Method and system for evaluating health state of vertical numerical control machine tool

Abstract

The invention relates to the technical field of machine learning, and discloses a vertical numerical control machine tool health state evaluation method and system. The method comprises the following steps: synchronously acquiring machine tool operation condition data and a multi-mode sensing signal; analyzing and generating a working condition feature vector, and performing time-frequency processing on the sensing signal to form a real-time state feature vector; based on a pre-trained machine tool part relational graph neural network model, generating a prediction state feature vector matched with the current working condition; the residual error between the real-time vector and the prediction vector is calculated, and the mahalanobis distance is solved by combining the health residual error covariance matrix to serve as a dynamic health evaluation value; and early warning is triggered according to the multi-level threshold value, and the residual component is analyzed to position the abnormal component during overrun. The system comprises a synchronous acquisition module, a feature processing module, a dynamic reference generation module, a health quantitative evaluation module and an early warning diagnosis module. According to the method, the evaluation accuracy, the early fault sensitivity and the fault positioning capability are improved through working condition self-adaptive dynamic reference and graph neural network modeling.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to a method and system for evaluating the health status of vertical CNC machine tools. Background Technology

[0002] As high-end manufacturing demands ever-increasing precision and equipment reliability, the real-time health assessment of vertical CNC machine tools, as core equipment, has become a crucial aspect of intelligent manufacturing systems. Traditional health monitoring methods primarily rely on sensor signals such as vibration, noise, and temperature, and set uniform, fixed thresholds for anomaly detection. These methods assume that the machine tool has a static "normal" baseline under all operating conditions, ignoring the highly dynamic nature of the actual machining process.

[0003] In fact, the physical signal characteristics of machine tools during normal operation vary under different load intensities, different tool types, and different combinations of materials being machined. For example, the vibration amplitude under heavy cutting conditions may be much greater than that under finishing conditions, but it is still within the range of healthy operation.

[0004] The core of health status evaluation for vertical CNC machine tools lies in accurately identifying early abnormal signs that deviate from the normal behavior pattern under the current operating conditions. Existing technologies generally adopt threshold alarm mechanisms based on rules or simple statistical models, lacking the ability to model the coupled effects of multi-dimensional operating conditions.

[0005] Since the modulation effect of changes in operating conditions on the normal signal distribution is not considered, the fixed threshold strategy is very likely to misjudge reasonable responses under high load as fault precursors, resulting in a large number of false alarms; conversely, under light load or low-sensitivity operating conditions, weak but destructive abnormal signals may be missed because they are below the threshold.

[0006] This situation of both false alarms and missed alarms severely undermines the credibility and practicality of the early warning system, leading to inaccurate maintenance decisions, wasting human resources, and potentially delaying intervention in critical faults.

[0007] Existing technologies generally suffer from weak model generalization ability and poor adaptability when dealing with the contradiction between the diversity of operating conditions and the dynamic nature of signals. Most solutions have failed to establish a knowledge base of normal behavior covering typical operating conditions, and also lack the ability to match the current operating conditions online and dynamically adjust the judgment criteria.

[0008] Therefore, in complex and ever-changing real-world processing scenarios, there is an urgent need for a health status evaluation method that can automatically learn normal operating modes under different working conditions and thereby achieve accurate deviation measurement, so as to break through the inherent limitations of the fixed threshold mechanism and improve the accuracy and timeliness of anomaly detection. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for evaluating the health status of vertical CNC machine tools, aiming to solve the technical problem in the prior art where the dynamic and ever-changing operating conditions of machine tools lead to a large number of false alarms and missed alarms when using fixed thresholds for health status monitoring, resulting in low reliability of the early warning system and inability to effectively guide predictive maintenance.

[0010] To achieve the above objectives, the present invention provides a method for evaluating the health status of a vertical CNC machine tool, the method comprising the following steps: Real-time synchronous acquisition of operating condition data and multimodal sensor signal data of vertical CNC machine tools. The operating condition data includes currently executed CNC code instructions, tool library information and workpiece material parameters obtained from the CNC system kernel. The multimodal sensor signal data includes vibration signals, temperature signals, acoustic signals and spindle drive current signals of key machine tool components. The operating condition data is structured and analyzed to extract a condition feature vector containing spindle speed, feed rate, depth of cut, tool identification number, and material physical properties. Simultaneously, the multimodal sensor signal data is processed by time-frequency domain transformation to extract a real-time state feature vector containing the vibration signal spectrum amplitude at each measuring point, spindle bearing temperature, sound pressure level in the cutting area, and effective value of spindle drive current. Based on a machine tool component relationship graph neural network model pre-trained using historical health operation data, the working condition feature vector is used as the input of the model. The model generates a predicted state feature vector that accurately corresponds to the current working condition based on the physical topological relationship and mechanical transmission characteristics between the functional components of the machine tool. The predicted state feature vector characterizes the multimodal sensing signal response that the machine tool should have when it is in an absolutely healthy state under this specific working condition. Calculate the residual vector between the real-time state feature vector and the predicted state feature vector, where each dimension of the residual vector corresponds to the deviation between a sensing signal and a theoretical health benchmark. Based on the residual vector and the health status residual covariance matrix learned during the model training phase, the Mahalanobis distance of the residual vector is calculated, and the Mahalanobis distance value is used as a dynamic health status evaluation value characterizing the degree of deviation of the machine tool from the health benchmark. The dynamic health status evaluation value is continuously compared with the preset multi-level dynamic alarm threshold to generate an early warning level corresponding to the current health status of the machine tool. When the dynamic health status evaluation value is greater than the highest level alarm threshold, the amplitude of each component in the residual vector is analyzed to determine the physical source component causing the abnormal state.

[0011] As one embodiment of the present invention, the real-time synchronous acquisition of the operating condition data and multimodal sensor signal data of the vertical CNC machine tool specifically includes: Through the industrial fieldbus protocol interface, the kernel data area of ​​the CNC system of the vertical CNC machine tool can be directly accessed, the CNC code pointed to by the program segment pointer can be read in real time, and the spindle speed command, feed rate command and tool call command in it can be parsed. The material grade of the workpiece being processed can be obtained through the human-computer interaction interface or the interface with the manufacturing execution system, and the corresponding physical property parameters such as elastic modulus, density and thermal conductivity can be retrieved from the material property database. Triaxial piezoelectric accelerometers are installed on key structural components such as the machine tool spindle box, bed column, and worktable to collect vibration acceleration signals in three orthogonal directions. Platinum resistance temperature sensors are installed on the outside of the front and rear bearing housings of the spindle to monitor the bearing operating temperature. A high-sensitivity condenser microphone is installed inside the safety shield near the cutting area to collect air-conducted acoustic signals during the machining process; The three-phase input current of the spindle drive motor is measured non-invasively using a Hall effect current sensor.

[0012] All sensor signals are acquired by a multi-channel data acquisition device with synchronous sampling function, and are synchronized with the internal clock of the CNC system at the microsecond level through a network time protocol.

[0013] As one embodiment of the present invention, the time-frequency domain transformation processing of the multimodal sensing signal data specifically includes: The vibration acceleration signals collected at each measuring point are processed in frames. A Hanning window function is applied to each data frame, and then a fast Fourier transform is performed to extract the spectral amplitude distribution within a preset characteristic frequency band, which constitutes the vibration component of the real-time state feature vector. The resistance value output by the platinum resistance temperature sensor is calibrated by looking up a table and converted into a Celsius temperature value; The root mean square value of the acoustic signal collected by the condenser microphone is calculated and converted into sound pressure level in decibels. Calculate the instantaneous effective value of the three-phase current signal of the spindle drive motor.

[0014] All the values ​​obtained from the above processing are combined in a predetermined order to form a real-time state feature vector with fixed dimensions.

[0015] As one embodiment of the present invention, the machine tool component relationship graph neural network model is a deep learning model that abstracts the physical structure of the machine tool into a graph structure, wherein: The nodes in the diagram represent the core functional components of the machine tool, including the spindle unit, X-axis feed drive unit, Y-axis feed drive unit, Z-axis feed drive unit, tool changer unit, and bed structure. The edges of the graph represent the physical connection relationships and mechanical transmission paths between components, and the weight of the edges is determined by the connection stiffness and damping coefficient between the components. Each node is assigned a set of attribute features, which are assigned values ​​in the model input layer by the corresponding parameters in the working condition feature vector; The model operates through a multi-layer graph convolutional network, aggregating information from adjacent nodes layer by layer to simulate the transmission and response of load and motion in the entire machine tool structure under specific working conditions. The final output layer of the model passes through a fully connected network to generate a predicted state feature vector that matches the dimension of the input working condition vector. Each element of this vector corresponds to a theoretical health reading of a sensor under that working condition.

[0016] As one embodiment of the present invention, the step of calculating the Mahalanobis distance of the residual vector based on the residual vector and the health state residual covariance matrix learned during the model training phase follows the standard calculation formula of Mahalanobis distance in its specific mathematical operation process. The covariance matrix used is obtained during the offline training phase of the model by inputting massive historical health operation data covering the entire working condition range into the established model and statistically analyzing the overall distribution characteristics of the generated residual vector on all health samples. This covariance matrix captures the intrinsic correlation between the residuals of different sensor signals.

[0017] As one embodiment of the present invention, the step of determining the physical source component causing the abnormal state specifically includes: When the dynamic health status evaluation value is greater than the highest level alarm threshold, the system locks the residual vector at that moment; Normalize all elements in the residual vector and sort them by absolute value; Identify the N residual components with the largest absolute values, where N is a preset integer; Based on the preset mapping relationship between the first N residual components and the physical sensor, the corresponding physical component or physical phenomenon is determined and taken as the most likely source of abnormality to generate a diagnostic report.

[0018] The present invention also provides a health status evaluation system for vertical CNC machine tools, the system comprising: The module for synchronous acquisition of operating data and sensor signals is connected to the CNC system of the vertical CNC machine tool through an industrial bus interface and to multiple sensors deployed at various locations on the machine tool. It is used to perform the step of real-time synchronous acquisition of operating condition data and multimodal sensor signal data. The feature engineering and data processing module is connected to the synchronous acquisition module and is used to perform the steps of parsing the operating condition data into operating condition feature vectors and processing the multimodal sensor signal data into real-time state feature vectors. A dynamic health baseline generation module, which has a pre-trained machine tool component relationship graph neural network model embedded in it, and is connected to the feature engineering and data processing module, is used to receive the working condition feature vector and execute the step of generating the predicted state feature vector. A health status quantification and evaluation module, which is connected to the feature engineering and data processing module and the dynamic health benchmark generation module, is used to calculate the residual vector between the real-time state feature vector and the predicted state feature vector, and to generate the dynamic health status evaluation value based on Mahalanobis distance. A multi-level early warning and fault diagnosis module, which is connected to the health status quantification and evaluation module, is used to perform the steps of comparing dynamic health status evaluation values ​​with multi-level alarm thresholds and generating early warning levels, and analyzing residual vectors to determine the source of anomalies when the highest level alarm is triggered.

[0019] All modules of the system are deployed in an industrial control computer independent of the CNC system, and exchange internal data through a high-speed data bus.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing a dynamic health benchmark that is strictly correlated with the current actual operating conditions of the machine tool, this invention fundamentally solves the problem of high false alarm rate and false negative rate that the fixed threshold method inevitably produces when facing variable load and variable parameter conditions, and greatly improves the accuracy and reliability of health status evaluation. 2. The machine tool component relationship diagram neural network model can profoundly simulate the internal physical laws of machine tools as complex electromechanical systems. The prediction benchmarks generated by the model are not only highly accurate, but also have clear physical interpretability, which is better than traditional black box models. 3. By using Mahalanobis distance to measure the multidimensional residual vector, the intrinsic correlation between various sensor signals can be comprehensively considered. This results in higher sensitivity to early and subtle fault signs caused by multi-factor coupling, thus enabling earlier fault warning. 4. After an alarm is triggered, the structured analysis of the residual vector can directly point the anomaly to a specific physical component, realizing the leap from judging whether there is an anomaly to diagnosing where the anomaly is, providing accurate and direct data support for subsequent maintenance decisions, and improving the efficiency and pertinence of equipment maintenance. Attached Figure Description

[0021] Figure 1This is a schematic diagram of the overall technical architecture of a method and system for evaluating the health status of a vertical CNC machine tool proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the neural network model for the relationship diagram of machine tool components in this invention; Figure 3 This is a logical flowchart of the synchronous acquisition of operating condition data and multimodal sensing signals and the construction of feature vectors in this invention. Figure 4 This is a flowchart illustrating the logical flow of generating dynamic health benchmark prediction state feature vectors based on working condition feature vectors in this invention. Figure 5 This is a flowchart illustrating the logical process of residual vector calculation, Mahalanobis distance evaluation, and dynamic health status quantitative assessment in this invention. Figure 6 This is a logical flowchart of the multi-level early warning triggering and abnormal physical source diagnosis in this invention. Detailed Implementation

[0022] Please refer to Figures 1 to 6 This invention provides a method and system for evaluating the health status of vertical CNC machine tools. It aims to solve the technical problem in existing technologies where the dynamic and ever-changing operating conditions of machine tools lead to numerous false alarms and missed alarms when using fixed thresholds for health status monitoring, resulting in low reliability of the early warning system and its inability to effectively guide predictive maintenance. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0023] The method for evaluating the health status of vertical CNC machine tools includes the following steps: Real-time synchronous acquisition of operating condition data and multimodal sensor signal data of vertical CNC machine tools; The operating condition data is structured and parsed into operating condition feature vectors, and the multimodal sensor signal data is processed by time-frequency domain transformation to form real-time state feature vectors. Based on a pre-trained neural network model of machine tool component relationship graph, a predicted state feature vector is generated according to the working condition feature vector; Calculate the residual vector between the real-time state feature vector and the predicted state feature vector; The Mahalanobis distance is calculated based on the residual vector and the covariance matrix of the health status residuals, and is used as a dynamic health status evaluation value. The evaluation value is compared with the multi-level dynamic alarm threshold to generate an early warning level. When the value is greater than the highest level threshold, the residual vector is analyzed to determine the abnormal physical source component.

[0024] First, execute step S1: Real-time synchronous acquisition of the operating condition data and multimodal sensor signal data of the vertical CNC machine tool.

[0025] Operating condition data includes currently executing CNC code instructions, tool library information, and workpiece material parameters obtained from the CNC system kernel.

[0026] Multimodal sensing signal data includes vibration signals, temperature signals, acoustic signals, and spindle drive current signals of key machine tool components.

[0027] Specifically, the kernel data area of ​​the CNC system of the vertical CNC machine tool is directly accessed through the industrial fieldbus protocol interface. The CNC code pointed to by the program segment pointer is read in real time, and the spindle speed command, feed rate command and tool call command are parsed.

[0028] The material grade of the workpiece being processed is obtained through the human-computer interaction interface or the interface with the manufacturing execution system, and the corresponding physical property parameters such as elastic modulus, density and thermal conductivity are retrieved from the material property database.

[0029] Triaxial piezoelectric accelerometers are installed on key structural components such as the machine tool spindle box, bed column, and worktable to collect vibration acceleration signals in three orthogonal directions.

[0030] Platinum resistance temperature sensors are installed on the outside of the front and rear bearing housings of the spindle to monitor the bearing operating temperature.

[0031] A high-sensitivity condenser microphone is installed inside the safety shield near the cutting area to collect airborne acoustic signals during the machining process.

[0032] The three-phase input current of the spindle drive motor is measured non-invasively using a Hall effect current sensor.

[0033] All sensor signal acquisition is performed by a multi-channel data acquisition device with synchronous sampling function, and is synchronized with the internal clock of the CNC system at the microsecond level through network time protocol to ensure that the operating condition data and sensor signal data are strictly aligned in the time dimension.

[0034] Then, step S2 is executed: the operating condition data is structured and analyzed to extract a condition feature vector containing spindle speed, feed rate, depth of cut, tool identification number, and material physical properties; at the same time, the multimodal sensor signal data is processed by time-frequency domain transformation to extract a real-time state feature vector containing the vibration signal spectrum amplitude of each measuring point, spindle bearing temperature value, sound pressure level in the cutting area, and effective value of spindle drive current.

[0035] For operating condition data, the system first performs lexical and syntactic analysis on the CNC code to identify G code, M code and coordinate instructions, and extracts the spindle speed S value, feed rate F value and tool number T value from them.

[0036] The cutting depth is calculated by combining the current tool geometry parameters with the trajectory interpolation results in the workpiece coordinate system.

[0037] The physical properties of materials are indexed from the material grade to the local material database to obtain its three core parameters: elastic modulus, density, and thermal conductivity.

[0038] The above six parameters are arranged in a predefined order to form a fixed-dimensional working condition feature vector.

[0039] For multimodal sensing signal data, the vibration acceleration signals collected at each measuring point are divided into data frames with a length of 2048 sampling points and a frame shift of 1024 sampling points. A Hanning window function is applied to each frame to suppress spectral leakage. Then, a fast Fourier transform is performed to extract the spectral amplitude distribution within a preset characteristic frequency band. This frequency band covers the spindle rotation frequency and its harmonics, gear meshing frequency, and bearing fault characteristic frequency range. The amplitudes of 30 key frequency points are selected as vibration components.

[0040] The resistance value output by the platinum resistance temperature sensor is calibrated using a pre-stored resistance-temperature lookup table and converted into a Celsius temperature value.

[0041] The acoustic signal collected by the condenser microphone is bandpass filtered, its root mean square value is calculated, and then converted into sound pressure level in decibels.

[0042] The three-phase current signal of the spindle drive motor is synchronously sampled, and its instantaneous effective value is calculated.

[0043] All the values ​​obtained from the above processing are combined in a predetermined order to form a real-time state feature vector with a fixed dimension, which is strictly consistent with the dimension of the subsequent predicted state feature vector.

[0044] Next, step S3 is executed: Based on the machine tool component relationship graph neural network model that has been pre-trained using historical health operation data, the working condition feature vector is used as the input of the model. The model generates a predicted state feature vector that accurately corresponds to the current working condition according to the physical topological relationship and mechanical transmission characteristics between the various functional components of the machine tool.

[0045] The predicted state feature vector characterizes the multimodal sensing signal response that the machine tool should have when it is in an absolutely healthy state under this specific operating condition.

[0046] The machine tool component relationship graph neural network model is a deep learning model that abstracts the physical structure of the machine tool into a graph structure.

[0047] The nodes in the diagram represent the core functional components of the machine tool, including the spindle unit, X-axis feed drive unit, Y-axis feed drive unit, Z-axis feed drive unit, tool changer unit, and bed structure.

[0048] The edges of the graph represent the physical connection relationships and mechanical transmission paths between components. The weights of the edges are determined by the connection stiffness and damping coefficients between components. These parameters are obtained through finite element simulation or experimental modal analysis and are fixed during the model initialization stage.

[0049] Each node is assigned a set of attribute features, which are assigned values ​​in the model input layer by the corresponding parameters in the working condition feature vector.

[0050] For example, the spindle unit node receives the spindle speed and tool identification number, the X, Y, and Z axis feed drive units receive the corresponding feed rate and depth of cut information, and the bed structure node receives the material physical properties.

[0051] The model performs computations through a three-layer graph convolutional network. The graph convolution operation in each layer is defined as follows: For each node, aggregate the feature vectors of its first-order adjacent nodes, sum them by weight, and then pass them through a non-linear activation function.

[0052] The aggregation weights are determined by the physical weights of the edges and the learnable parameters. After three layers of information transmission, the feature vectors of each node have been integrated with the context information of the entire graph.

[0053] Finally, the feature vectors of all nodes are concatenated and passed through a three-layer fully connected network to output a predicted state feature vector with the same dimension as the real-time state feature vector.

[0054] The model is trained end-to-end using historical health operation data covering the entire operating condition range during the offline phase. The loss function is the mean square error between the predicted state feature vector and the true health state feature vector.

[0055] Then perform step S4: calculate the residual vector between the real-time state feature vector and the predicted state feature vector.

[0056] Each dimension of the residual vector corresponds to the deviation between a sensing signal and a theoretical health benchmark.

[0057] The specific calculation method is element-wise subtraction, i.e., residual vector. ,in This is the real-time state feature vector. This is the feature vector for predicting the state.

[0058] The residual vector directly reflects the degree to which the machine tool deviates from the ideal healthy state under the current operating conditions. The sign and magnitude of each component indicate whether the specific sensing signal is greater than or less than the healthy baseline, and the magnitude of the deviation.

[0059] Then, step S5 is executed: based on the residual vector and the health status residual covariance matrix learned during the model training phase, the Mahalanobis distance of the residual vector is calculated, and the Mahalanobis distance value is used as a dynamic health status evaluation value that characterizes the degree of deviation of the machine tool from the health benchmark.

[0060] The formula for calculating Mahalanobis distance is: ; For the residual vector, Let be the residual covariance matrix of the health status.

[0061] The covariance matrix is ​​obtained during the offline training phase of the model by inputting massive amounts of historical health operation data covering the entire operating condition range into the established model, and statistically analyzing the overall distribution characteristics of the generated residual vector across all healthy samples.

[0062] It captures the intrinsic correlation between the residuals of different sensing signals. For example, increased spindle vibration is often accompanied by increased bearing temperature and increased sound pressure level.

[0063] The advantage of Mahalanobis distance over Euclidean distance is that it takes into account the variance and covariance of data in various dimensions, can normalize and measure signals with different dimensions and fluctuation characteristics, and amplifies joint bias patterns that are rare in healthy conditions, thus making it more sensitive to early minor faults.

[0064] Finally, step S6 is executed: the dynamic health status evaluation value is continuously compared with the preset multi-level dynamic alarm threshold to generate an early warning level corresponding to the current health status of the machine tool, and when the dynamic health status evaluation value is greater than the highest level alarm threshold, the amplitude of each component in the residual vector is analyzed to determine the physical source component causing the abnormal state.

[0065] The multi-level dynamic alarm threshold includes three levels: Level 1 warning threshold corresponds to slight deviation, prompting the operator to pay attention; Level 2 warning threshold corresponds to abnormality, suggesting that a planned inspection be arranged; Level 3 warning threshold corresponds to serious fault risk, requiring immediate shutdown and maintenance.

[0066] These thresholds are not fixed values, but are dynamically adjusted according to the feature vector of the current working condition. For example, under heavy-load cutting conditions, the allowable Mahalanobis distance threshold is slightly larger than that under light-load conditions.

[0067] When the evaluation value is greater than the level 3 warning threshold, the system locks the residual vector at that moment, normalizes all elements in the vector so that the mean is zero and the standard deviation is one, and then sorts them in descending order of absolute value.

[0068] Identify the top three residual components with the largest absolute values, and determine the corresponding physical components or physical phenomena based on the preset mapping relationship—for example, the first dimension corresponds to the amplitude of the spindle X-axis vibration spectrum, the second dimension corresponds to the temperature of the spindle front bearing, and the third dimension corresponds to the cutting sound pressure level.

[0069] If the first three items are all spindle-related signals, the diagnostic report will indicate that "there is an abnormality in the spindle unit, which may be caused by bearing wear, dynamic balance failure or tool clamping looseness," and recommend that the spindle system be checked first.

[0070] The vertical CNC machine tool health status evaluation system includes five core functional modules. All modules are deployed in an industrial control computer independent of the CNC system and exchange internal data through a high-speed data bus.

[0071] The first module is the synchronous acquisition module for operating data and sensor signals. This module is connected to the CNC system of the vertical CNC machine tool through an industrial bus interface and to multiple sensors deployed throughout the machine tool. It is used to perform the function of real-time synchronous acquisition of operating condition data and multimodal sensor signal data.

[0072] The second module is the feature engineering and data processing module, which is connected to the synchronous acquisition module. It has a built-in signal processing algorithm library and is used to perform functions such as parsing the operating condition data into operating condition feature vectors and processing multimodal sensor signal data into real-time state feature vectors.

[0073] The third module is the dynamic health baseline generation module. This module contains a pre-trained neural network model of machine tool component relationship graph and is connected to the feature engineering and data processing module. It is used to receive working condition feature vectors and perform the function of generating predictive state feature vectors.

[0074] The fourth module is the health status quantitative evaluation module. This module is connected to the feature engineering and data processing module and the dynamic health benchmark generation module. It has a built-in Mahalanobis distance calculation unit and covariance matrix storage unit, which are used to calculate the residual vector and generate dynamic health status evaluation values.

[0075] The fifth module is the multi-level early warning and fault diagnosis module. This module is connected to the health status quantitative evaluation module and has built-in threshold comparison logic and residual analysis engine. It is used to perform functions such as comparing dynamic health status evaluation values ​​with multi-level alarm thresholds and generating early warning levels, as well as analyzing residual vectors to determine the source of anomalies when the highest level alarm is triggered.

[0076] Through the collaborative work of the above modules, the system achieves closed-loop management of the entire process from data acquisition, feature extraction, dynamic benchmark generation, health quantification to intelligent early warning and diagnosis.

[0077] This embodiment solves the problem of high false alarm and false negative rates that inevitably occur when the fixed threshold method faces variable load and variable parameter conditions by introducing a dynamic health benchmark that is strictly related to the current actual operating conditions of the machine tool.

[0078] By employing a neural network model of machine tool component relationships, it is possible to profoundly simulate the internal physical laws of machine tools as complex electromechanical systems. The generated prediction benchmarks are not only highly accurate, but also have clear physical interpretability.

[0079] Using Mahalanobis distance to measure multidimensional residual vectors can comprehensively consider the intrinsic correlation between various sensing signals and has higher sensitivity to early and weak fault signs caused by multi-factor coupling.

[0080] After an alarm is triggered, the structured analysis of the residual vector can directly point the anomaly to a specific physical component, thus achieving a leap from judging whether there is an anomaly to diagnosing where the anomaly is, providing accurate and direct data support for subsequent maintenance decisions.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the health status of a vertical CNC machine tool, characterized in that, include: Real-time synchronous acquisition of operating condition data and multimodal sensor signal data of vertical CNC machine tools; The operating condition data is structured and analyzed to extract a condition feature vector containing spindle speed, feed rate, depth of cut, tool identification number, and material physical properties. Simultaneously, the multimodal sensor signal data is processed by time-frequency domain transformation to extract a real-time state feature vector containing the vibration signal spectrum amplitude at each measuring point, spindle bearing temperature, sound pressure level in the cutting area, and effective value of spindle drive current. Based on a machine tool component relationship graph neural network model pre-trained using historical health operation data, the working condition feature vector is used as the input to the model. The machine tool component relationship graph neural network model generates a predicted state feature vector that accurately corresponds to the current working condition based on the physical topological relationship and mechanical transmission characteristics between the various functional components of the machine tool. Calculate the residual vector between the real-time state feature vector and the predicted state feature vector, where each dimension of the residual vector corresponds to the deviation between a sensing signal and a theoretical health benchmark. Based on the residual vector and the health status residual covariance matrix learned during the model training phase, the Mahalanobis distance of the residual vector is calculated, and the Mahalanobis distance value is used as a dynamic health status evaluation value characterizing the degree of deviation of the machine tool from the health benchmark. The dynamic health status evaluation value is continuously compared with the preset multi-level dynamic alarm threshold to generate an early warning level corresponding to the current health status of the machine tool. When the dynamic health status evaluation value is greater than the highest level alarm threshold, the amplitude of each component in the residual vector is analyzed to determine the physical source component causing the abnormal state.

2. The method for evaluating the health status of a vertical CNC machine tool according to claim 1, characterized in that, Real-time synchronous acquisition of operating condition data and multimodal sensor signal data of vertical CNC machine tools, including: Through the industrial fieldbus protocol interface, the kernel data area of ​​the CNC system of the vertical CNC machine tool can be directly accessed, the CNC code pointed to by the program segment pointer can be read in real time, and the spindle speed command, feed rate command and tool call command in it can be parsed. The material grade of the workpiece being processed can be obtained through the human-computer interaction interface or the interface with the manufacturing execution system, and the corresponding physical property parameters such as elastic modulus, density and thermal conductivity can be retrieved from the material property database. Triaxial piezoelectric accelerometers are installed on key structural components such as the machine tool spindle box, bed column, and worktable to collect vibration acceleration signals in three orthogonal directions. Platinum resistance temperature sensors are installed on the outside of the front and rear bearing housings of the spindle to monitor the bearing operating temperature. A high-sensitivity condenser microphone is installed inside the safety shield near the cutting area to collect air-conducted acoustic signals during the machining process; The three-phase input current of the spindle drive motor is non-invasively measured using a Hall effect current sensor. All sensor signals are acquired by a multi-channel data acquisition device with synchronous sampling function, and are synchronized with the internal clock of the CNC system at the microsecond level through a network time protocol.

3. The method for evaluating the health status of a vertical CNC machine tool according to claim 2, characterized in that, The time-frequency domain transformation processing of the multimodal sensing signal data includes: The vibration acceleration signals collected at each measuring point are processed in frames. A Hanning window function is applied to each data frame, and then a fast Fourier transform is performed to extract the spectral amplitude distribution within a preset characteristic frequency band, which constitutes the vibration component of the real-time state feature vector. The resistance value output by the platinum resistance temperature sensor is calibrated by looking up a table and converted into a Celsius temperature value; The root mean square value of the acoustic signal collected by the condenser microphone is calculated and converted into sound pressure level in decibels. Calculate the instantaneous effective value of the three-phase current signal of the spindle drive motor; All the values ​​obtained from the above processing are combined in a predetermined order to form a real-time state feature vector with fixed dimensions.

4. The method for evaluating the health status of a vertical CNC machine tool according to claim 1, characterized in that, The machine tool component relationship graph neural network model is a deep learning model that abstracts the physical structure of the machine tool into a graph structure, wherein: The nodes in the diagram represent the core functional components of the machine tool, including the spindle unit, X-axis feed drive unit, Y-axis feed drive unit, Z-axis feed drive unit, tool changer unit, and bed structure. The edges of the graph represent the physical connection relationships and mechanical transmission paths between components, and the weight of the edges is determined by the connection stiffness and damping coefficient between the components. Each node is assigned a set of attribute features, which are assigned values ​​in the model input layer by the corresponding parameters in the working condition feature vector; The model operates through a multi-layer graph convolutional network, aggregating information from adjacent nodes layer by layer to simulate the transmission and response of load and motion in the entire machine tool structure under specific working conditions. The final output layer of the model passes through a fully connected network to generate a predicted state feature vector that matches the dimension of the input working condition vector. Each element of this vector corresponds to a theoretical health reading of a sensor under that working condition.

5. The method for evaluating the health status of a vertical CNC machine tool according to claim 4, characterized in that, The machine tool component relationship graph neural network model is operated through a three-layer graph convolutional network. The graph convolution operation of each layer is defined as follows: For each node, the feature vectors of its first-order adjacent nodes are aggregated, weighted, and then passed through a non-linear activation function. The aggregation weights are determined by the physical weights of the edges and the learnable parameters. After three layers of information transmission, the feature vectors of each node are concatenated and passed through a three-layer fully connected network to output the predicted state feature vector.

6. The method for evaluating the health status of a vertical CNC machine tool according to claim 1, characterized in that, The health status residual covariance matrix is ​​obtained during the offline training phase of the model by inputting massive amounts of historical health operation data covering the entire operating range into the established model and statistically analyzing the overall distribution characteristics of the generated residual vector on all health samples. This covariance matrix captures the intrinsic correlation between the residuals of different sensor signals.

7. The method for evaluating the health status of a vertical CNC machine tool according to claim 6, characterized in that, Calculating the Mahalanobis distance of the residual vector includes: The Mahalanobis distance is calculated using the following formula: ； in Let be the residual vector. Let be the residual covariance matrix of the health state.

8. The method for evaluating the health status of a vertical CNC machine tool according to claim 1, characterized in that, The multi-level dynamic alarm threshold includes three levels of warning thresholds: a first-level warning threshold, a second-level warning threshold, and a third-level warning threshold. The third-level warning threshold corresponds to a serious fault risk, requiring immediate shutdown and maintenance. Furthermore, the multi-level dynamic alarm threshold is dynamically adjusted based on the current operating condition feature vector.

9. The method for evaluating the health status of a vertical CNC machine tool according to claim 1, characterized in that, Identify the physical source component causing the abnormal state, including: When the dynamic health status evaluation value is greater than the highest level alarm threshold, the system locks the residual vector at that moment; Normalize all elements in the residual vector and sort them by absolute value; Identify the N residual components with the largest absolute values, where N is a preset integer; Based on the preset mapping relationship between the first N residual components and the physical sensor, the corresponding physical component or physical phenomenon is determined and taken as the most likely source of abnormality to generate a diagnostic report.

10. A health status evaluation system for vertical CNC machine tools, characterized in that, The health status evaluation of a vertical CNC machine tool is implemented using the health status evaluation method according to any one of claims 1 to 9. The health status evaluation system for a vertical CNC machine tool includes: The operation data and sensor signal synchronous acquisition module is used to connect to the CNC system of the vertical CNC machine tool through an industrial bus interface, and to connect to multiple sensors deployed in various parts of the machine tool, in order to perform the step of real-time synchronous acquisition of operation condition data and multimodal sensor signal data. The feature engineering and data processing module is used to parse operating condition data into operating condition feature vectors and process multimodal sensor signal data into real-time state feature vectors. The dynamic health benchmark generation module has a pre-trained neural network model of machine tool component relationship graph embedded in it, which is used to receive the working condition feature vector and generate the predicted state feature vector. The health status quantitative evaluation module is used to calculate the residual vector between the real-time status feature vector and the predicted status feature vector, and to generate the dynamic health status evaluation value based on Mahalanobis distance. The multi-level early warning and fault diagnosis module is used to compare the dynamic health status evaluation value with the multi-level dynamic alarm threshold and generate an early warning level, and analyze the residual vector to determine the abnormal source component when the highest level alarm is triggered.

Images