Feeding system health state monitoring method based on digital twinning and interpretable graph neural network
By combining a digital twin system with an interpretable graph neural network, a multi-source data fusion model is constructed and self-learning is achieved. This solves the problems of limited information and poor model interpretability in the health status monitoring of CNC machine tool feed systems, and enables efficient health status assessment and predictive maintenance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies for monitoring the health status of CNC machine tool feed systems suffer from limited information sources, poor model interpretability, difficulty in achieving system-wide health awareness, and a lack of self-learning and self-updating mechanisms.
By employing a digital twin system combined with an interpretable graph neural network, a health relationship graph model based on multi-source heterogeneous data is constructed. The interpretable graph neural network is then used for data fusion and identification, and a model self-learning and parameter update mechanism is established to realize the dynamic evolution of health status.
It achieves unified modeling from component-level health identification to system-level performance cognition, improving the accuracy and reliability of health assessment and supporting intelligent operation and maintenance of machine tools.
Smart Images
Figure CN121901835A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent manufacturing and health monitoring technology for high-end CNC machine tools, specifically involving a method for monitoring the health status of a feed system based on digital twins and interpretable graph neural networks. Background Technology
[0002] CNC machine tools are key equipment in modern manufacturing, and their machining quality and production efficiency are directly affected by the stability of their operating performance. The feed system, as the core unit for achieving high-precision motion and dynamic response in CNC machine tools, consists of servo motors, ball screw pairs, guide rail pairs, support structures, and control units. During long-term service, the feed system is prone to problems such as friction and wear, widened clearances, temperature drift, and servo control errors under the coupled effects of heat, force, and control factors. This leads to decreased positioning accuracy, weakened dynamic stability, and in severe cases, even system failure.
[0003] Currently, health monitoring of feed systems mainly focuses on single-signal analysis or experience-based threshold discrimination. Vibration spectrum analysis can identify some structural anomalies, but it is difficult to reflect the influence of thermal or control factors; current and torque characteristic methods can reflect changes in electrical performance, but they are insufficient for mechanical coupling responses; temperature and displacement monitoring can reveal thermal drift trends, but cannot infer the source of performance degradation. Therefore, existing technologies suffer from problems such as limited information sources, poor model interpretability, and difficulty in achieving system-wide health assessment.
[0004] On the other hand, digital twin technology in recent years has provided new technical means for machine tool condition monitoring. By establishing a virtual model and interacting with the physical system, operating condition simulation and performance prediction can be achieved. However, existing digital twin models mostly focus on geometric accuracy or thermal error compensation, and their modeling granularity remains at the system-level overall level, failing to delve into multi-source signal fusion, component coupling relationship modeling, and interpretable intelligent diagnosis. The models also lack self-learning and self-updating mechanisms, and cannot achieve cognitive evolution with changes in operating conditions.
[0005] Therefore, there is an urgent need for a multi-source data fusion and health status monitoring method for CNC machine tool feed systems, which can achieve interpretable diagnosis and adaptive learning in the context of digital twin systems, thereby improving the accuracy and reliability of health assessment and providing support for intelligent operation and maintenance of machine tools. Summary of the Invention
[0006] To address the aforementioned problems, this invention proposes a method for monitoring the health status of a feed system based on digital twins and interpretable graph neural networks. This method uses digital twins as the core carrier, collects multi-source heterogeneous data from the feed system, constructs a health relationship graph model reflecting the coupling relationships between components, and utilizes an interpretable graph neural network with an attention mechanism to achieve multi-source data fusion, health status identification, and degradation path inference. Simultaneously, a model self-learning and parameter update mechanism is established within the digital twin system to achieve dynamic evolution of health cognition. This invention enables a unified modeling framework from component-level health identification to system-level performance cognition, providing an effective technical means for the intelligent maintenance of high-end CNC machine tools.
[0007] To achieve the above objectives, the present invention adopts the following technical solution.
[0008] A method for monitoring the health status of a feed system based on digital twins and interpretable graph neural networks, comprising:
[0009] Step S110: Collect multi-source heterogeneous data and extract features;
[0010] Step S120: Based on the multi-source heterogeneous data obtained in step S110, construct a health relationship graph G=(V,E,A) driven by a hybrid mechanism and data, where V is the set of nodes, E is the set of edges, and A is the adjacency matrix with quantized edge weights.
[0011] Step S130: Construct an interpretable graph neural network model based on the health relationship graph;
[0012] Step S140: Perform health identification and degradation path inference using the trained interpretable graph neural network model;
[0013] Step S150: Execute the digital twin-driven health status monitoring and model self-learning update mechanism: Integrate an interpretable graph neural network model into the digital twin system and establish a bidirectional mapping through virtual and real data interaction; When there is a deviation between the prediction results of the digital twin system and the measured data, parameter calibration and graph structure update are automatically triggered to realize the model's self-learning and adaptive optimization, thereby forming a closed loop for feeding the system's health cognition.
[0014] A computing device includes: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method.
[0015] A readable storage medium storing program instructions that, when read and executed by a computing device, cause the computing device to perform the method.
[0016] The advantages and beneficial effects of this invention are as follows:
[0017] This invention, centered on a digital twin system, achieves unified modeling and bidirectional interaction between the physical entity and virtual model of the feed system, enabling comprehensive characterization of dynamic relationships between multi-source data. By introducing a health relationship graph structure and interpretable graph neural networks, this invention achieves hierarchical diagnosis from component-level health identification to system-level performance understanding. This invention features strong interpretability, computational efficiency, and good scalability, and can perform online self-learning and evolutionary updates based on actual operating data. This method can output system health indices and degradation path topologies, providing quantitative basis for predictive maintenance and service performance evaluation of machine tools. Attached Figure Description
[0018] Figure 1 This is a flowchart of a method for monitoring the health status of a feed system based on digital twins and interpretable graph neural networks, according to the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0020] This invention provides a method for monitoring the health status of a feed system based on digital twins and interpretable graph neural networks. Figure 1 The flowchart of the method is shown, using the feed system of a high-end CNC horizontal machining center as an example. The feed system consists of a servo motor, ball screw pair, guideway pair, support end seat, and servo control unit. During long-term operation of the machine tool, frictional wear, thermal drift, and control loop deviations can lead to performance degradation. To achieve precise perception and intelligent maintenance of the system, this invention uses a digital twin system as its core to realize multi-source data fusion, structured health relationship modeling, interpretable graph neural network diagnosis, and self-learning optimization. Figure 1 As shown, the specific implementation method is as follows:
[0021] Step S110: Collect multi-source heterogeneous data and extract features. Signals are collected from multiple sensor channels of the horizontal machining center's feed system, and a feature matrix reflecting the mechanical, electrical, thermal, and control states is constructed to provide input data for health relationship modeling.
[0022] Step S110-1: Acquire vibration, electrical, thermal, control, and environmental signals from the feed system. Sensors used to acquire these signals include a triaxial accelerometer, thermocouple temperature sensors, current and voltage detection modules, position sensors, and ambient temperature and humidity sensors. Specifically, this includes: installing a triaxial accelerometer at the lead screw support end and guide rail slider to capture vibration signals; arranging thermocouples in the servo motor housing and the middle section of the lead screw to measure temperature changes; acquiring three-phase current, voltage, torque, and speed signals on the servo driver side; recording speed commands, position errors, and feedback positions in the CNC system; and acquiring temperature, humidity, and external disturbance signals in the machining environment.
[0023] Step S110-2: Perform time synchronization, filtering and denoising, normalization, and feature extraction on the acquired signal. Extracted signal features include root mean square (RMS), spectral kurtosis, temperature rise rate, and mean servo error. Construct a fused feature vector at sampling time t.
[0024] ,
[0025] in, These represent vibration, electrical, thermal, control, and environmental characteristics, respectively, with d being the feature dimension; continuous sampling constitutes the feature matrix X=[x1,x2,...,x...]. T ].
[0026] The signals are synchronized by a unified acquisition system and subjected to filtering, noise reduction, normalization, and interpolation correction to ensure consistency of data from different sources. Subsequently, statistical and spectral features reflecting the operating status are extracted from various signals, including root mean square value, peak energy, envelope entropy, temperature rise rate, torque fluctuation coefficient, and mean servo error, for use in constructing subsequent health relationship diagrams. This step achieves time synchronization and feature alignment of multiple signals, enabling the unified modeling of different physical characteristics of the system under the same time reference.
[0027] Step S120: Based on the multi-source heterogeneous data obtained in step S110, construct a health relationship graph G=(V, E, A) driven by a hybrid mechanism and data approach, where V is the set of nodes, E is the set of edges, and A is the adjacency matrix with quantified edge weights. This graph structure describes the coupling and health dependencies between components of the horizontal machining center's feed system, realizing the mapping and association between data and structure.
[0028] Step S120-1: Based on the structural topology of the feed system, determine the node set V={v1,v2,...,v...} of the health relationship graph. n Each node represents a physical component or monitoring location, including motors, lead screws, nuts, guide rails, supports, and environmental nodes. Node features are assigned values from the corresponding feature vectors obtained in step S110. The edge set E={e} is determined. ij} represents node v iWith v j There is a health link between them.
[0029] Step S120-2: Establish the connection relationship between nodes and calculate the edge weight w ij Edge weights reflect the strength of healthy coupling between nodes, and their calculation formula is as follows:
[0030] ,
[0031] in, For v i With v j The physical coupling coefficient, including force transmission, heat conduction, control feedback coupling, and environmental coupling, is determined based on the mechanical connection, heat conduction path, and control loop relationship. The statistical correlation between signals can be calculated using the coherence coefficient; These are the weighting coefficients.
[0032] Step S120-3: Put all w ij Form an adjacency matrix A = [w ij ], used as the structural input for interpretable graph neural networks. The adjacency matrix A = [w] is formed by calculating all edge weights. ij The adjacency matrix not only reflects the static coupling relationship between components, but can also be dynamically updated according to the operating conditions, thereby capturing the changing trend of the health status of the feed system.
[0033] In this model, node features reflect local health status, edge weights express coupling strength, and the entire graph structure comprehensively describes system-level health relationships, providing physical and semantic support for subsequent interpretable graph neural network training.
[0034] Based on the structural topology and energy and signal transmission paths of the feeding system, a system health relationship diagram is established with components and functional modules as nodes. By integrating physical coupling information and signal correlation characteristics, the connection weights between nodes are calculated to generate an adjacency matrix reflecting the health coupling strength between components, thus realizing the representation of the system-level health structure.
[0035] Step S130: Construct an interpretable graph neural network model based on the health relationship graph. Establish an interpretable graph neural network model to achieve health identification and feature contribution analysis.
[0036] Step S130-1: Using the health relationship graph G as input, construct an interpretable graph neural network model structure. The model includes an input layer, a graph convolutional layer, an attention layer, a feature contribution layer, and an output layer. The input layer receives the health relationship graph G. The graph convolutional layer extracts local structural features by aggregating information from neighboring nodes. The attention layer uses learnable parameters to calculate the influence weights between nodes, constructing an attention weight matrix to identify key coupling paths. The feature contribution layer analyzes the degree of influence of different input features on the prediction results. The output layer generates the health probability distribution of each node using the Softmax function.
[0037] Step S130-2: The loss function of the model is defined as:
[0038] ,
[0039] in, This is the node health classification error term; This is an attention distribution constraint term used to maintain physical consistency; The weighting coefficients are used for balancing. By optimizing the weighting parameters through backpropagation, a balance between health recognition accuracy and interpretability is achieved, resulting in a health recognition model with interpretability.
[0040] During the training phase, historical data samples from a horizontal machining center under various operating conditions were selected for learning. The training samples covered normal, slightly degraded, and abnormal operating conditions. Through multiple rounds of iterative optimization, the model was able to output both healthy classification results and provide interpretable analysis of features and paths.
[0041] Step S140: Perform health identification and degradation path inference using the trained interpretable graph neural network model.
[0042] Step S140-1: Input the real-time collected feed system data into the trained model, and output the health probability of each node and the system health index, as well as the health probability vector of each node. The three terms represent the probabilities of healthy, degraded, and faulty states, respectively; the system health index is calculated as follows:
[0043] ,
[0044] Where N represents the total number of nodes. When the health index drops below the threshold, the system automatically generates a health warning. The health index threshold can be dynamically adjusted according to the type of feed system and historical operating conditions to adapt to the health assessment needs under different environments.
[0045] Step S140-2: Extract the set of edges with higher weights based on the attention weight matrix to construct a degradation propagation topology graph. The weights reflect the correlation between health changes between nodes. When the weight of a connecting edge continuously increases, it indicates that the health change may propagate along that path. When the motor temperature rise causes an increase in the vibration amplitude at the lead screw end and the weight between two nodes increases, it can be identified as a propagation path of thermal degradation. This step demonstrates the causal chain of "starting component - propagation path - affected component," which is used for degradation source localization and propagation trend analysis. The results are visualized in the digital twin platform to assist in maintenance decision-making.
[0046] Step S150: Execute the digital twin-driven health status monitoring and model self-learning update mechanism: Integrate an interpretable graph neural network model into the digital twin system and establish a bidirectional mapping through virtual and real data interaction. When there is a deviation between the prediction results of the digital twin system and the measured data, parameter calibration and graph structure update are automatically triggered to achieve self-learning and adaptive optimization of the model, thereby forming a closed loop for feeding the system's health cognition.
[0047] Step S150-1: Achieve bidirectional interaction between the digital twin and the physical world. The digital twin system consists of a physical layer, a data layer, a model layer, and a virtual layer. The physical layer collects real-time operating signals from the feed system of the horizontal machining center; the data layer is responsible for cleaning and synchronization; the model layer runs an interpretable graph neural network to achieve health recognition; and the virtual layer compares the predicted results with the measured data.
[0048] Step S150-2: Establish a self-learning update mechanism. When predicting the health index HI... pred Compared with the measured health index HI meas The deviation exceeds the threshold When the system automatically triggers a model update, the update process includes: updating the edge weights of the adjacency matrix of the health relationship graph, updating the network parameters of the interpretable graph neural network, updating the signal feature extraction parameters, and using new samples for incremental training to fine-tune the network weights so that the model adapts to the new operating conditions.
[0049] Step S150-3: Output health assessment results. The digital twin system outputs node health levels, system health indices, and degradation path topology in real time, providing maintenance personnel with structured decision-making support, including adjusting lubrication plans, optimizing feed rates, or performing calibration operations. During operation, the digital twin system continuously outputs system health indices, node health levels, and degradation propagation path information to guide lubrication work, feed system parameter optimization, and maintenance cycle adjustments.
[0050] Through this mechanism, the digital twin system can achieve self-learning evolution of the health model during actual service, ensuring the accuracy and robustness of health identification results.
[0051] In summary, this invention proposes a health status monitoring method for feed systems based on digital twins and interpretable graph neural networks, establishing a complete technical system comprised of multimodal perception, structural modeling, intelligent diagnosis, and self-learning evolution. This method achieves explicit expression of coupling relationships between components through a health relationship graph model, deep fusion and traceable reasoning of multi-source data through interpretable graph neural networks, and continuous model updates and cognitive optimization through the virtual-real interaction of the digital twin system. Compared with existing technologies, this invention can accurately identify degradation characteristics and track health evolution paths under complex coupling and time-varying operating conditions of feed systems, significantly improving the scientific and intelligent level of machine tool service performance cognition and predictive maintenance.
[0052] The present invention also provides a computing device, comprising: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method described thereon.
[0053] The present invention also provides a readable storage medium storing program instructions, which, when read and executed by a computing device, cause the computing device to perform the method described thereon.
[0054] The contents not described in detail in this specification are existing technologies known to those skilled in the art.
[0055] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered to fall within the protection scope of the present invention.
Claims
1. A method for monitoring the health status of a feed system based on digital twins and interpretable graph neural networks, characterized in that, include: Step S110: Collect multi-source heterogeneous data and extract features; Step S120: Based on the multi-source heterogeneous data obtained in step S110, construct a health relationship graph G=(V,E,A) driven by a hybrid mechanism and data, where V is the set of nodes, E is the set of edges, and A is the adjacency matrix with quantized edge weights. Step S130: Construct an interpretable graph neural network model based on the health relationship graph; Step S140: Perform health identification and degradation path inference using the trained interpretable graph neural network model; Step S150: Execute the digital twin-driven health status monitoring and model self-learning update mechanism: Integrate an interpretable graph neural network model into the digital twin system and establish a bidirectional mapping through virtual and real data interaction; When there is a deviation between the prediction results of the digital twin system and the measured data, parameter calibration and graph structure update are automatically triggered to realize the model's self-learning and adaptive optimization, thereby forming a closed loop for feeding the system's health cognition.
2. The method according to claim 1, characterized in that, Step S110 includes: Step S110-1: Collect vibration, electrical, thermal, control, and environmental signals of the feed system; Step S110-2: Perform time synchronization, filtering and denoising, normalization, and feature extraction on the acquired signal. Extracted signal features include root mean square (RMS), spectral kurtosis, temperature rise rate, and mean servo error. Construct a fused feature vector at sampling time t. ， in, These represent vibration, electrical, thermal, control, and environmental characteristics, respectively, with d being the feature dimension; continuous sampling constitutes the feature matrix X=[x1,x2,...,x...]. T ].
3. The method according to claim 2, characterized in that, In step S110-1, a triaxial acceleration sensor is installed at the lead screw support end and the guide rail slider to capture vibration signals; a thermocouple is arranged in the servo motor housing and the middle section of the lead screw to measure temperature changes; three-phase current, voltage, torque and speed signals are collected on the servo driver side; speed commands, position errors and feedback positions are recorded in the CNC system; and temperature, humidity and external disturbance signals are collected in the machining environment.
4. The method according to claim 1, characterized in that, Step S120 includes: Step S120-1: Based on the topology of the feed system structure, determine the node set V={v1,v2,...,v...} of the health relationship graph. n Each node represents a system component or monitoring location, including motors, lead screws, nuts, guide rails, supports, and environmental nodes; node features are assigned values by the corresponding feature vectors obtained in step S110; the edge set E={e} is determined. ij } represents node v i With v j There is a health link between them; Step S120-2: Establish the connection relationship between nodes and calculate the edge weight w ij Edge weights reflect the strength of healthy coupling between nodes, and their calculation formula is as follows: ， in, Indicates component v i With v j The degree of physical coupling, including force transmission, heat conduction, control feedback coupling, and environmental coupling; Indicates the statistical correlation between signals; These are weighting coefficients; Step S120-3: Put all w ij Form an adjacency matrix A = [w ij ], used as the structural input for interpretable graph neural networks.
5. The method according to claim 1, characterized in that, Step S130 includes: Step S130-1: Using the health relationship graph G as input, construct an interpretable graph neural network model structure, including an input layer, a graph convolutional layer, an attention layer, a feature contribution layer, and an output layer; the input layer receives the health relationship graph G; the graph convolutional layer aggregates neighborhood information; the attention layer calculates the influence weights between nodes through learnable parameters and constructs an attention weight matrix; the feature contribution layer evaluates the influence of features on the output; the output layer outputs the health status probability of the nodes. Step S130-2: The loss function of the model is defined as: ， in, For node health classification error, For attention distribution constraints, The weights are used as balance coefficients; by optimizing the weight parameters of the interpretable graph neural network model through backpropagation, a health recognition model with interpretability is obtained.
6. The method according to claim 1, characterized in that, Step S140 includes: Step S140-1: Input the real-time collected feed system data into the trained model, and output the health probability of each node and the system health index, as well as the health probability vector of each node. The three terms represent the probabilities of healthy, degraded, and faulty states, respectively; the system health index is calculated as follows: ， Where N is the total number of nodes; Step S140-2: Extract the set of edges with higher weights based on the attention weight matrix, construct a degradation propagation topology graph, and display the causal chain of "starting component - propagation path - affected component" for degradation source localization and propagation trend analysis.
7. The method according to claim 1, characterized in that, Step S150 includes: Step S150-1: Realize two-way interaction between the digital twin and the real world. The digital twin system consists of a physical layer, a data layer, a model layer and a virtual layer. The physical layer collects real-time data into the system, the data layer is responsible for cleaning and synchronization, the model layer performs health recognition based on an interpretable graph neural network, and the virtual layer compares the prediction results with the measured data. Step S150-2: Establish a self-learning update mechanism; when predicting the health index HI pred Compared with the measured health index HI meas The deviation exceeds the threshold When the model is updated, it automatically triggers the following updates: updating the edge weights of the adjacency matrix of the health relationship graph, updating the network parameters of the interpretable graph neural network, and updating the signal feature extraction parameters, thereby achieving the adaptive evolution of the digital twin model. Step S150-3: Output health cognition results: The digital twin system outputs the node health level, system health index and degradation propagation path in real time, providing a quantitative basis for maintenance decisions.
8. The method according to claim 6, characterized in that, Step S140-1 includes: when the system health index drops below the health index threshold, the system automatically generates a health warning; the health index threshold is dynamically adjusted according to the type of the feeding system and historical operating conditions to adapt to the health judgment requirements under different environments.
9. A computing device, characterized in that, include: At least one processor and a memory storing program instructions; When the program instructions are read and executed by the processor, the computing device performs the method as described in any one of claims 1-8.
10. A readable storage medium storing program instructions, characterized in that, When the program instructions are read and executed by the computing device, the computing device performs the method as described in any one of claims 1-8.