Intelligent machining and detecting integrated system for automobile die and control method

Abstract

The invention discloses an intelligent machining and detecting integrated system for an automobile mold and a control method, and relates to the field of automobile mold machining and detecting. The multi-modal data acquisition module is used for acquiring processing and detection data by using a nanoscale sensor and terahertz equipment; the edge-cloud collaboration module processes the data in real time and completes data analysis at an edge end; the digital twinning processing module realizes micron-level precision control by predicting tool wear dynamic compensation; the multi-modal detection module fuses three-dimensional scanning and spectral data, and accurately recognizes surface and subsurface defects; and the knowledge-driven decision-making module combines reinforcement learning and a knowledge graph optimization processing strategy and stores evidence through a block chain. Through the multi-modal data fusion and digital twinning technology, the size deviation is controlled, the processing efficiency is improved, and the defect detection accuracy is high; the tool loss is reduced, the detection time is shortened, die reworking is reduced, the production quality and efficiency are remarkably improved, and the manufacturing cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automobile die machining and detection technology, and particularly relates to an automobile die intelligent machining and detection integrated system and a control method. BACKGROUND

[0002] In the automobile manufacturing industry, as core process equipment, the machining and detection quality of the die directly affects the production precision and production efficiency of automobile parts. The traditional automobile die machining adopts a combination of manual programming and numerical control machining equipment. In the machining process, problems such as tool wear and mismatched cutting parameters are difficult to monitor and adjust in real time, resulting in large fluctuations in key indicators such as die surface roughness and dimensional accuracy. When machining large cover dies, if the sudden change in cutting force is not handled in time, local overcutting or undercutting often occurs, resulting in a high scrap rate. At the same time, manual programming relies on the experience of engineers, and when faced with complex curved surface dies, the programming efficiency is low and the error rate is high, and the programming cycle of a single die can be as long as several weeks.

[0003] Existing die detection technology also faces many challenges. Traditional three-coordinate measuring machines have slow detection speeds, and it takes a long time to detect the full size of large dies, making it difficult to meet the rapid detection needs of mass production. While high-precision detection equipment such as industrial CT can identify internal defects, the cost of a single device is over 5 million yuan, and the radiation protection requirements are strict, making it difficult to be widely applied in production sites. In addition, surface defect detection is mostly done manually or with simple optical detection, which has a high rate of missing small cracks and sand holes, leading to frequent rework of dies during the die testing stage and prolonging the product development cycle.

[0004] With the increasing speed of product updates and the demand for personalized customization in the automotive industry, the limitations of traditional die machining and detection modes have become increasingly apparent. While there are some intelligent devices on the current market, there is a common problem of data silos, and there is a lack of data interoperability between machining equipment, detection instruments, and management systems, making it impossible to achieve closed-loop control of machining-detection-feedback. For example, size deviation data found through detection cannot be automatically fed back to the machining system for parameter correction, and manual reprogramming and adjustment are required, resulting in low production efficiency. At the same time, existing technologies lack deep modeling and intelligent prediction capabilities for the machining process, making it difficult to predict potential risks before machining and to achieve dynamic optimization of process parameters, making it difficult to meet the manufacturing requirements of high precision, high efficiency, and intelligence for automobile dies. SUMMARY

[0005] The automobile die intelligent machining and detection integrated system and control method proposed by the present application solve the problems mentioned in the above existing technologies.

[0006] To achieve the above purpose, the present application adopts the following technical scheme: an automobile die intelligent machining and detection integrated system and control method, comprising:

[0007] Multi-modal data acquisition module: Deploy multiple types of sensors on the spindle, tool magazine and worktable of the five-axis machining center, introduce sensor layout optimization algorithm to dynamically adjust the installation position, and use quantum key distribution technology to transmit data;

[0008] Edge-cloud collaboration module: Deploy edge computing clusters and build a hierarchical processing architecture; propose an adaptive feature extraction algorithm, use variational modal decomposition combined with convolutional attention mechanism to identify tool wear features, improve the FastPointRCNN network to detect defects in three-dimensional point cloud data, and introduce cross-modal attention mechanism to fuse laser scanning and terahertz data; Real-time data processing on the edge, use federated edge learning framework to update model parameters with the cloud;

[0009] Digital twin machining module: Build a digital twin of the mold machining, integrate the machine tool dynamics model and material removal simulation algorithm; Propose a predictive machining compensation strategy, optimize the compensation parameters through a long short-term memory network combined with a genetic algorithm, the formula is C i = f (ΔL i , ω i , α), C i is the i-th compensation amount, ΔL i is the predicted wear, ω i is the machining load coefficient, and α is the compensation coefficient; The system adapts to the machining mode and dynamically adjusts the cutting parameters according to real-time detection data;

[0010] Multi-modal detection module: Develop a data fusion platform using a cross-modal alignment network; Propose a three-dimensional topological analysis method for defects, calculate the defect connectivity index to evaluate the risk of defect expansion, L c is the length of the defect connectivity path, and L t is the total path length; Use a deformable template matching algorithm to detect dimensional deviations and generate an interactive inspection report;

[0011] Knowledge-driven decision-making module: Build an automotive mold machining knowledge graph and propose a reinforcement learning-knowledge graph joint optimization algorithm to verify feasibility, the formula is R total = β1R rl + β2R kg , R total is the total reward, R rl is the reinforcement learning reward, R kg is the knowledge graph matching reward, and β1 and β2 are weight coefficients; Establish a blockchain digital twin notarization system to store and trace the entire machining process.

[0012] Further, it also includes:

[0013] Multimodal data sensing module: Adds molecular vibration spectroscopy sensor to monitor material chemical changes in real time; Edge-cloud collaboration module develops lightweight model compression technology, compresses deep learning model parameters through knowledge distillation, and deploys them to the edge; Closed-loop optimization decision module introduces digital thread technology to connect design, processing, and testing data, enabling full lifecycle traceability and collaborative optimization.

[0014] Furthermore, it also includes:

[0015] Intelligent processing control module: Based on digital twin, it develops virtual trial cutting function to predict interference risks by simulating the processing process; Multimodal detection module integrates AI quality inspection assistant, uses natural language processing technology to analyze inspection reports and automatically generate rectification suggestions; System integration module digital mainline interacts with PLM and MES systems in real time.

[0016] Furthermore, the edge-cloud collaboration module adopts an edge autonomous architecture. When the network is interrupted, the edge continues to work based on the pre-trained model cached locally. The detection accuracy during offline periods is controlled through an asynchronous update strategy of model parameters.

[0017] Furthermore, the cross-modal alignment network in the multimodal detection module adopts a Transformer architecture combined with an attention mechanism to optimize alignment error during the fusion of data from different detection modalities.

[0018] Furthermore, the reinforcement learning-knowledge graph joint algorithm in the knowledge-driven decision-making module mines implicit relationships in the knowledge graph through graph convolutional networks, thereby optimizing the efficiency of policy generation.

[0019] Furthermore, it also includes:

[0020] Human-computer collaborative interaction module: Develop a hybrid interaction system based on gesture recognition and brain-computer interface to retrieve processing data through thought commands; integrate AR remote collaboration function to annotate mold defects in real time through AR glasses and guide on-site operation.

[0021] Furthermore, it also includes the following steps:

[0022] Data acquisition steps: Sensors synchronously acquire multi-source data at sampling frequency, and synchronize data using a timestamp alignment algorithm; Quantum key distribution technology is used to encrypt and transmit data, and the receiving end verifies data integrity through quantum state measurement; Digital twins are used to optimize sensor layout;

[0023] Collaborative processing steps: The vibration signal is subjected to variational mode decomposition at the edge to extract IMF component features; point cloud data is used to detect defects through FastPointRCNN, and the detection results are corrected by combining terahertz data; federated edge learning algorithm is used to synchronize model parameters with the cloud.

[0024] Machining control steps: Construct a digital twin of the mold machining process to simulate the material removal process in real time; use an LSTM network to predict tool wear trends and a genetic algorithm to optimize compensation parameters; when abnormal cutting force is detected, automatically switch to adaptive machining mode and adjust cutting parameters to maintain machining accuracy;

[0025] Inspection and evaluation steps: The mold is scanned simultaneously by a multi-beam structured light scanner and a terahertz device, and the two types of data are fused by CrossModalNet; the risk level is assessed by a 3D topology analysis algorithm for defects, and the dimensional deviation is detected by a deformable template matching algorithm; an AR marker inspection report is generated, marking the defect location and rectification suggestions;

[0026] Optimize decision-making steps: DRQN ​​generates processing strategies, knowledge graphs verify feasibility; calculate total reward R. total =β1R rl +β2R kg Select the optimal solution; after verification through digital twin simulation, push the strategy to the processing system, and store all data on the blockchain.

[0027] Furthermore, the collaborative processing step compresses the teacher model parameters and deploys them to the edge by distilling the knowledge of the deep learning model.

[0028] Furthermore, it also includes:

[0029] Human-machine collaborative optimization steps: Access the digital twin model via gestures or brain-computer interface to view the processing simulation results; remotely label defects via AR, and the system automatically converts suggestions into processing parameter adjustment instructions.

[0030] Compared with existing technologies, the beneficial effects of this invention are:

[0031] In terms of machining accuracy, nanoscale sensors monitor tool wear and machining vibration in real time, and combined with digital twin-driven compensation algorithms, mold size deviations are reduced, surface roughness Ra value is optimized, and mold surface quality and assembly accuracy are significantly improved.

[0032] In terms of production efficiency, the system achieves adaptive control of the machining process. When abnormal cutting force is detected, the cutting parameters are automatically adjusted to reduce downtime caused by equipment failure or improper parameters, thereby improving machining efficiency. At the same time, edge-cloud collaborative processing technology shortens the processing time of vibration signals and point cloud data, improves detection efficiency, and reduces the full-size inspection time of a single mold.

[0033] In terms of cost control and quality assurance, the system extends tool life and reduces tool replacement costs through predictive tool management. The fusion of terahertz wave and structured light detection technology significantly improves the accuracy of surface and subsurface defect detection, achieving a near-zero false negative rate, reducing mold rework frequency, and lowering the manufacturing cost per mold. Furthermore, the knowledge-driven closed-loop optimization decision module, through the combination of reinforcement learning and knowledge graphs, improves the efficiency of process parameter optimization, shortens the development cycle of new product molds, effectively enhances the company's competitiveness in the automotive mold manufacturing field, and promotes the industry's development towards intelligence and high precision. Attached Figure Description

[0034] Figure 1 This is a schematic block diagram of an integrated intelligent processing and inspection system for automotive molds proposed in this invention;

[0035] Figure 2 This is a schematic block diagram of an integrated intelligent processing and inspection control method for automotive molds proposed in this invention;

[0036] Figure 3 This is a diagram illustrating the comparison of detection accuracy for multimodal data.

[0037] Figure 4 This is a diagram illustrating the comparison of edge-cloud collaborative processing efficiency.

[0038] Figure 5 A schematic diagram showing the optimization and comparison of processing accuracy and efficiency;

[0039] Figure 6 This is a diagram comparing the time spent on defect detection with the improvement in efficiency.

[0040] Figure 7 A schematic diagram illustrating the convergence speed of the knowledge-driven optimization strategy. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0043] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0044] Reference Figures 1 to 7 A smart integrated system and control method for machining and inspecting automotive molds, comprising:

[0045] Multimodal Data Acquisition Module: In the automotive mold processing workshop, the five-axis machining center serves as the core equipment, with the spindle, tool magazine, and worktable being key sensor deployment points. The spindle is equipped with an MTI-2100 nanometer-level capacitive displacement sensor from Miitech (Germany), employing a non-contact measurement principle with a resolution of 0.1nm and a measurement range of ±250μm. It monitors tool wear in real time at a sampling frequency of 500Hz, capturing micron-level dimensional changes in the tool. The tool magazine and worktable are equipped with Analog Devices' ADXL355 MEMS triaxial accelerometer, with a range of ±50g and a frequency response range of 0-10kHz, acquiring chatter signals generated during cutting to provide data support for subsequent machining status analysis. The inspection station is equipped with an ATOS Q multi-beam structured light scanner from GOM (Germany) and a TPS3000 terahertz wave inspection device from Teraview (UK). The ATOS Q scanner employs blue light scanning technology, achieving a single-frame scanning accuracy of ±5μm to acquire 3D point cloud data of the mold surface. The TPS3000 terahertz wave inspection equipment operates in the 0.1-10THz frequency band, with a penetration depth of 0.5-5mm, detecting defects such as delamination and cracks on the subsurface of the mold. The two devices work together to achieve comprehensive mold inspection. In the sensor placement phase, digital twin technology is used to construct complete digital models of the machine tool, cutting tools, and workpiece in a virtual environment, simulating machining conditions under different cutting parameters (such as cutting speed, feed rate, and depth of cut). A genetic algorithm is used to optimize the sensor installation positions, using the detection sensitivity of key parameters (such as tool wear, cutting force, and vibration frequency) as the objective function. After 100 iterations, the sensor detection coverage of key parameters has been increased from 68% in the traditional layout to 98%. To ensure data transmission security, ID Quantique's Clavis4 quantum key distribution system is used. Based on the principle of quantum entanglement, it generates encryption keys at a rate of 1 Mbps. Through the non-cloning property of quantum states, it effectively resists quantum computer attacks and ensures the confidentiality and integrity of data during transmission.

[0046] Edge-Cloud Collaboration Module: The edge computing cluster consists of two NVIDIA Jetson AGX Orin development platforms and one Xilinx Alveo U280 FPGA accelerator card. The Jetson AGX Orin is equipped with an NVIDIA Carmel ARM 64-bit CPU and a Volta architecture GPU; the Xilinx Alveo U280 FPGA accelerator card is programmed hardware according to specific algorithm requirements. For the acquired vibration signals, Variational Mode Decomposition (VMD) is used for processing. The vibration signal is decomposed into five Intrinsic Mode Function (IMF) components, each representing vibration characteristics within a different frequency range. Convolutional Attention (CBAM) is used to filter features from the IMF components. CBAM includes channel attention and spatial attention modules, automatically identifying 12 key features such as tool wear and chatter by learning the importance weights of different feature channels and spatial locations. For 3D point cloud data processing, an improved FastPoint R-CNN network is used for defect detection. A cross-modal attention mechanism is introduced to fuse the point cloud data acquired by laser scanning with terahertz detection data. During network training, a transfer learning strategy was employed, using publicly available ShapeNet and ModelNet datasets for pre-training and fine-tuning with mold defect data collected from actual enterprise applications. After 100 rounds of iterative training, the defect recognition accuracy improved from 95.7% to 99.2%. The edge computing layer handled over 90% of the real-time data processing tasks. A federated edge learning framework was used, with model parameters synchronized with the cloud every 10 minutes. During local training, the batch size was set to 32, the initial learning rate was 0.001, and an exponential decay strategy with a decay rate of 0.99 was adopted. Redis was deployed as an edge cache database, with a data caching strategy. For frequently accessed data such as tool parameters and process templates, an LRU (Least Recently Used) eviction algorithm was used to store them locally, achieving a cache hit rate of over 95%. During network interruptions, the pre-trained model based on the local cache continued to operate at the edge, employing an asynchronous model parameter update strategy to ensure that the detection accuracy drop during offline periods did not exceed 5%.

[0047] Digital Twin Machining Module: Based on the Lagrange equation, and combining structural parameters such as machine tool mass, stiffness, and damping with kinematic relationships, a machine tool dynamics model is established. A finite element-discrete element coupled method (FEM-DEM) is used for material removal simulation. In the finite element analysis, the mold workpiece is divided into more than 100,000 elements; in the discrete element analysis, the chips during the cutting process are granularized. The simulation step size is set to 0.01s to simulate physical phenomena such as material deformation, cutting force changes, and cutting heat transfer during mold machining. A Long Short-Term Memory (LSTM) network is constructed to predict tool wear trends. Data such as collected cutting force, vibration signals, and spindle speed are used as input. After 1000 training cycles, the model is trained on actual enterprise production data, and the prediction error of tool wear is controlled within ±5%. A genetic algorithm is used to optimize the compensation parameters according to formula C. i =f(ΔL) i ,ω i ,α) Calculate the compensation amount, C i Let ΔL be the compensation amount for the i-th time. i To predict wear, ω i α is the machining load coefficient, and α is the compensation coefficient. The genetic algorithm is set to a population size of 50, a crossover probability of 0.8, and a mutation probability of 0.02. After 50 generations of evolution, the optimal compensation parameters are obtained. When the cutting force fluctuation exceeds a set threshold of ±5%, a mode switch is automatically triggered. By combining real-time detection data with simulation results from the digital twin model, cutting parameters are dynamically adjusted to ensure machining accuracy is controlled within ±3μm. The system supports virtual trial cutting based on digital twins. Before machining, the toolpath is imported into the digital twin model for simulation. By detecting interference between the tool, workpiece, and fixture, potential problems are identified in advance, reducing the traditional 2-hour trial cutting time to 15 minutes, effectively improving machining preparation efficiency.

[0048] Multimodal Detection Module: A deep learning-based multimodal detection data fusion platform was developed, employing a Transformer architecture combined with an attention mechanism for cross-modal alignment. The network encoder is configured with 6 layers, each containing a multi-head attention mechanism (8 attention heads) and a feedforward neural network. A specialized contrastive loss function is designed to optimize the alignment accuracy between different modalities. During training, a triplet loss function is used, combining 3D point cloud, CT tomographic images, and terahertz spectral data into data triplets. After 200 rounds of training, the alignment error between different modalities was reduced to 0.03 mm. A 3D defect topology analysis method is introduced, calculating the defect connectivity index. Assess the risk of defect propagation, L c L is the length of the defective connected path. tThis represents the total path length. In actual inspection, when CI > 0.7, the system automatically marks the defect as high-risk and generates a detailed risk assessment report, including information such as defect location, size, and expansion trend. Size deviation detection employs a deep learning-based deformable template matching algorithm (DeformableAlignNet). Building upon traditional template matching, it introduces deformable convolutional layers, enabling adaptive adjustment of the template shape to match complex curved surface molds. During training, 100,000 sets of mold inspection data accumulated by the company were used, including standard mold models and deviation data from actual inspections. After 150 rounds of iterative training, compared to the traditional Iterative Closest Point (ICP) algorithm, the detection efficiency is improved by 5 times and the detection accuracy by 20% in complex curved surface mold inspection. After inspection, the system automatically generates an interactive inspection report containing AR markers. With the help of AR devices such as Microsoft HoloLens 2, operators can use gesture control or voice commands to overlay and display inspection results on real molds. This includes information such as defect locations highlighted in red and dimensional deviation values ​​marked in yellow, enabling 3D visualization and interaction of inspection results. This allows operators to quickly locate problems and make repairs.

[0049] The knowledge-driven decision-making module utilizes knowledge graph technology to integrate over 1 million entities (covering process parameters, failure cases, material properties, equipment maintenance knowledge, etc.) and over 5 million relationships in the automotive mold processing field, stored and managed using the Neo4j graph database. Natural language processing (NLP) technology extracts knowledge from enterprise process documents, industry standards and specifications, and academic literature, and a knowledge graph is constructed using a combination of manual annotation and semi-supervised learning. A reinforcement learning-knowledge graph joint optimization algorithm is proposed. The deep reinforcement learning part adopts a DRQN ​​architecture, with an experience replay buffer size of 500,000 entries, and the target network is updated every 1000 steps. During optimization, DRQN ​​uses processing efficiency, product quality, and tool life as reward functions to generate initial processing strategies. Then, the feasibility of the strategies is verified using the knowledge graph. Graph Convolutional Networks (GCNs) are used to mine implicit relationships in the knowledge graph, such as the potential connection between process parameters and mold quality. Based on formula R... total =β1R rl +β2R kg Verify feasibility, R total For the total reward, R rl To enhance learning rewards, R kg The reward for matching knowledge graphs is β1 and β2, which are weight coefficients. Reasoning is performed on the knowledge graph using GCN; if the strategy violates process rules or may lead to mold quality problems, the reward is reduced. kgThis prompts the algorithm to regenerate strategies, improving the efficiency of strategy generation by 60%. A blockchain digital twin certification system based on Hyperledger Fabric is established, with a block size of 1MB, containing 10 endorsing nodes, and employing the Practical Byzantine Fault-Tolerant (PBFT) consensus algorithm to ensure data consistency and immutability. Digital twin data from the processing process, including design models, processing parameters, test results, and equipment operating status, are stored on the blockchain. Each block contains the hash value of the previous block, forming a complete blockchain data chain. This supports judicial-grade data traceability; in the event of quality disputes or process improvement needs, the entire processing data can be quickly queried and verified through the blockchain.

[0050] This invention also includes the following modules:

[0051] Multimodal data sensing module: Added molecular vibrational spectroscopy sensor (8cm resolution) －1 The system monitors material chemical changes in real time during the cutting process; the edge-cloud collaborative module develops lightweight model compression technology, which reduces deep learning model parameters by 70% through knowledge distillation and deploys them to the edge; the closed-loop optimization decision module introduces digital thread technology, which connects design, processing, and testing data into a complete digital link to achieve full lifecycle traceability and collaborative optimization.

[0052] This invention also includes the following modules:

[0053] Intelligent processing control module: Develops a virtual trial cutting function based on digital twins, which predicts interference risks by simulating the processing and reduces the trial cutting time from 2 hours to 15 minutes; Inspection and evaluation module integrates an AI quality inspection assistant, which uses natural language processing technology to analyze inspection reports and automatically generate rectification suggestions; System integration module supports a digital twin-driven digital thread, realizing real-time data interaction with PLM and MES systems (data synchronization latency <500ms).

[0054] In this invention, the edge-cloud collaborative module adopts an edge autonomous architecture. When the network is interrupted, the locally pre-trained model combines transfer and incremental learning, using publicly available and enterprise historical data for pre-training in the cloud. It is then fine-tuned using real-time mold processing data (such as vibration, temperature, and process parameters) collected at the edge, balancing versatility and scenario adaptability to ensure accurate detection and judgment during network interruptions. When the network is normal, the edge collects and preprocesses mold processing data (geometric dimensions, process parameters, equipment status, etc.), asynchronously transmitting feature vectors and gradient information to the cloud. After cloud integration and optimization, the updated edge model is asynchronously downloaded, mitigating the impact of network interruptions and allowing the model to adapt to scenario changes, resulting in an offline detection accuracy decrease of less than 5%. Simultaneously, the edge constructs a model health monitoring and self-repair mechanism, monitoring output deviations, resource usage, and other indicators in real time. In case of performance degradation or anomalies, timely intervention can be initiated, further improving the architecture's reliability and adaptability, ensuring the continuous and stable operation of mold processing detection tasks.

[0055] In this invention, the cross-modal alignment network in the multimodal detection module employs a Transformer architecture combined with an attention mechanism to capture long-distance dependencies of multi-source heterogeneous data such as 3D point clouds, CT tomographic images, and terahertz spectra. The attention mechanism is refined into intramodal and cross-modal collaboration; the former mines details of a single modality, while the latter uses multi-head attention to associate features of different modalities. During data fusion, the original multimodal data is preprocessed, such as point cloud voxelization, CT image denoising and normalization, and terahertz spectrum correction and extraction. The preprocessed data is input into the network, where the encoder performs multi-layer self-attention encoding alignment, optimizing weights to reduce differences, achieving an alignment error of 0.03mm, accurately capturing mold deviations and defects. To adapt to real-time requirements, the Transformer is lightweight and optimized, adjusting the number of attention heads and introducing hierarchical encoding to improve inference speed. Relying on edge-cloud collaborative computing power, edge preprocessing and cloud-based network operation ensure efficient and stable detection, laying a solid foundation for accurate intelligent processing and detection.

[0056] In this invention, the reinforcement learning-knowledge graph joint algorithm in the knowledge-driven decision-making module mines implicit relationships in the knowledge graph through Graph Convolutional Networks (GCNs), improving policy generation efficiency by 60%. This is achieved by constructing a knowledge graph for automotive mold processing containing a large number of entities and relationships, integrating multi-dimensional knowledge. Knowledge is stored in triples, and deep relationships require mining, making GCNs crucial. GCNs aggregate and transform neighbor node features for graph-structured data, extracting implicit features of nodes and relationships. In the knowledge graph, entities are nodes and relationships are edges. Through multiple rounds of operations, GCNs transform discrete relationships into continuous feature vectors, capturing implicit patterns in mold processing knowledge. Reinforcement learning (such as DRQN) generates processing strategies, but traditional methods involve many trial-and-error steps and are inefficient. Introducing the knowledge mined by GCNs forms a "knowledge-guided + reinforcement learning" model. GCN output features serve as state supplements or reward corrections, integrating total rewards to guide policy iteration and pruning ineffective explorations, improving policy generation efficiency and facilitating the leap from traditional to intelligent decision-making in mold processing.

[0057] This invention also includes the following modules:

[0058] Human-computer collaborative interaction module: Develop a hybrid interaction system based on gesture recognition (Leap Motion) and brain-computer interface (EEG) to support engineers in retrieving processing data through thought commands; integrate AR remote collaboration function, allowing experts to annotate mold defects in real time through AR glasses and guide on-site operations.

[0059] This invention also includes the following steps:

[0060] Data acquisition steps: Various sensors acquire data at set frequencies. A nanoscale capacitive displacement sensor acquires tool wear data at 500Hz, a MEMS triaxial accelerometer acquires vibration signals at 10kHz, a multi-beam structured light scanner performs a one-time comprehensive scan after mold fixation (scanning time approximately 3 minutes), and a terahertz wave detection device simultaneously performs subsurface detection (detection time approximately 2 minutes). The acquired data is synchronized using a timestamp alignment algorithm to ensure consistency across multiple data sources in the time dimension. Quantum key distribution (QKD) technology is used to encrypt and transmit the acquired data. Both the sending and receiving ends are equipped with quantum key generation devices, generating symmetric encryption keys through quantum entanglement. Before transmission, the data is encrypted using AES-256 with this key. Upon receiving the data, the receiving end decrypts it using the same key. Quantum state measurement technology verifies data integrity; if data tampering is detected, a retransmission is requested, ensuring data security during transmission. For sensor layout optimization, detailed simulations and adjustments are performed based on a digital twin model. Three-dimensional models of the machine tool, cutting tool, and workpiece are constructed in a virtual environment, assigning material properties and physical parameters. The physical field distribution (such as stress field, temperature field, and vibration field) under different processing conditions is simulated, and the sensor installation position is iteratively optimized using a genetic algorithm. The fitness function of the genetic algorithm is defined as the weighted sum of the detection sensitivity of key parameters. After 100 generations of evolutionary calculations, the optimal sensor layout scheme is finally determined, which enables the key parameter detection coverage to reach 98%. Compared with the traditional layout, it can acquire more comprehensive key information in the processing process.

[0061] Collaborative Processing Steps: Upon receiving the collected data, the edge computing device immediately initiates the data preprocessing process. The vibration signal undergoes mean removal and normalization, and is decomposed using the Variational Mode Decomposition (VMD) algorithm. The number of decomposition layers K is set to 5, and the penalty factor α is set to 2000, decomposing the signal into 5 IMF components. A hybrid network composed of a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (BiLSTM) network is used to extract features from the IMF components. The CNN is used to extract local features, while the BiLSTM is used to capture long-term dependencies in the time series. Features are weighted using a Convolutional Attention (CBAM) mechanism to highlight key features, achieving accurate identification of tool wear, chatter, and other conditions. For the 3D point cloud data, point cloud filtering is performed to remove noise and outliers, using a combination of voxel filtering and statistical filtering. An improved FastPointRCNN network is used for defect detection, introducing a cross-modal attention mechanism to fuse laser scanning point cloud data with terahertz detection data. During network inference, features are extracted from two modalities of data separately. Weights of the features from both modalities are calculated using an attention mechanism, and the weighted features are fused and input into the detection head for defect identification. For model updates, a federated edge learning framework is employed. The edge devices train the model on their local datasets. Every 10 training batches, the gradients of the model parameters are calculated and sent to the cloud server. The cloud server receives gradient information from multiple edge devices and aggregates the models using a federated averaging algorithm (FedAvg) to update the global model parameters. The updated global model parameters are then distributed to each edge device, which continues training with the new parameters. Simultaneously, to reduce the amount of data transmitted to the model, knowledge distillation is performed on the deep learning model. Knowledge from the teacher model (a complex model trained on a large dataset) is transferred to the student model (a lightweight model deployed at the edge), reducing the number of model parameters by 70% and decreasing inference latency from 300ms to 100ms.

[0062] Machining Control Steps: In the machining preparation stage, based on the mold design drawings and process requirements, a machine tool dynamics model is established using the Lagrange equation, considering parameters such as the mass, moment of inertia, stiffness, and damping of each machine tool component, as well as the contact force model between the tool and the workpiece. A material removal simulation model is established using the Finite Element-Discrete Element (FEM-DEM) coupling method, dividing the mold workpiece into more than 100,000 finite element elements. The chips generated during the cutting process are granulated into discrete element particles, and the simulation step size is set to 0.01s, constructing an accurate digital twin of the mold machining process. During machining, a Long Short-Term Memory (LSTM) network is used to predict tool wear trends. The input layer of the LSTM network receives time-series data such as cutting force, vibration signals, spindle speed, and feed rate. After processing through two hidden layers (128 neurons each), the output layer predicts the tool wear amount over a future period. The model is trained on actual production data from the enterprise. After 1000 training cycles, the prediction error is controlled within ±5%. The compensation parameters are optimized using a genetic algorithm. The genetic algorithm population size is set to 50, crossover probability 0.8, and mutation probability 0.02. With machining accuracy and tool life as optimization objectives, the optimal compensation parameters are obtained after 50 generations of evolutionary calculations. The system features an adaptive machining mode. Force sensors installed on the machine tool spindle and feed axis monitor cutting forces in real time. When the cutting force fluctuation exceeds a set threshold of ±5%, adaptive adjustment is automatically triggered. Real-time detection data is input into a digital twin model for simulation to predict the machining effect after adjusting the cutting parameters. Cutting parameters are dynamically adjusted based on the simulation results, such as reducing the cutting speed by 20% or the feed rate by 15%, and the adjusted parameters are sent to the machine tool control system to ensure machining accuracy is controlled within ±3μm. The system supports a virtual trial cutting function based on digital twins. Before formal machining, the tool path is imported into the digital twin model for simulated machining. By detecting interference between the tool, workpiece, and fixture, potential problems are identified in advance, reducing the traditional 2-hour trial cutting time to 15 minutes, effectively improving machining preparation efficiency.

[0063] Inspection and Evaluation Steps: After mold processing is completed, a multimodal inspection process is initiated. A multi-beam structured light scanner, using blue light scanning technology, performs an all-around scan of the mold surface with a scanning accuracy of ±5μm and a scanning time of approximately 3 minutes, acquiring 3D point cloud data of the mold surface. A terahertz wave inspection device simultaneously performs subsurface inspection of the mold, operating in the 0.1-10THz frequency band with a penetration depth of 0.5-5mm, and an inspection time of approximately 2 minutes, detecting internal defects such as delamination and cracks. A CT tomography scanner (if more detailed internal inspection is required) performs tomographic scanning of the mold, acquiring a 2D image sequence of the internal structure, with a scan layer thickness of 0.5mm and a scanning time of approximately 5 minutes. The 3D point cloud, CT tomographic images, and terahertz spectral data are fused using a cross-modal alignment network (CrossModalNet). CrossModalNet employs a Transformer architecture with a 6-layer encoder, each layer containing a multi-head attention mechanism (8 attention heads) and a feedforward neural network. During training, a contrastive learning method is employed to align identical regions in different modalities. By minimizing the contrastive loss function, the alignment error between different modalities is reduced to 0.03 mm. A 3D defect topology analysis method is used to process the fused detection data. Defect regions are extracted using image processing algorithms such as threshold segmentation and region growing, and the defect connectivity index (CI) is calculated to assess the risk of defect propagation. When CI > 0.7, the system automatically marks the defect as high-risk and generates a detailed risk assessment report.

[0064] Optimized decision-making steps: DRQN ​​incorporates long short-term memory units to process machining time-series data. After training in simulated scenarios, it outputs initial solutions adapted to different mold characteristics and stages, covering cutting parameters, path planning, and other aspects. A knowledge graph is used for feasibility verification. It uses over 1 million entities and over 5 million relationships, linking process standards, failure cases, and other knowledge. On one hand, it checks parameter compliance, judging whether strategy parameters meet safety and accuracy specifications based on mold materials and process requirements; on the other hand, it uses relational reasoning to predict risks, mining potential strategy defects from historical cases, outputting knowledge graph matching rewards, and quantifying strategy adaptability. By combining the performance rewards from reinforcement learning feedback and the knowledge graph adaptation rewards, the weights are dynamically adjusted (e.g., compliance is emphasized in trial machining, efficiency in mass production), selecting the optimal solution that balances innovation potential and risk avoidance. The solution is verified through digital twin simulation, integrating multi-physics modeling to simulate the entire machining process, outputting results such as surface error and tool wear. Once the results meet the standards, they are pushed to the machining system. Meanwhile, the entire process data is stored on the blockchain and traceable through the consortium blockchain, which lays a solid foundation of reliable data for quality traceability and process iteration, and builds a closed loop for intelligent processing decision-making.

[0065] In this invention, the collaborative processing step utilizes knowledge distillation technology to perform knowledge distillation on the deep learning model. Knowledge distillation transfers the complex knowledge of the cloud-based "teacher model" to the edge, simplifying it into a "student model." In the automotive mold scenario, the cloud-based teacher model integrates multiple networks, has tens of millions of parameters, and accurately processes multimodal data, but is difficult to deploy directly due to limited resources at the edge. The student model is optimized for edge hardware and adopts a lightweight structure. During execution, the cloud-based teacher model infers from the full set of mold data, outputting soft labels containing implicit knowledge, which are then enriched through temperature scaling. During the distillation training phase, the student model targets both soft and hard labels, learning synchronously using a loss function, and compressing the number of parameters by 70% through quantization and pruning. Previously, the teacher model's inference latency was 300ms; after compression, the student model, through hardware parallel optimization, reduces the latency to 100ms, meeting the real-time requirements of the production line, ensuring processing accuracy and efficiency, and can also be extended to scenarios such as parameter optimization and equipment maintenance, driving the evolution of manufacturing models.

[0066] This invention also includes the following steps:

[0067] Human-machine collaborative optimization steps: Operators access the digital twin model via gestures or brain-computer interface to view the processing simulation results; experts remotely collaborate via AR to annotate defects, and the system automatically converts suggestions into processing parameter adjustment instructions, achieving a 50% improvement in human-machine collaborative optimization efficiency.

[0068] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An integrated intelligent processing and inspection system for automotive molds, characterized in that, Includes the following modules: Multimodal data acquisition module: Multiple types of sensors are deployed on the spindle, tool magazine and worktable of the five-axis machining center. Sensor layout optimization algorithm is introduced to dynamically adjust the installation position and quantum key distribution technology is used to transmit data. Edge-Cloud Collaboration Module: Deploys an edge computing cluster and constructs a hierarchical processing architecture; proposes an adaptive feature extraction algorithm, adopts variational mode decomposition combined with convolutional attention mechanism to identify tool wear features, detects defects in 3D point cloud data by improving the FastPointRCNN network, and introduces a cross-modal attention mechanism to fuse laser scanning and terahertz data; processes data in real time at the edge and adopts a federated edge learning framework to jointly update model parameters with the cloud. Digital Twin Machining Module: Constructs a digital twin of mold machining, integrates machine tool dynamics model and material removal simulation algorithm to propose predictive machining compensation strategy, and optimizes compensation parameters through long short-term memory network combined with genetic algorithm, formula C. i =f(ΔL) i ω i , α), C i Let ΔL be the compensation amount for the i-th time. i To predict wear, ω i The machining load coefficient is α, and the compensation coefficient is α. The system adopts an adaptive machining mode, which dynamically adjusts the cutting parameters based on real-time detection data. The multimodal detection module develops a data fusion platform that employs cross-modal aligned network fusion. A three-dimensional topology analysis method for defects is proposed, which calculates the defect connectivity index. Assess the risk of defect propagation, L c L is the length of the defective connected path. t The total path length is used; a deformable template matching algorithm is employed to detect dimensional deviations and generate an interactive inspection report. Knowledge-driven decision-making module: Constructs a knowledge graph for automotive mold processing, proposes a joint optimization algorithm of reinforcement learning and knowledge graph, and uses formula R... total =β1R rl +β2R kg Verify feasibility, R total For the total reward, R rl To enhance learning rewards, R kg For knowledge graph matching rewards, β1 and β2 are weight coefficients; establish a blockchain digital twin certification system to process and trace the entire process of evidence storage.

2. The intelligent processing and inspection integrated system for automotive molds according to claim 1, characterized in that, Also includes: Multimodal data sensing module: Added molecular vibrational spectroscopy sensor to monitor material chemical changes in real time; The edge-cloud collaboration module develops lightweight model compression technology, compresses deep learning model parameters through knowledge distillation, and deploys them to the edge; the closed-loop optimization decision module introduces digital thread technology to connect design, processing, and testing data, enabling full lifecycle traceability and collaborative optimization.

3. The intelligent processing and testing integrated system for automotive molds according to claim 1, characterized in that, Also includes: Intelligent processing control module: Based on digital twin, it develops virtual trial cutting function to predict interference risks by simulating the processing process; Multimodal detection module integrates AI quality inspection assistant, uses natural language processing technology to analyze inspection reports and automatically generate rectification suggestions; System integration module digital mainline interacts with PLM and MES systems in real time.

4. The intelligent processing and inspection integrated system for automotive molds according to claim 1, characterized in that, The edge-cloud collaboration module adopts an edge autonomous architecture. When the network is interrupted, the edge continues to work based on the pre-trained model cached locally. The detection accuracy is controlled during offline periods through an asynchronous update strategy of model parameters.

5. The intelligent processing and inspection integrated system for automotive molds according to claim 1, characterized in that, In the multimodal detection module, the cross-modal alignment network adopts a Transformer architecture combined with an attention mechanism to optimize alignment error during the fusion of data from different detection modalities.

6. The intelligent processing and inspection integrated system for automotive molds according to claim 1, characterized in that, In the knowledge-driven decision-making module, the reinforcement learning-knowledge graph joint algorithm mines the implicit relationships in the knowledge graph through graph convolutional networks, thereby optimizing the efficiency of policy generation.

7. The intelligent processing and inspection integrated system for automotive molds according to claim 1, characterized in that, Also includes: Human-computer collaborative interaction module: Develop a hybrid interaction system based on gesture recognition and brain-computer interface to retrieve processing data through thought commands; integrate AR remote collaboration function to annotate mold defects in real time through AR glasses and guide on-site operation.

8. A method for integrated intelligent processing and inspection control of automotive molds based on the system described in any one of claims 1-7, characterized in that, Includes the following steps: Data acquisition steps: Sensors synchronously acquire multi-source data at sampling frequency, and synchronize data using a timestamp alignment algorithm; Quantum key distribution technology is used to encrypt and transmit data, and the receiving end verifies data integrity through quantum state measurement; Digital twins are used to optimize sensor layout; Collaborative processing steps: The vibration signal is subjected to variational mode decomposition at the edge to extract IMF component features; point cloud data is used to detect defects through FastPointRCNN, and the detection results are corrected by combining terahertz data; federated edge learning algorithm is used to synchronize model parameters with the cloud. Machining control steps: Construct a digital twin of the mold machining process to simulate the material removal process in real time; use an LSTM network to predict tool wear trends and a genetic algorithm to optimize compensation parameters; when abnormal cutting force is detected, automatically switch to adaptive machining mode and adjust cutting parameters to maintain machining accuracy; Detection and evaluation steps: The mold is scanned simultaneously by a multi-beam structured light scanner and a terahertz device, and the two types of data are fused by CrossModalNet; The risk level is assessed using a 3D topology analysis algorithm for defects, and dimensional deviations are detected using a deformable template matching algorithm. An AR marker inspection report is generated, indicating the location of defects and rectification suggestions. Optimize decision-making steps: DRQN ​​generates processing strategies, knowledge graphs verify feasibility; calculate total reward R. total =β1R rl +β2R kg Select the optimal solution; after verification through digital twin simulation, push the strategy to the processing system, and store all data on the blockchain.

9. The integrated intelligent processing and inspection control method for automotive molds according to claim 8, characterized in that, The collaborative processing step involves distilling knowledge from deep learning models, compressing the teacher model parameters, and deploying them to the edge.

10. The integrated intelligent processing and inspection control method for automotive molds according to claim 8, characterized in that, Also includes: Human-machine collaborative optimization steps: Access the digital twin model via gestures or brain-computer interface to view the processing simulation results; remotely label defects via AR, and the system automatically converts suggestions into processing parameter adjustment instructions.

Images