relationship spaces from different sources (mold structure and surface quality). This semantic relationship space fusion module first aligns the feature vectors and semantic dimensions in the two spaces to ensure that they are compared and analyzed under the same framework. Then, the module uses a fusion strategy (such as weighted summation, feature splicing, etc.) to merge the information in the two spaces to generate a comprehensive target mold structure semantic relationship space. This space not only contains the structural information of the mold (such as shape, size, component relationship, etc.), but also incorporates the surface quality information of the mold (such as scratches, pits, stains, etc.), providing more comprehensive and in-depth data support for subsequent health status estimation.
[0136] Finally, the server uses the third artificial intelligence network to estimate the health status of the mold structure in the fused semantic relationship space of the target mold structure, and generates health status knowledge data of the target mold structure image data.
[0137] In detail, the server takes the target mold structure semantic relationship space as the input of the third artificial intelligence network. This third artificial intelligence network may be a health state estimator based on a multi-layer perceptron MLP, a deep forest or other complex models. Specifically, by learning the mapping relationship between the semantic relationship space and the health state, the health state knowledge data of the mold is output. These data may include information in multiple dimensions such as health index, defect type, defect location, defect severity, etc., providing a scientific basis for the maintenance, repair or replacement of the mold. Through this process, the server not only improves the accuracy of health state estimation, but also enhances its application capability and practical value in complex industrial environments.
[0138] In a possible implementation, step S130 includes:
[0139] In step S131, the mold structure vector sequence obtained from the first artificial intelligence network is used as the input of the second artificial intelligence network, and the mold structure vector sequence captures the key features of the mold structure, including shape features, size features, and material property features.
[0140] In this embodiment, the server first uses the mold structure vector sequence obtained from the first artificial intelligence network (such as convolutional neural network CNN) as the input of the second artificial intelligence network (such as long short-term memory network LSTM or graph neural network GNN). These mold structure vector sequences are extracted from the mold structure image through multi-layer convolution and pooling operations, capturing the key features of the mold, including shape features, size features, and material property features.
[0141] Step S132: in the second artificial intelligence network, a mold structure semantic relationship space is initialized, wherein the mold structure semantic relationship space is a high-dimensional vector space, wherein each dimension in the mold structure semantic relationship space represents a semantic concept related to the health status of the mold structure, and the semantic concepts include cracks, deformation, wear, and corrosion.
[0142] After receiving the mold structure vector sequence, the server initializes a mold structure semantic relationship space inside the second artificial intelligence network. This mold structure semantic relationship space is a high-dimensional vector space used to accommodate and represent various semantic concepts related to the mold structure health status.
[0143] In detail, the server allocates an area in memory to construct the mold structure semantic relationship space. The mold structure semantic relationship space is designed to have enough dimensions to accommodate rich semantic information. Each dimension is assigned a specific semantic concept label, such as crack, deformation, wear, corrosion, etc. These concepts are closely related to the health status of the mold structure. At this time, the mold structure semantic relationship space is still empty, waiting for the mapping and filling of the mold structure vector sequence.
[0144] Step S133, using the second artificial intelligence network to map the input mold structure vector sequence to the mold structure semantic relationship space, specifically capturing the feature relationship between feature vectors in the mold structure vector sequence through nonlinear transformation, and mapping the feature relationship to the corresponding dimension of the semantic relationship space through linear transformation. In the mapping process, the relationship dependency information between the feature vectors is extracted through the second artificial intelligence network, and the relationship dependency information is represented as a vector in the semantic relationship space, the relationship dependency information includes interaction information and mutual influence information between the feature vectors, so as to generate a final mold structure semantic relationship space, each feature vector in the mold structure semantic relationship space contains the feature vector of the mold structure and the relationship dependency information between the feature vectors, and each dimension in the mold structure semantic relationship space corresponds to a semantic concept related to the health status of the mold structure.
[0145] In this embodiment, the server uses the second artificial intelligence network to map the input mold structure vector sequence into the mold structure semantic relationship space. This process involves two key steps: nonlinear transformation and linear transformation.
[0146] In detail, the server processes the input mold structure vector sequence through the nonlinear activation functions (such as ReLU, Sigmoid, etc.) in the second artificial intelligence network. These activation functions can capture the complex relationship between feature vectors in the vector sequence, such as nonlinear dependence, interaction, etc. After the nonlinear transformation, the relationship between the feature vectors becomes more obvious and prominent.
[0147] Next, the server maps the nonlinearly transformed feature relationships to the corresponding dimensions of the semantic relationship space through linear transformations (such as fully connected layers, matrix multiplication, etc.). This process is similar to projecting the feature vector into a new coordinate system, where each new coordinate axis corresponds to a semantic concept. Through linear transformations, the server can match feature relationships with semantic concepts one by one, thereby forming a clear and orderly feature representation in the semantic relationship space.
[0148] During the mapping process, the server extracts the relationship dependency information between feature vectors through the second artificial intelligence network and represents this relationship dependency information as vectors in the semantic relationship space. This relationship dependency information includes the interaction information and mutual influence information between feature vectors, which is crucial to understanding the overall health status of the mold structure.
[0149] In detail, there is a complex network structure (such as LSTM memory cells, GNN nodes and edges) inside the server to capture and represent the relationship dependencies between feature vectors. These structures analyze and extract the interactions and mutual influences between feature vectors through specific algorithms and mechanisms (such as attention mechanisms, gating mechanisms, etc.). Finally, these relationship dependency information is encoded into vector form in the semantic relationship space, and together with the feature vector of the mold structure, it constitutes a complete semantic representation. After the above steps, the server generates a final mold structure semantic relationship space. Each feature vector in this mold structure semantic relationship space contains the original feature vector of the mold structure and the relationship dependency information between these feature vectors. At the same time, each dimension in the mold structure semantic relationship space corresponds to a semantic concept related to the health status of the mold structure. That is, now, the semantic relationship space in the server is full of rich information. The vector on each dimension represents the state or feature of the mold structure under a certain semantic concept. By observing and analyzing the vector distribution and change trend in this space, the server can make accurate assessment and prediction of the health status of the mold. This semantic relationship space not only provides strong data support for subsequent health status estimation, but also provides a scientific basis for mold maintenance, repair and replacement.
[0150] Figure 2 The hardware structure of the mold visual inspection system 100 provided in the embodiment of the present application for implementing the above-mentioned artificial intelligence-based mold structure health monitoring method is shown, as shown in FIG. Figure 2 As shown, the mold vision inspection system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .
[0151] In one possible design, the mold vision inspection system 100 can be a single mold vision inspection system or a mold vision inspection system group. The mold vision inspection system group can be centralized or distributed (for example, the mold vision inspection system 100 can be a distributed system). In some embodiments, the mold vision inspection system 100 can be local or remote. For example, the mold vision inspection system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the mold vision inspection system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the mold vision inspection system 100 can be implemented on a mold vision inspection system. By way of example only, the mold vision inspection system can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.
[0152] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store data and / or instructions that the mold vision inspection system 100 uses to execute or use to complete the exemplary methods described in this application.
[0153] During the specific implementation process, at least one processor 110 executes computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the mold structure health monitoring method based on artificial intelligence in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.
[0154] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned mold vision inspection system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.
[0155] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned artificial intelligence-based mold structure health monitoring method is implemented.
[0156] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.
Claims
1. A mold structure health monitoring method based on artificial intelligence, It is characterized in that The method comprises: Acquire target mold structure image data for health status knowledge extraction; Performing mold structure vector representation on the target mold structure image data according to the first artificial intelligence network to generate a mold structure vector sequence; Loading the mold structure vector sequence into a second artificial intelligence network to perform image semantic domain transformation to generate a mold structure semantic relationship space; The mold structure health state is estimated for the mold structure semantic relationship space according to the third artificial intelligence network, and the health state knowledge data of the target mold structure image data is generated; wherein, the first artificial intelligence network, the second artificial intelligence network and the third artificial intelligence network are generated by optimizing the combined network parameters according to the reference training cost, and the reference training cost includes the first training cost and the second training cost; the first training cost is generated by calculating the feature distance between the first example mold structure semantic relationship space and the second example mold structure semantic relationship space, the first example mold structure semantic relationship space is generated by performing image semantic domain transformation on the example mold structure image after performing mold structure vector representation through the first artificial intelligence network and then performing image semantic domain transformation on the second artificial intelligence network, and the second example mold structure semantic relationship space is generated by performing image semantic domain transformation on the example image semantic features corresponding to the example mold structure image; the second training cost is generated by calculating the feature distance between the estimated health state knowledge data and the labeled health state knowledge data, and the estimated health state knowledge data is generated by the third artificial intelligence network estimating the mold structure health state of the first example mold structure semantic relationship space; Before estimating the mold structure health status of the mold structure semantic relationship space according to the third artificial intelligence network and generating the health status knowledge data of the target mold structure image data, the method further includes: Performing surface quality recognition on the target mold structure image data to generate surface quality semantic features; Performing image semantic domain transformation on the surface quality semantic features to generate a target semantic relationship space; The target semantic relationship space is merged with the mold structure semantic relationship space to generate a target mold structure semantic relationship space; The method of estimating the mold structure health status of the mold structure semantic relationship space according to the third artificial intelligence network to generate health status knowledge data of the target mold structure image data includes: Estimating the mold structure health status of the target mold structure semantic relationship space according to the third artificial intelligence network to generate health status knowledge data of the target mold structure image data; and The step of loading the mold structure vector sequence into a second artificial intelligence network for image semantic domain transformation to generate a mold structure semantic relationship space includes: Using the mold structure vector sequence obtained from the first artificial intelligence network as the input of the second artificial intelligence network, wherein the mold structure vector sequence captures key features of the mold structure, including shape features, size features, and material property features; In the second artificial intelligence network, a mold structure semantic relationship space is initialized, wherein the mold structure semantic relationship space is a high-dimensional vector space, wherein each dimension in the mold structure semantic relationship space represents a semantic concept related to the health state of the mold structure, and the semantic concepts include crack, deformation, wear, and corrosion; The second artificial intelligence network is used to map the input mold structure vector sequence to the mold structure semantic relationship space, specifically, the feature relationship between the feature vectors in the mold structure vector sequence is captured by nonlinear transformation, and the feature relationship is mapped to the corresponding dimension of the semantic relationship space by linear transformation. In the mapping process, the relationship dependency information between the feature vectors is extracted by the second artificial intelligence network, and the relationship dependency information is represented as a vector in the semantic relationship space, and the relationship dependency information includes the interaction information and mutual influence information between the feature vectors, so as to generate a final mold structure semantic relationship space, each feature vector in the mold structure semantic relationship space contains the feature vector of the mold structure and the relationship dependency information between the feature vectors, and each dimension in the mold structure semantic relationship space corresponds to a semantic concept related to the health status of the mold structure.
2. The mold structure health monitoring method based on artificial intelligence according to claim 1, It is characterized in that The combined network parameter optimization process of the first artificial intelligence network, the second artificial intelligence network and the third artificial intelligence network includes: Acquire sample network learning data, wherein the sample network learning data includes a plurality of sample mold structure images and annotated health status knowledge data corresponding to each sample mold structure image; Performing mold structure vector representation on the example mold structure image according to the first artificial intelligence network to generate an example mold structure vector sequence, and performing image semantic domain transformation on the example mold structure vector sequence according to the second artificial intelligence network to generate the first example mold structure semantic relationship space; Performing mold structure image recognition on the example mold structure image to generate corresponding example image semantic features, and performing image semantic domain transformation on the example image semantic features according to the fourth artificial intelligence network to generate the second example mold structure semantic relationship space, wherein the image processing result of the fourth artificial intelligence network matches the input characteristic dimension of the third artificial intelligence network; Calculating a semantic dependency feature distance based on the first example mold structure semantic relationship space and the second example mold structure semantic relationship space to generate a first training cost; Estimating the mold structure health state of the first example mold structure semantic relationship space according to the third artificial intelligence network to generate first estimated health state knowledge data, and performing feature distance calculation based on the first estimated health state knowledge data and the labeled health state knowledge data to generate a second training cost; Network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network are optimized according to the first training cost and the second training cost.
3. The mold structure health monitoring method based on artificial intelligence according to claim 2, It is characterized in that The performing network parameter optimization on the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost and the second training cost includes: Obtaining a first training importance value corresponding to the first training cost, and obtaining a second training importance value corresponding to the second training cost; Performing weight fusion on the first training cost and the second training cost according to the first training importance value and the second training importance value to generate a first reference training cost; Network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network are optimized based on the first reference training cost.
4. The mold structure health monitoring method based on artificial intelligence according to claim 2, It is characterized in that After performing mold structure image recognition on the example mold structure image to generate corresponding example image semantic features, and performing image semantic domain transformation on the example image semantic features according to the fourth artificial intelligence network to generate the second example mold structure semantic relationship space, the method further includes: Determine a key component semantic feature in the example image semantic feature, and determine a semantic influence relationship subspace corresponding to the key component semantic feature in the second example mold structure semantic relationship space; Performing feature domain mapping on each semantic influence relationship subspace and the first example mold structure semantic relationship space to generate a key component structure semantic relationship space corresponding to each key component; Switching the semantic influence relationship subspace corresponding to each key component in the second example mold structure semantic relationship space to the corresponding key component structure semantic relationship space to generate a cross-domain semantic relationship space; Estimating the mold structure health state of the cross-domain semantic relationship space according to the third artificial intelligence network to generate second estimated health state knowledge data, and estimating the mold structure health state of the second example mold structure semantic relationship space according to the third artificial intelligence network to generate third estimated health state knowledge data; Calculate a third training cost based on the second estimated health state knowledge data and the third estimated health state knowledge data; The performing network parameter optimization on the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost and the second training cost includes: Network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network are optimized according to the first training cost, the second training cost, and the third training cost.
5. The mold structure health monitoring method based on artificial intelligence according to claim 4, It is characterized in that After the method of estimating the mold structure health state of the cross-domain semantic relationship space according to the third artificial intelligence network to generate the second estimated health state knowledge data, and estimating the mold structure health state of the second example mold structure semantic relationship space according to the third artificial intelligence network to generate the third estimated health state knowledge data, the method further includes: a fourth training cost generated by performing feature distance calculation based on the second estimated health status knowledge data and the annotated health status knowledge data, and a fifth training cost generated by performing feature distance calculation based on the third estimated health status knowledge data and the annotated health status knowledge data; The performing network parameter optimization on the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost, the second training cost, and the third training cost includes: Calculating a second reference training cost according to the first training cost, the second training cost, the third training cost, the fourth training cost, and the fifth training cost; Network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network are optimized based on the second reference training cost.
6. The mold structure health monitoring method based on artificial intelligence according to claim 5, It is characterized in that After performing mold structure image recognition on the example mold structure image to generate corresponding example image semantic features, and performing image semantic domain transformation on the example image semantic features according to the fourth artificial intelligence network to generate the second example mold structure semantic relationship space, the method further includes: The first example mold structure semantic relationship space is loaded into the fifth artificial intelligence network to extract the semantic relationship link based on the change of the structure health state to generate a first semantic relationship link vector, and the second example mold structure semantic relationship space is loaded into the fifth artificial intelligence network to extract the semantic relationship link based on the change of the structure health state to generate a second semantic relationship link vector; Performing attention processing on the first semantic relationship link vector according to a sixth artificial intelligence network to generate a third semantic relationship link vector, and performing attention processing on the second semantic relationship link vector according to the sixth artificial intelligence network to generate a fourth semantic relationship link vector; Calculating a sixth training cost based on the third semantic relationship link vector and the fourth semantic relationship link vector; The calculating the second reference training cost according to the first training cost, the second training cost, the third training cost, the fourth training cost and the fifth training cost comprises: A second reference training cost is calculated based on the first training cost, the second training cost, the third training cost, the fourth training cost, the fifth training cost, and the sixth training cost.
7. The mold structure health monitoring method based on artificial intelligence according to claim 6, It is characterized in that After performing attention processing on the first semantic relationship link vector according to the sixth artificial intelligence network to generate a third semantic relationship link vector, and performing attention processing on the second semantic relationship link vector according to the sixth artificial intelligence network to generate a fourth semantic relationship link vector, the method further includes: performing mold structure health state estimation on the third semantic relationship link vector according to the third artificial intelligence network to generate fourth estimated health state knowledge data, and performing mold structure health state estimation on the fourth semantic relationship link vector according to the third artificial intelligence network to generate fifth estimated health state knowledge data; calculating a seventh training cost based on the fourth estimated health state knowledge data and the fifth estimated health state knowledge data; The calculating a second reference training cost based on the first training cost, the second training cost, the third training cost, the fourth training cost, the fifth training cost, and the sixth training cost comprises: Get the iteration threshold value; Calculating a third training importance value corresponding to each training cost according to the iteration threshold value; The first training cost, the second training cost, the third training cost, the fourth training cost, the fifth training cost, the sixth training cost and the seventh training cost are weighted and fused according to the third training importance value corresponding to each training cost to generate a second reference training cost.
8. The mold structure health monitoring method based on artificial intelligence according to claim 2, It is characterized in that After performing mold structure image recognition on the example mold structure image to generate corresponding example image semantic features, and performing image semantic domain transformation on the example image semantic features according to the fourth artificial intelligence network to generate the second example mold structure semantic relationship space, the method further includes: Loading the semantic relationship space of the first example mold structure into the fifth artificial intelligence network to extract semantic relationship links based on changes in the structural health state, and loading the extracted semantic relationship link vector into the sixth artificial intelligence network for attention processing to generate a fifth semantic relationship link vector; Loading the second example mold structure semantic relationship space into the fifth artificial intelligence network to extract semantic relationship links based on changes in the structural health state, and loading the extracted semantic relationship link vector into the sixth artificial intelligence network for attention processing to generate a sixth semantic relationship link vector; Calculating an eighth training cost based on the fifth semantic relationship link vector and the sixth semantic relationship link vector; The performing network parameter optimization on the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network according to the first training cost and the second training cost includes: Network parameters of the first artificial intelligence network, the second artificial intelligence network, and the third artificial intelligence network are optimized according to the first training cost, the second training cost, and the eighth training cost.
9. A mold visual inspection system, It is characterized in that The mold visual inspection system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based mold structure health monitoring method described in any one of claims 1 to 8.
