Method and system for identifying operation faults of CNC machine tools

Abstract

The present invention relates to the field of data processing technology, and discloses a method and system for identifying operational faults of numerically controlled machine tools, which are used to improve the accuracy of operational fault identification of numerically controlled machine tools. The method includes: performing federated learning modeling and virtual environment verification on feature data streams to obtain a digital twin model of normal operation of the machine tool; performing data extraction on the digital twin model of normal operation of the machine tool to obtain an operation data set, and performing reinforcement learning on the operation data set to obtain a distributed fault mode library; performing edge computing anomaly detection on the feature data stream to obtain hierarchical collaborative preliminary fault diagnosis data; performing augmented reality-assisted interactive confirmation on the hierarchical collaborative preliminary fault diagnosis data to obtain confirmation data, and performing fault matching based on the confirmation data and the knowledge graph of the distributed fault mode library to obtain target fault data; performing fault response strategy analysis on the target fault data to obtain a target response strategy corresponding to the target fault data.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for identifying operation faults of a numerically controlled machine tool. Background Art

[0002] As an important equipment in modern manufacturing, the operating status of CNC machine tools directly affects production efficiency and product quality. Traditional fault identification methods mainly rely on manual experience and simple threshold judgment, which is difficult to cope with complex industrial environments and diverse fault types. With the development of Industry 4.0, some intelligent fault diagnosis methods have been proposed, such as data-driven machine learning methods and model-based fault diagnosis methods. These methods have improved the accuracy and efficiency of fault identification to a certain extent.

[0003] However, existing fault identification methods still have some shortcomings. First, most methods do not fully consider data security and privacy protection issues, and sensitive information is easily leaked during data collection and processing. Second, existing methods often use centralized data processing and twin construction methods, which are difficult to adapt to distributed industrial environments. Furthermore, most methods lack the ability to analyze and predict the dynamic evolution of faults and cannot provide timely and effective fault response strategies. Finally, existing methods need to be improved in terms of human-computer interaction and decision support, and it is difficult to make full use of expert experience and real-time feedback from operators. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and system for identifying operating faults of a numerically controlled machine tool, so as to improve the accuracy of identifying operating faults of a numerically controlled machine tool.

[0005] The present invention provides an operation fault identification method for a numerically controlled machine tool, comprising: performing distributed edge acquisition and homomorphic encryption preprocessing on multi-source heterogeneous sensor data of the numerically controlled machine tool to obtain a feature data stream; performing federated learning modeling and virtual environment verification on the feature data stream to obtain a digital twin model of normal operation of the machine tool; performing data extraction on the digital twin model of normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data and simulated fault data, and performing reinforcement learning on the operation data set to obtain a distributed fault mode library; performing edge computing anomaly detection on the real-time feature data stream collected and analyzed in real time through the digital twin model of normal operation of the machine tool and the distributed fault mode library to obtain hierarchical collaborative preliminary fault diagnosis data; performing augmented reality-assisted interactive confirmation on the hierarchical collaborative preliminary fault diagnosis data to obtain confirmation data and performing data feature mapping on the confirmation data to obtain a feature vector to be matched, and performing fault matching based on the feature vector to be matched and the knowledge graph of the distributed fault mode library to obtain target fault data; performing fault response strategy analysis on the target fault data to obtain a target response strategy corresponding to the target fault data.

[0006] In the present invention, the step of performing distributed edge acquisition and homomorphic encryption preprocessing on the multi-source heterogeneous sensor data of the CNC machine tool to obtain a feature data stream includes: distributing the multi-source heterogeneous sensors of the CNC machine tool to obtain an original sensor network; allocating edge computing nodes to the original sensor network to obtain an edge computing network topology; real-time acquisition of sensor data through the edge computing network topology to obtain an original data stream; performing data cleaning and outlier detection on the original data stream to obtain a preprocessed data stream; performing denoising on the preprocessed data stream through a wavelet transform algorithm to obtain a denoised data stream; performing time domain and frequency domain feature extraction on the denoised data stream to obtain a multidimensional feature vector; performing dimensionality reduction processing on the multidimensional feature vector through a principal component analysis algorithm to obtain an optimized feature vector; performing data standardization and normalization processing on the optimized feature vector to obtain a standardized feature vector; encrypting the standardized feature vector through a homomorphic encryption algorithm to obtain encrypted feature data; and performing distributed storage and index construction on the encrypted feature data to obtain the feature data stream.

[0007] In the present invention, the step of performing federated learning modeling and virtual environment verification on the feature data stream to obtain a digital twin model of the normal operation of the machine tool includes: performing data segmentation on the feature data stream to obtain multiple local data subsets; performing privacy protection processing on the local data subsets through a differential privacy algorithm to obtain a privacy-protected data subset; performing local twin construction on the privacy-protected data subset to obtain an initial local twin; performing model parameter aggregation on the initial local twin through a secure aggregation protocol to obtain a global twin; performing model distillation on the global twin to obtain a compressed twin; and performing a federated learning algorithm on the compressed twin. Iterative optimization is performed to obtain an optimized global twin; a virtual environment is constructed for the optimized global twin to obtain a virtual operating environment for the machine tool; multi-scenario simulation is performed on the virtual operating environment of the machine tool through a Monte Carlo simulation method to obtain a simulation data set; the simulation data set is homomorphically encrypted to obtain encrypted simulation data, and the encrypted simulation data is input into the optimized global twin to obtain global twin output data, and data labels are compared between the global twin output data and the simulation data set to obtain label difference data; parameters of the label difference data are tuned through a Bayesian optimization algorithm to obtain a digital twin model of the normal operation of the machine tool.

[0008] In the present invention, data extraction is performed on the digital twin model of the normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data and simulated fault data, and reinforcement learning is performed on the operation data set to obtain a distributed fault mode library step, including: setting parameter thresholds for the digital twin model of the normal operation of the machine tool to obtain a normal operation parameter range; randomly sampling the normal operation parameter range through the Monte Carlo method to obtain a sampling data set, and performing statistical deviation analysis on the sampling data to obtain operation deviation data, wherein the operation deviation data includes: mean deviation data, variance deviation data and distribution deviation data; data Cleaning and standardization are performed to obtain normalized historical fault data; data enhancement is performed on the normalized historical fault data by using a generative adversarial network algorithm to obtain an expanded historical fault data set; a fault injection experiment is performed on the normal operation digital twin model of the machine tool to obtain simulated fault data; the operation deviation data, the expanded historical fault data set and the simulated fault data are labeled to obtain a comprehensive operation data set; reinforcement learning training is performed on the comprehensive operation data set by using a deep Q learning algorithm to obtain an initial fault mode library; data compression is performed on the initial fault mode library to obtain a lightweight fault mode library; distributed storage and consensus mechanism construction are performed on the lightweight fault mode library to obtain the distributed fault mode library.

[0009] In the present invention, the step of performing edge computing anomaly detection on the real-time feature data stream collected and analyzed in real time through the normal operation digital twin model of the machine tool and the distributed fault mode library to obtain hierarchical collaborative preliminary fault diagnosis data includes: homomorphically encrypting the real-time feature data stream to obtain a real-time encrypted data stream, and performing real-time segmentation processing on the real-time feature data stream to obtain time series data fragments; performing dynamic feature extraction on the time series data fragments through a sliding window algorithm to obtain a dynamic feature sequence; performing operation feature sequence extraction on the normal operation digital twin model of the machine tool to obtain an operation feature sequence, and performing comparative analysis on the dynamic feature sequence and the operation feature sequence to obtain a preliminary anomaly indicator; performing anomaly quantification on the preliminary anomaly indicator through a local anomaly factor algorithm to obtain an anomaly degree evaluation. score; perform threshold judgment on the abnormality score to obtain an abnormal trigger signal and an abnormal type label; based on the abnormal trigger signal and the abnormal type label, quickly search the distributed fault mode library to obtain a candidate fault mode set; perform multi-dimensional matching on the candidate fault mode set through a fuzzy inference algorithm to obtain a multi-dimensional feature vector, obtain fault time domain characteristics and fault frequency domain characteristics, and perform fault mode membership analysis on the fault time domain characteristics and fault frequency domain characteristics to obtain a fault possibility distribution; perform hierarchical clustering analysis on the fault possibility distribution to obtain hierarchical fault categories; perform causal relationship inference on the hierarchical fault categories through a Bayesian network to obtain a fault propagation path; perform a comprehensive analysis on the fault propagation path and the hierarchical fault categories to obtain the hierarchical collaborative preliminary fault diagnosis data.

[0010] In the present invention, the step of performing augmented reality-assisted interactive confirmation on the preliminary fault diagnosis data of the hierarchical collaboration to obtain confirmation data, and performing fault matching based on the confirmation data and the knowledge graph of the distributed fault mode library to obtain target fault data includes: visualizing the preliminary fault diagnosis data of the hierarchical collaboration to obtain a three-dimensional fault model; matching the three-dimensional fault model with the actual machine tool through a spatial registration algorithm to obtain an augmented reality overlay view; designing a human-computer interaction interface for the augmented reality overlay view to obtain an interactive fault display interface; and recognizing the operator's interactive operations through a gesture recognition algorithm to obtain interactive instructions. sequence; semantically parse the interactive instruction sequence to obtain fault confirmation information; perform credibility analysis on the fault confirmation information through a fuzzy comprehensive evaluation method to obtain the confirmation data, and perform data feature mapping on the confirmation data to obtain a feature vector to be matched; construct a knowledge graph for the distributed fault mode library to obtain a fault knowledge graph; embed and learn the fault knowledge graph through a graph neural network algorithm to obtain a fault feature vector; calculate the similarity between the feature vector to be matched and the fault feature vector to obtain a fault matching result; comprehensively evaluate the fault matching result through a multi-criteria decision-making algorithm to obtain the target fault data.

[0011] In the present invention, the step of performing fault response strategy analysis on the target fault data to obtain the target response strategy corresponding to the target fault data includes: evaluating the fault severity of the target fault data to obtain a fault risk level; classifying the fault risk level through a decision tree algorithm to obtain a preliminary response plan set; performing historical data association analysis on the preliminary response plan set to obtain a plan effect evaluation index; optimizing and calculating the plan effect evaluation index through a simulated annealing algorithm to obtain a preferred response plan; performing resource demand analysis on the preferred response plan to obtain a resource allocation list; optimizing and allocating the resource allocation list through a linear programming algorithm to obtain an optimal resource scheduling plan; performing time series arrangement on the optimal resource scheduling plan to obtain a response strategy execution schedule; dynamically adjusting the response strategy execution schedule through a reinforcement learning algorithm to obtain an adaptive execution plan; performing human-machine collaborative task allocation on the adaptive execution plan to obtain a collaborative work instruction; and performing semantic optimization processing on the collaborative work instruction to obtain the target response strategy.

[0012] The present invention also provides an operation fault identification system for a numerically controlled machine tool, comprising:

[0013] The encryption module is used to perform distributed edge acquisition and homomorphic encryption preprocessing on multi-source heterogeneous sensor data of CNC machine tools to obtain feature data streams;

[0014] A verification module, used to perform federated learning modeling and virtual environment verification on the characteristic data stream to obtain a digital twin model of normal operation of the machine tool;

[0015] An extraction module is used to extract data from the digital twin model of the normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data and simulated fault data, and to perform reinforcement learning on the operation data set to obtain a distributed fault mode library;

[0016] A detection module is used to perform edge computing anomaly detection on the real-time feature data stream collected and analyzed in real time through the normal operation digital twin model of the machine tool and the distributed fault mode library to obtain hierarchical and collaborative preliminary fault diagnosis data;

[0017] A matching module, used to perform augmented reality-assisted interactive confirmation on the hierarchical collaborative preliminary fault diagnosis data, obtain confirmation data, perform data feature mapping on the confirmation data, obtain a feature vector to be matched, and perform fault matching based on the feature vector to be matched and the knowledge graph of the distributed fault mode library to obtain target fault data;

[0018] The analysis module is used to perform fault response strategy analysis on the target fault data to obtain a target response strategy corresponding to the target fault data.

[0019] In the technical solution provided by the present invention, the security and privacy protection capabilities of data collection are significantly improved through distributed edge collection and homomorphic encryption preprocessing technology, and the leakage of sensitive information is effectively prevented; the introduction of federated learning modeling and virtual environment verification mechanism not only protects the data privacy of each participant, but also improves the generalization and robustness of the model; the reinforcement learning method is used to build a distributed fault mode library, which enhances the system's adaptive learning ability for new faults; through edge computing anomaly detection and multi-model fusion diagnosis, fast and accurate fault identification is achieved, greatly improving the real-time and reliability of the system; the fault confirmation and matching method combining augmented reality technology and knowledge graph greatly improves the intuitiveness and efficiency of human-computer interaction, enabling operators to understand and handle fault situations more quickly and accurately; the fault matching mechanism based on multi-criteria decision-making improves the accuracy and Interpretability; by introducing multiple optimization algorithms such as decision trees, simulated annealing, and linear programming, the intelligent generation and dynamic adjustment of fault response strategies are realized, greatly improving the efficiency and pertinence of fault handling; the reinforcement learning algorithm is used to adaptively adjust the execution plan, enhancing the system's adaptability to complex and dynamic environments; through human-machine collaborative task allocation and knowledge distillation technology, it effectively combines human experience and machine intelligence to improve the overall effect of fault handling; the distributed architecture of this method and the application of edge computing technology have significantly improved the scalability and processing capabilities of the system, and can better adapt to large-scale and complex industrial environments; by building a complete fault identification and processing process, from data collection, twin construction, fault diagnosis to strategy formulation and execution, a closed-loop intelligent fault management system is formed, which greatly improves the operation reliability and maintenance efficiency of CNC machine tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 It is a flow chart of a method for identifying operation faults of a numerically controlled machine tool in an embodiment of the present invention;

[0022] Figure 2 The figure is a schematic diagram of an operation fault identification system for a CNC machine tool according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0025] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , Figure 1 is a flow chart of a method for identifying an operation fault of a CNC machine tool according to an embodiment of the present invention. Figure 1 As shown, the following steps are included:

[0027] S101, performing distributed edge acquisition and homomorphic encryption preprocessing on multi-source heterogeneous sensor data of CNC machine tools to obtain feature data streams;

[0028] S102, performing federated learning modeling and virtual environment verification on the feature data stream to obtain a digital twin model of normal operation of the machine tool;

[0029] S103, extracting data from the digital twin model of normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data, and simulated fault data, and performing reinforcement learning on the operation data set to obtain a distributed fault mode library;

[0030] S104, performing edge computing anomaly detection on the feature data stream through the normal operation digital twin model of the machine tool and the distributed fault mode library to obtain hierarchical and collaborative preliminary fault diagnosis data;

[0031] S105, performing augmented reality-assisted interactive confirmation on the hierarchical collaborative preliminary fault diagnosis data to obtain confirmation data, and performing fault matching based on the confirmation data and the knowledge graph of the distributed fault mode library to obtain target fault data;

[0032] S106: Perform fault response strategy analysis on the target fault data to obtain a target response strategy corresponding to the target fault data.

[0033] It should be noted that distributed edge collection and homomorphic encryption preprocessing of multi-source heterogeneous sensor data are performed. This step involves deploying a variety of sensors at key parts of the machine tool, such as vibration sensors, temperature sensors, and current sensors, which collect data in real time through edge computing nodes. The collected raw data undergoes preprocessing steps such as cleaning, noise reduction, and feature extraction, and then is encrypted using a homomorphic encryption algorithm to ensure the security of the data during transmission and storage. Homomorphic encryption allows direct calculations on encrypted data without decryption, thereby protecting sensitive information. Federated learning modeling and virtual environment verification are performed on the encrypted feature data stream. Federated learning is a distributed machine learning technology that allows multiple participants to jointly build twins without sharing raw data. In this process, each edge node is built using a local data twin, and then the model parameters (not the raw data) are transmitted to the central server for aggregation to obtain a global twin. This global twin is then verified through a virtual environment, which simulates various possible machine tool operation scenarios to ensure the robustness and generalization ability of the model.

[0034] The operation data set is extracted from the normal operation digital twin model of the machine tool, including operation deviation data, historical fault data and simulated fault data. The operation deviation data reflects the parameter fluctuation range of the machine tool during normal operation, the historical fault data comes from the actual fault records in the past, and the simulated fault data is generated through fault injection experiments. These data are trained by the reinforcement learning algorithm to construct a distributed fault pattern library. Reinforcement learning can continuously optimize fault identification and processing strategies through trial and error and reward mechanisms. During real-time operation, the system uses the normal operation digital twin model of the machine tool and the distributed fault pattern library to perform edge computing anomaly detection on the feature data stream collected in real time. Edge computing allows calculations to be performed near the data source, reducing data transmission delays and improving response speed. Anomaly detection algorithms such as local anomaly factor (LOF) can quickly identify abnormal data points and combine the diagnosis results of multiple models to obtain preliminary fault diagnosis data.

[0035] In order to improve the accuracy of diagnosis, the system uses augmented reality technology to assist operators in interactive confirmation. Augmented reality technology superimposes virtual fault information on the actual machine tool view, allowing operators to intuitively understand the fault situation. The operator's confirmation information is matched with the knowledge graph in the distributed fault mode library. The knowledge graph represents the complex relationship between faults through a graph structure, and uses a graph neural network algorithm to extract and match fault features, and finally obtain the target fault data.

[0036] Conduct in-depth analysis of target fault data and formulate corresponding response strategies. This process involves multiple optimization algorithms, such as decision trees for fault classification, simulated annealing algorithms for solution optimization, and linear programming algorithms for resource scheduling. The response strategy finally generated not only considers the severity of the fault, but also factors such as resource allocation and execution time to ensure the feasibility and effectiveness of the strategy.

[0037] For example, when the spindle of a CNC machine tool was running, the vibration sensor detected abnormal vibration signals. These signals were pre-processed and encrypted by the edge node and then transmitted to the central processing unit. Based on the pre-trained normal operation model, the system found that the vibration amplitude exceeded the normal range. At the same time, the temperature sensor also showed that the spindle temperature was slightly elevated. The system input these abnormal data into the distributed fault mode library for matching, and preliminarily diagnosed that it might be a spindle bearing failure. Through the augmented reality interface, the operator saw that the spindle bearing position was marked in red and detailed vibration and temperature data were displayed. The operator confirmed the abnormality and immediately generated a step-by-step response strategy including shutdown inspection, bearing replacement and recalibration. The entire process from abnormality detection to strategy generation took only a few minutes, greatly reducing the risk of potential downtime and failure expansion.

[0038] By executing the above steps, the security and privacy protection capabilities of data collection are significantly improved through distributed edge collection and homomorphic encryption preprocessing technology, effectively preventing the leakage of sensitive information; the introduction of federated learning modeling and virtual environment verification mechanism not only protects the data privacy of all participants, but also improves the generalization and robustness of the model; the reinforcement learning method is used to build a distributed fault mode library, which enhances the system's adaptive learning ability for new faults; through edge computing anomaly detection and multi-model fusion diagnosis, fast and accurate fault identification is achieved, greatly improving the real-time and reliability of the system; the fault confirmation and matching method combining augmented reality technology and knowledge graphs greatly improves the intuitiveness and efficiency of human-computer interaction, enabling operators to understand and handle fault situations more quickly and accurately; the fault matching mechanism based on multi-criteria decision-making improves the accuracy and reliability of fault diagnosis Interpretability; by introducing multiple optimization algorithms such as decision trees, simulated annealing, and linear programming, the intelligent generation and dynamic adjustment of fault response strategies are realized, greatly improving the efficiency and pertinence of fault handling; the reinforcement learning algorithm is used to adaptively adjust the execution plan, enhancing the system's adaptability to complex and dynamic environments; through human-machine collaborative task allocation and knowledge distillation technology, it effectively combines human experience and machine intelligence to improve the overall effect of fault handling; the distributed architecture of this method and the application of edge computing technology have significantly improved the scalability and processing capabilities of the system, and can better adapt to large-scale and complex industrial environments; by building a complete fault identification and processing process, from data collection, twin construction, fault diagnosis to strategy formulation and execution, a closed-loop intelligent fault management system is formed, which greatly improves the operation reliability and maintenance efficiency of CNC machine tools.

[0039] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0040] (1) Distributed deployment of multi-source heterogeneous sensors of CNC machine tools to obtain the original sensor network;

[0041] (2) Allocate edge computing nodes to the original sensor network to obtain the edge computing network topology;

[0042] (3) Real-time sensor data collection through edge computing network topology to obtain the original data stream;

[0043] (4) Perform data cleaning and outlier detection on the original data stream to obtain the preprocessed data stream;

[0044] (5) Performing denoising on the preprocessed data stream by using a wavelet transform algorithm to obtain a denoised data stream;

[0045] (6) Extract time domain and frequency domain features from the denoised data stream to obtain a multi-dimensional feature vector;

[0046] (7) Using the principal component analysis algorithm to reduce the dimensionality of the multi-dimensional feature vector, we can obtain the optimized feature vector;

[0047] (8) Performing data standardization and normalization processing on the optimized feature vector to obtain a standardized feature vector;

[0048] (9) Encrypting the standardized feature vector through a homomorphic encryption algorithm to obtain encrypted feature data;

[0049] (10) Distributed storage and index construction are performed on the encrypted feature data to obtain a feature data stream.

[0050] Specifically, the multi-source heterogeneous sensors are deployed in a distributed manner to obtain the original sensor network. This step involves installing different types of sensors at key parts of the machine tool, such as vibration sensors, temperature sensors, pressure sensors, and current sensors. These sensors are connected by wire or wireless means to form a perception network covering the entire machine tool. The original sensor network is assigned edge computing nodes to obtain the edge computing network topology. Edge computing is a technology that distributes computing tasks to the edge of the network, which can reduce data transmission delays and improve real-time performance. In this step, multiple edge computing nodes are deployed near the machine tool based on the distribution and data characteristics of the sensors, and each node is responsible for processing sensor data in a specific area or a specific type.

[0051] The sensor data is collected in real time through the edge computing network topology to obtain the original data stream. Each edge node obtains data from the sensor according to the preset sampling frequency to form a continuous data stream. These raw data contain various physical quantity information during the operation of the machine tool, but may also contain noise and outliers. Data cleaning and outlier detection are performed on the raw data stream to obtain a preprocessed data stream. Data cleaning includes operations such as removing duplicate data and filling missing values. Outlier detection uses statistical methods or machine learning algorithms to identify and process data points that are significantly deviated from the normal range to ensure the quality and reliability of the data.

[0052] The preprocessed data stream is denoised by the wavelet transform algorithm to obtain a denoised data stream. Wavelet transform is a time-frequency analysis method that can effectively remove noise while retaining the important features of the signal. It decomposes the signal into wavelet coefficients of different scales, removes the noise-related wavelet coefficients by threshold processing, and then reconstructs the signal to obtain the denoised data. The denoised data stream is subjected to time domain and frequency domain feature extraction to obtain a multidimensional feature vector. Time domain features include statistics such as mean, variance, and peak value; frequency domain features are obtained by Fourier transform, including spectrum energy distribution, main frequency, etc.; time-frequency domain features combine time domain and frequency domain information, such as wavelet packet energy features. These features together constitute a multidimensional feature space that describes the operating status of the machine tool.

[0053] The multi-dimensional feature vector is reduced in dimension by the principal component analysis (PCA) algorithm to obtain the optimized feature vector. PCA is a commonly used dimensionality reduction technique that projects the original high-dimensional data into a new low-dimensional space through linear transformation while retaining the main information of the data. This step not only reduces the dimension of the data, but also reduces the computational complexity of subsequent processing. The optimized feature vector is subjected to data standardization and normalization to obtain a standardized feature vector. Standardization converts features of different dimensions to the same scale, usually converting the data into a distribution with a mean of 0 and a variance of 1. Normalization scales the data to a specific interval, such as [0, 1]. These operations ensure the comparability between different features, which is conducive to the subsequent machine learning algorithm processing. The standardized feature vector is encrypted by the homomorphic encryption algorithm to obtain encrypted feature data. Homomorphic encryption is a special encryption method that allows calculations to be performed directly on encrypted data without decrypting it first. This ensures the security of data during transmission and storage, while also supporting the analysis and processing of encrypted data.

[0054] Finally, the encrypted feature data is distributedly stored and indexed to obtain the feature data stream. Distributed storage improves the reliability and access efficiency of data, while indexing speeds up data retrieval.

[0055] For example, the spindle of a CNC machine tool is equipped with vibration sensors, temperature sensors, and current sensors. These sensors are connected to the nearest edge computing node via a wireless network. The edge node collects 1,000 vibration data, 10 temperature data, and 100 current data per second. After the raw data is cleaned, 0.1% of the outliers are removed. Then, the vibration signal is decomposed into 5 layers using the db4 wavelet, which effectively reduces the high-frequency noise. In the feature extraction stage, the RMS value and peak factor in the time domain, the spectral entropy and main frequency in the frequency domain, and the wavelet energy characteristics in the time and frequency domain are calculated. Through PCA analysis, the original 20-dimensional feature vector is reduced to 8 dimensions, retaining 95% of the information. Standardization processing maps all eigenvalues ​​to the [-1, 1] interval. Finally, the Paillier homomorphic encryption algorithm is used to encrypt the feature vector, and the encrypted data is distributed and stored on multiple nodes, and an index structure based on the B+ tree is established.

[0056] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0057] (1) Segment the feature data stream to obtain multiple local data subsets;

[0058] (2) Perform privacy protection on the local data subset using a differential privacy algorithm to obtain a privacy-protected data subset;

[0059] (3) Construct a local twin for the privacy-preserving data subset to obtain the initial local twin;

[0060] (4) Aggregate the model parameters of the initial local twin through a secure aggregation protocol to obtain a global twin;

[0061] (5) Perform model distillation on the global twin to obtain the compressed twin;

[0062] (6) Iteratively optimize the compressed twins through the federated learning algorithm to obtain the optimized global twins;

[0063] (7) Construct a virtual environment for the optimized global twin to obtain the virtual operating environment of the machine tool;

[0064] (8) Use the Monte Carlo simulation method to simulate multiple scenarios of the machine tool virtual operating environment and obtain a simulation data set;

[0065] (9) Perform homomorphic encryption on the simulation data set to obtain encrypted simulation data, and input the encrypted simulation data into the optimized global twin to obtain global twin output data, and perform data label comparison on the global twin output data and the simulation data set to obtain label difference data;

[0066] (10) The parameters of the label difference data are tuned through the Bayesian optimization algorithm to obtain the digital twin model of the normal operation of the machine tool.

[0067] It should be noted that the feature data stream is segmented to obtain multiple local data subsets. This step divides the overall data according to dimensions such as time series or machine tool components to ensure that each subset contains enough information to train the local model. Next, the local data subset is privacy-protected by the differential privacy algorithm to obtain a privacy-protected data subset. Differential privacy is a mathematically rigorous privacy protection method that adds carefully designed random noise to the original data so that individual information cannot be inferred from the statistical results while maintaining the overall statistical properties of the data. A local twin is constructed for the privacy-protected data subset to obtain an initial local twin. This step is performed independently on each edge node, using machine learning algorithms such as deep neural networks to train a model that can reflect local features based on local data. Next, the model parameters of the initial local twin are aggregated through a secure aggregation protocol to obtain a global twin. The secure aggregation protocol allows the parameters of multiple local models to be securely merged without leaking the original data of each party, thereby obtaining a global twin that integrates all local information.

[0068] Perform model distillation on the global twin to obtain a compressed twin. Model distillation is a technique that transfers knowledge from a complex model (teacher model) to a simple model (student model). This step results in a lighter model with similar performance, which is conducive to deployment on edge devices. Subsequently, the compressed twin is iteratively optimized through a federated learning algorithm to obtain an optimized global twin. Federated learning is a distributed machine learning paradigm that allows multiple parties to jointly build twins without sharing original data, and continuously improves model performance through multiple rounds of local training and global aggregation. A virtual environment is constructed for the optimized global twin to obtain a virtual operating environment for the machine tool. This step creates a digital twin system that simulates various working states and parameters of a real machine tool. The Monte Carlo simulation method is used to perform multi-scenario simulations on the virtual operating environment of the machine tool to obtain a simulation data set. The Monte Carlo method simulates various possible operating scenarios, including normal operation and various fault conditions, through a large number of random sampling to generate comprehensive simulation data.

[0069] Homomorphic encryption allows specific computing operations to be performed in an encrypted state without decrypting the data. A commonly used homomorphic encryption method is the Paillier encryption algorithm. This algorithm can be used to encrypt various parameters in the simulation data set (such as temperature, vibration frequency, pressure, etc.) to obtain encrypted simulation data. The encrypted simulation data is input into the optimized global twin. After obtaining the output data of the global twin, it needs to be compared with the original simulation data set. This step requires decrypting the output data first and then comparing it with the original unencrypted simulation data. The comparison process can use various indicators, such as mean square error, mean absolute error, or more complex statistical methods to evaluate the difference between the model output and the expected output. Homomorphic encryption allows certain operations to be performed in the encrypted domain, but it also increases the computational complexity and data size. Therefore, in practical applications, it may be necessary to find a balance between privacy protection and computational efficiency.

[0070] Finally, the parameters of the label difference data are tuned through the Bayesian optimization algorithm to obtain the digital twin model of the normal operation of the machine tool. Bayesian optimization is an efficient global optimization method that builds a probabilistic model of the objective function and strikes a balance between exploration and utilization, thereby quickly finding the optimal model parameter configuration.

[0071] For example, a CNC machine tool manufacturer has 10 machine tools of the same model distributed in different factories. The edge computing node of each machine tool collects one year of operation data, including vibration, temperature, current and other information. First, the data of each machine tool is divided into 12 subsets by month. Then, the Laplace mechanism is used to implement differential privacy, and noise with a mean of 0 and a standard deviation of 0.1 is added to each data point. Each edge node uses these privacy-preserving data to train a 5-layer deep neural network as a local model. Then, the secure multi-party computing technology is used to aggregate the parameters of each local model to obtain an initial global twin. Through knowledge distillation, this complex global twin is compressed into a 3-layer lightweight model. Then 10 rounds of federated learning iterations are performed, and in each round, the edge node updates the model with new local data and then performs global aggregation. Based on the optimized global twin, a virtual machine tool environment with 100 key parameters is constructed. Using the Monte Carlo method, 10,000 different operation scenarios are simulated in this environment, generating a large amount of simulation data. These simulation data were input into the optimized global twin, and the prediction error of the model in various scenarios was analyzed. Finally, the Bayesian optimization algorithm was used to find the model parameter configuration that minimized the prediction error through 100 iterations, thus obtaining the final digital twin model of the normal operation of the machine tool.

[0072] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0073] (1) Setting parameter thresholds for the digital twin model of normal operation of the machine tool to obtain the normal operation parameter range;

[0074] (2) Randomly sample the normal operating parameter range using the Monte Carlo method to obtain a sampled data set, and perform statistical deviation analysis on the sampled data to obtain operating deviation data, wherein the operating deviation data includes: mean deviation data, variance deviation data, and distribution deviation data;

[0075] (3) Perform data cleaning and standardization on historical fault data to obtain standardized historical fault data;

[0076] (4) The normalized historical fault data is enhanced by using a generative adversarial network algorithm to obtain an expanded historical fault data set;

[0077] (5) Conduct fault injection experiments on the digital twin model of the normal operation of the machine tool to obtain simulated fault data;

[0078] (6) Label the operation deviation data, the expanded historical fault data set, and the simulated fault data to obtain a comprehensive operation data set;

[0079] (7) Perform data sharding and federation distribution on the comprehensive operation data set to obtain distributed data subsets;

[0080] (8) Perform reinforcement learning training on the distributed data subset through the deep Q learning algorithm to obtain the initial fault mode library;

[0081] (9) Compressing the data of the initial fault mode library to obtain a lightweight fault mode library;

[0082] (10) Distributed storage and consensus mechanism are constructed for the lightweight fault mode library to obtain a distributed fault mode library.

[0083] Specifically, the parameter thresholds of the normal operation digital twin model of the machine tool are set to obtain the normal operation parameter range. This step sets the upper and lower thresholds for each key parameter based on the design specifications and long-term operation data of the machine tool, and defines the boundaries of the normal operation of the machine tool. Subsequently, the normal operation parameter range is randomly sampled by the Monte Carlo method to obtain the operation deviation data. The Monte Carlo method is a numerical calculation method based on random sampling. It performs a large number of random samplings within the set parameter range to generate a data set that reflects the parameter fluctuations under normal operation. These parameter ranges are randomly sampled using the Monte Carlo method. The core of the Monte Carlo method is to simulate the behavior of complex systems through a large number of random samplings. In this scenario, a large number of random samples can be generated using uniform distribution or more complex probability distributions (such as normal distribution). For example, 10,000 groups of samples can be generated, each containing random values ​​of the above three parameters. For example, suppose that the three parameters of the machine tool spindle speed, feed rate and cutting depth are of interest. The normal operation range may be: spindle speed 1000-5000 RPM, feed rate 100-500 mm / min, and cutting depth 0.5-3 mm. After generating the sample data set, it is necessary to perform statistical deviation analysis on these data to obtain the running deviation data. This analysis process includes calculating the mean deviation, variance deviation and distribution deviation. The mean deviation reflects the difference between the average level of the parameter and the expected value, the variance deviation indicates the change in the degree of parameter fluctuation, and the distribution deviation reflects the change in the shape of the parameter distribution. To calculate the mean deviation, first calculate the sample mean of each parameter and then compare it with the expected normal operating mean. For example, if the expected mean of spindle speed is 3000 RPM and the sample mean is 3100 RPM, then the mean deviation is 100 RPM. The same method applies to other parameters. The calculation of variance deviation involves calculating the sample variance of each parameter and comparing it with the expected normal operating variance. If the sample variance is significantly larger than the expected variance, it may mean that the instability of the machine tool operation has increased. For example, if the expected variance of feed speed is 50 (mm / min)² and the sample variance is 70 (mm / min)², then the variance deviation is 20 (mm / min)². The analysis of distribution deviations is more complex and usually requires the use of statistical tests, such as the chi-square test or the Kolmogorov-Smirnov test, to compare the difference between the sample distribution and the expected distribution. This can help detect unusual changes in the shape of the parameter distribution, such as a change from a normal distribution to a skewed distribution or a bimodal distribution. In practical applications, this process may be implemented using a programming language such as Python. For example, the NumPy library is used to generate random samples and the SciPy library is used for statistical analysis. The code may include functions to generate random samples, functions to calculate statistics, and functions to perform statistical tests.

[0084] A fault injection experiment was conducted on the digital twin model of the normal operation of the machine tool to obtain simulated fault data. Fault injection is a method of actively simulating various fault conditions. By artificially introducing various fault parameters in the normal operation model, a large amount of simulated fault data was generated. Subsequently, the operation deviation data, the expanded historical fault data set, and the simulated fault data were labeled by the data fusion algorithm to obtain a comprehensive operation data set. The data fusion algorithm comprehensively considers the characteristics and weights of data from different sources to form a data set that comprehensively reflects the various operating states of the machine tool. The comprehensive operation data set was subjected to data sharding and federated distribution to obtain a distributed data subset. Data sharding divides a large data set into multiple small subsets, while federated distribution ensures that these subsets are stored in different nodes in a dispersed manner, which not only improves the parallelism of data processing but also enhances data security. Then, the distributed data subsets were subjected to reinforcement learning training by the deep Q learning algorithm to obtain the initial fault mode library. Deep Q learning is a reinforcement learning method that combines deep neural networks with Q learning. It can learn optimal strategies in large-scale, high-dimensional state spaces and effectively capture complex fault modes.

[0085] The initial fault mode library is compressed to obtain a lightweight fault mode library. Model pruning reduces the model size by removing unimportant neurons or connections; model compression further reduces the storage space required for the model through quantization, encoding and other technologies, making the fault mode library more suitable for deployment on edge devices. Finally, the lightweight fault mode library is distributedly stored and a consensus mechanism is constructed to obtain a distributed fault mode library. Distributed storage improves data reliability and access efficiency, while the consensus mechanism ensures the consistency and credibility of fault mode data distributed on different nodes.

[0086] For example, the spindle system of a certain model of CNC machine tool contains 10 key parameters such as speed, temperature, and vibration. Based on historical data and expert experience, the normal operating range of these parameters is set, such as the speed range of 0-12000rpm and the temperature range of 20-80℃. The Monte Carlo method is used to generate 1 million sets of parameter combinations within these ranges to form operating deviation data. The 1000 fault data recorded in the past 5 years are cleaned, 50 obviously erroneous records are removed, and all parameters are standardized to the interval [-1, 1]. Using the GAN algorithm, 5000 new fault data samples are generated based on these 950 real fault data. In a virtual environment, 2000 different fault scenarios are simulated by changing parameters such as suddenly increasing the vibration amplitude or sharply rising the temperature. After labeling these data, a comprehensive data set containing more than 8000 states is obtained. This data set is divided into 100 subsets and distributed and stored on different edge nodes. Using the deep Q learning algorithm, an initial library containing 50 common fault modes is trained on these distributed data. Through pruning and compression, the model size was reduced from the original 500MB to 50MB. Finally, this lightweight fault mode library was deployed to a distributed system consisting of 10 nodes, using a blockchain-based consensus mechanism to ensure data consistency and security.

[0087] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0088] (1) Perform homomorphic encryption on the real-time feature data stream to obtain a real-time encrypted data stream, and perform real-time segmentation processing on the real-time feature data stream to obtain time series data segments;

[0089] (2) Extract dynamic features from time series data segments using a sliding window algorithm to obtain a dynamic feature sequence;

[0090] (3) Extracting the operating feature sequence of the digital twin model of the normal operation of the machine tool to obtain the operating feature sequence, and performing comparative analysis between the dynamic feature sequence and the operating feature sequence to obtain preliminary abnormality indicators;

[0091] (4) Quantify the initial abnormal indicators through the local abnormal factor algorithm to obtain the abnormality degree score;

[0092] (5) Perform threshold judgment on the abnormality degree score to obtain the abnormal trigger signal and abnormal type label;

[0093] (6) Based on the abnormal trigger signal and abnormal type label, the distributed fault mode library is quickly searched to obtain the candidate fault mode set;

[0094] (7) Perform multi-dimensional matching on the candidate fault mode set through fuzzy reasoning algorithm to obtain a multi-dimensional feature vector, obtain fault time domain features and fault frequency domain features, and perform fault mode membership analysis on the fault time domain features and fault frequency domain features to obtain fault possibility distribution;

[0095] (8) Perform hierarchical cluster analysis on the fault probability distribution to obtain hierarchical fault categories;

[0096] (9) Use Bayesian networks to infer the causal relationship of hierarchical fault categories and obtain the fault propagation path;

[0097] (10) Comprehensively analyze the fault propagation paths and hierarchical fault categories to obtain preliminary hierarchical collaborative fault diagnosis data.

[0098] Specifically, implementing this process involves two main steps: homomorphic encryption and time series data segmentation processing, which need to be performed simultaneously to ensure the real-time and security of the data. First, homomorphic encryption of real-time feature data streams is key to protecting sensitive information. Homomorphic encryption is a special encryption technique that allows specific computing operations to be performed in an encrypted state without first decrypting the data. This is particularly useful when processing real-time data streams because it allows necessary data analysis while protecting data privacy. A commonly used homomorphic encryption method is the Paillier encryption system. In practical applications, Python's phe library can be used to implement Paillier encryption. Suppose there is a real-time data stream that contains various parameters of a CNC machine tool, such as temperature, vibration, pressure, etc. An encryption function can be created for each data point that encrypts the data using a public key. The process might look like this: first generate a pair of public and private keys, and then encrypt each data point using the public key. It is important to note that the encryption process increases computational overhead, so in real-time systems, high-performance hardware or optimized algorithms may be required to ensure that encryption does not cause significant delays.

[0099] At the same time, the real-time feature data stream needs to be segmented to obtain time series data fragments. This process usually involves sliding window technology. Sliding window is a commonly used time series analysis method that can capture the time correlation of data. In this method, a fixed-size window is defined, which "slides" over time, and each movement contains the latest data points. For example, if your data arrives at a rate of 100 samples per second, you might choose a window size of 5 seconds, which means that each time series data fragment will contain 500 data points. One method is to segment before encryption and then encrypt each segment. Another method is to use a special data structure that supports homomorphic encryption, allowing window operations to be performed in an encrypted state. Whichever method you choose, you need to carefully consider the trade-off between performance and security.

[0100] In real-time applications, you might use a stream processing framework such as Apache Flink or Apache Kafka Streams to process real-time data streams. These frameworks provide built-in window operations and data processing functions that can be easily integrated with custom encryption functions. For example, you can create a processing function that receives the original data stream, encrypts each data point, and then applies a sliding window operation to create time series data fragments.

[0101] For example, suppose you are monitoring the temperature of a spindle on a CNC machine. The temperature sensor sends a reading every second. You want to encrypt this data and create 10-second time series data segments. The steps to implement this process might be as follows: First, initialize the Paillier encryption system and generate public and private keys. Then, create a stream processing pipeline that receives temperature readings, encrypts each reading using the public key, and puts the encrypted data into a sliding window of size 10. Whenever the window is full (that is, every 10 seconds), output a time series data segment containing 10 encrypted temperature readings.

[0102] Next, the dynamic feature extraction of the time series data fragments is performed through the sliding window algorithm to obtain the dynamic feature sequence. The sliding window algorithm moves a fixed-size window on the data and calculates the statistical features of the data in the window, such as the mean, variance, peak value, etc., so as to capture the local change trend of the data. Extracting the operation feature sequence of the digital twin model of the normal operation of the machine tool is the key first step. As a virtual representation of the actual machine tool, the digital twin model contains various parameters and behavioral characteristics of the machine tool under normal operation. To extract the operation feature sequence, it is necessary to simulate the operation state of the machine tool under various normal working conditions and record the changes of key parameters. These parameters may include spindle speed, feed rate, cutting force, temperature, vibration, etc. Time series analysis techniques, such as deep learning models such as autoregressive integrated moving average model (ARIMA) or long short-term memory network (LSTM) can be used to capture the change pattern of parameters over time. For example, the LSTM model can be used to predict the change trend of various machine tool parameters in the future period of time, and these prediction results constitute the operation feature sequence. Specifically, historical data, such as parameter records for the past 8 hours, can be input to let the model predict the parameter changes in the next hour. The operating characteristic sequence obtained in this way not only contains the absolute values ​​of the parameters, but also their changing trends and mutual relationships.

[0103] It is necessary to compare and analyze the dynamic feature sequence and the operating feature sequence. The dynamic feature sequence is a data sequence collected in real time from the actual running machine tool, while the operating feature sequence is the ideal state sequence predicted by the digital twin model. The purpose of the comparative analysis is to find the difference between the actual operating state and the expected normal state, so as to identify potential anomalies. This process can use a variety of statistical and machine learning methods. A common method is to calculate the Euclidean distance or Mahalanobis distance between the two sequences, which can give an overall similarity measure. For example, if there are n parameters and each parameter has m points in the time series, the two sequences can be regarded as two points in the n*m ​​dimensional space, and then the distance between the two points is calculated. The larger the distance, the more the actual operating state deviates from the expected normal state. Another more detailed method is to use the Dynamic Time Warping (DTW) algorithm. DTW can handle the inconsistency of the two sequences in the time scale and find the best correspondence. This is particularly useful for dealing with slight time delays or advances that may occur in the operation of machine tools. Through DTW, the degree of difference between the actual state and the expected state at each time point can be obtained.

[0104] Next, the local outlier factor (LOF) algorithm is used to quantify the initial outlier indicators and obtain an outlier degree score. The LOF algorithm calculates the local density comparison between the data point and its neighboring points, assigns an outlier score to each data point, and effectively identifies local outliers.

[0105] Threshold judgment is performed on the abnormality score to obtain an abnormal trigger signal. This step sets one or more thresholds. When the abnormality score exceeds the threshold, the abnormal signal is triggered. Based on the abnormal trigger signal, the distributed fault mode library is quickly retrieved to obtain a candidate fault mode set. The candidate fault mode set is matched in multiple dimensions using a fuzzy reasoning algorithm. Fuzzy reasoning is a reasoning method based on fuzzy logic. It can handle uncertainty and ambiguity and is very suitable for fault diagnosis of complex systems. In this process, a series of fuzzy rules need to be defined first, which associate various operating parameters of the machine tool with possible fault modes. For example, a simple fuzzy rule may be "If the vibration amplitude is high and the temperature rises quickly, there may be a bearing fault." These rules are usually formulated by domain experts based on experience and historical data. Next, these fuzzy rules are used to match the candidate fault mode set in multiple dimensions. The "multi-dimensional" here refers to considering multiple aspects of machine tool operation, such as vibration, temperature, sound, energy consumption, etc. Each dimension may have multiple related parameters. For example, the vibration dimension may include parameters such as vibration amplitude, frequency, acceleration, etc. Through fuzzy reasoning, the matching degree of each candidate fault mode in each dimension can be obtained, and these matching degrees constitute a multi-dimensional feature vector. In order to better understand the dynamic behavior of machine tools, it is necessary to analyze the time domain characteristics and frequency domain characteristics of faults separately. The time domain characteristics reflect the change law of faults over time, while the frequency domain characteristics reveal the periodicity and energy distribution of faults. In order to obtain the time domain characteristics, the original signal can be statistically analyzed to calculate indicators such as mean, standard deviation, peak value, and peak factor. These indicators can reflect the severity and development trend of the fault. For example, if the peak factor of the vibration signal suddenly increases, it may mean that a component is loose or cracked. The extraction of frequency domain features usually involves Fourier transform or wavelet transform. Through these transforms, the time domain signal can be converted to the frequency domain to observe the energy distribution of different frequency components. In the frequency domain, abnormal peaks at certain specific frequencies may correspond to specific types of faults. For example, bearing faults usually produce obvious peaks at their characteristic frequencies. Various spectrum analysis techniques, such as power spectral density estimation, cepstrum analysis, etc., can be used to extract these frequency domain features.

[0106] After obtaining the time domain and frequency domain features, it is necessary to perform fault mode membership analysis on these features to obtain the fault probability distribution. Membership analysis is the core concept of fuzzy set theory, which describes the degree to which an element belongs to a set. In the scenario, it is necessary to calculate the degree to which the observed features belong to each fault mode. This can be achieved by defining appropriate membership functions. Membership functions can be simple triangular functions, trapezoidal functions, or more complex Gaussian functions or bell-shaped functions. The choice of which membership function depends on the specific application scenario and expert knowledge.

[0107] For example, suppose you are diagnosing a spindle failure in a CNC milling machine. You may consider the following dimensions: vibration, temperature, sound, and machining accuracy. In the vibration dimension, you may extract time domain features such as RMS value and peak factor, and frequency domain features such as amplitude at characteristic frequencies. For temperature, you may focus on the temperature rise rate and maximum temperature. Sound features may include sound pressure level and energy in a specific frequency band. Machining accuracy may involve dimensional error and surface roughness.

[0108] Assume that through fuzzy reasoning, three possible fault modes are obtained: bearing fault, imbalance, and gear wear. A set of membership functions can be defined for each fault mode, which describe the degree to which each eigenvalue belongs to the fault. For example, for bearing fault, the membership function of high-frequency vibration may have a peak around 5000Hz, while for imbalance, the membership function of low-frequency vibration may have a peak at the rotor speed frequency.

[0109] By calculating the membership of the observed eigenvalues ​​on these membership functions, the probability of each fault mode can be obtained. For example, the probability of bearing failure is 0.7, the probability of imbalance is 0.2, and the probability of gear wear is 0.1. This constitutes the fault probability distribution.

[0110] The fuzzy reasoning algorithm is based on fuzzy set theory. It processes uncertainty and fuzzy information by setting fuzzy rules and membership functions, and calculates the probability of each fault mode. A hierarchical clustering analysis is performed on the fault probability distribution to obtain hierarchical fault categories. Hierarchical clustering analysis calculates the similarity between fault modes and aggregates similar fault modes into higher-level fault categories, forming a tree-like fault classification structure.

[0111] The Bayesian network is used to infer the causal relationship of the hierarchical fault categories and obtain the fault propagation path. The Bayesian network is a probabilistic graph model that represents variables through nodes and the dependencies between variables through edges. It can simulate the causal relationship in complex systems. This step uses the Bayesian network to infer the propagation relationship between faults and identify the root cause and possible development trend of the fault. Finally, the fault propagation path and hierarchical fault categories are comprehensively analyzed to obtain preliminary fault diagnosis data of hierarchical collaboration. This step comprehensively considers the fault category, severity and propagation path to form a comprehensive fault diagnosis report.

[0112] For example, the spindle system of a CNC machine tool monitors parameters such as speed, temperature, and vibration in real time. The system collects 100 data points per second and forms a time series data segment every 10 seconds. Using a sliding window of 5 seconds, each sliding for 1 second, the mean, standard deviation, peak-to-peak value and other features of the data in the window are calculated to form a dynamic feature sequence. Comparing these features with the normal operation model, it is found that the vibration amplitude exceeds the normal range by 20%. The LOF algorithm is used to calculate the abnormality and obtain an abnormal score of 0.85 (normal range 0-1). Because the score exceeds the preset threshold of 0.8, an abnormal signal is triggered. The system quickly retrieves the distributed fault mode library and finds 10 possible fault modes. Through fuzzy reasoning, the probability distribution of these 10 faults is calculated, among which "bearing wear" has the highest probability of 0.7. Hierarchical clustering classifies these faults into two categories: "bearing problem" and "assembly problem". Bayesian network analysis shows that bearing wear may cause the bearing seat to loosen, which in turn causes the vibration of the entire spindle system to increase. Based on this information, the system generated a preliminary diagnostic report, indicating that there may be bearing problems in the spindle system, recommending a focused inspection of bearing wear, and warning of possible chain failures.

[0113] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0114] (1) Visualize the hierarchical collaborative preliminary fault diagnosis data to obtain a three-dimensional fault model;

[0115] (2) Match the 3D fault model and the actual machine tool through a spatial registration algorithm to obtain an augmented reality overlay view;

[0116] (3) Design a human-computer interaction interface for the augmented reality overlay view to obtain an interactive fault display interface;

[0117] (4) Identify the operator's interactive operations through gesture recognition algorithms and obtain the interactive command sequence;

[0118] (5) Perform semantic analysis on the interactive instruction sequence to obtain fault confirmation information;

[0119] (6) Perform credibility analysis on the fault confirmation information through fuzzy comprehensive evaluation method to obtain confirmation data, and perform data feature mapping on the confirmation data to obtain the feature vector to be matched;

[0120] (7) Construct a knowledge graph for the distributed fault mode library to obtain a fault knowledge graph;

[0121] (8) Embed the fault knowledge graph through the graph neural network algorithm to obtain the fault feature vector;

[0122] (9) Calculate the similarity between the feature vector to be matched and the fault feature vector to obtain the fault matching result;

[0123] (10) The fault matching results are comprehensively evaluated through a multi-criteria decision-making algorithm to obtain the target fault data.

[0124] Specifically, the hierarchical collaborative preliminary fault diagnosis data is visualized to obtain a three-dimensional fault model. This step converts the abstract fault data into an intuitive three-dimensional model, and uses computer graphics technology to map information such as the fault location and severity to the three-dimensional structure of the machine tool. Then, the three-dimensional fault model and the actual machine tool are matched through a spatial registration algorithm to obtain an augmented reality overlay view. The spatial registration algorithm accurately superimposes the virtual fault model on the image or video stream of the actual machine tool through methods such as feature point matching and posture estimation.

[0125] Then, the human-computer interaction interface is designed for the augmented reality overlay view to obtain an interactive fault display interface. This step designs intuitive operation controls and information display areas, allowing operators to interact with fault information through touch, gestures, etc. The operator's interactive operations are recognized by the gesture recognition algorithm to obtain the interaction instruction sequence. The gesture recognition algorithm uses deep learning models, such as convolutional neural networks (CNN) or long short-term memory networks (LSTM), to recognize the operator's gesture actions from images or videos captured by the camera and convert them into corresponding instructions.

[0126] The interactive instruction sequence is semantically parsed to obtain fault confirmation information. Semantic parsing converts the operator's interactive behavior into semantic information that can be understood by the machine, such as confirming the fault type, adjusting fault parameters, etc. The fault confirmation information is analyzed for credibility through the fuzzy comprehensive evaluation method to obtain confirmation data. The fuzzy comprehensive evaluation method considers multiple factors, such as the operator's professionalism, the consistency of the confirmation operation, etc., and assigns a credibility score to each confirmation information.

[0127] Determine the set of evaluation factors and the set of comments. The set of evaluation factors includes various factors that affect the credibility of fault confirmation information, such as data source reliability, data collection time, data consistency, operator experience, etc. The set of comments is a qualitative description of credibility, such as "very credible", "relatively credible", "generally credible", "not very credible" and "uncredible". Next, it is necessary to establish a fuzzy relationship matrix, which describes the membership of each evaluation factor to different comments. For example, if the data source is very reliable, then its membership to the comment "very credible" may be very high. Then, it is necessary to determine the weight of each evaluation factor. This can be achieved through expert scoring, hierarchical analysis (AHP) or other methods. The distribution of weights reflects the importance of each factor in the overall judgment. For example, the reliability of the data source may be more important than the data collection time, so it will be given a higher weight. With the weight vector and the fuzzy relationship matrix, the comprehensive evaluation result can be obtained through fuzzy synthesis operation, which reflects the overall credibility of the fault confirmation information.

[0128] For example, suppose that the confirmation information of a bearing fault is being evaluated. The following factors may be considered: the quality of vibration data (F1), the consistency of temperature data (F2), the experience level of the operator (F3), and the matching degree of historical fault records (F4). The comment set can be {very credible, relatively credible, generally credible, not very credible, uncredible}. Through expert evaluation and historical data analysis, the following fuzzy relationship matrix may be obtained:

[0129] F1: [0.7, 0.2, 0.1, 0, 0]; F2: [0.5, 0.3, 0.2, 0, 0]; F3: [0.6, 0.3,0.1, 0, 0]; F4: [0.8, 0.1, 0.1, 0, 0].

[0130] Assuming that the weight vector obtained by the AHP method is [0.3, 0.2, 0.2, 0.3], then through the fuzzy synthesis operation, the final judgment result may be [0.68, 0.21, 0.11, 0, 0]. This means that the credibility of the fault confirmation information is very high, and there is a 68% probability that it is "very credible".

[0131] After obtaining the confirmed data, the next step is to perform data feature mapping to obtain the feature vector to be matched. This process involves converting the confirmed data into a standardized form that is easy to compare and match. The feature mapping method may vary depending on the application scenario, but it usually includes steps such as data normalization, dimensionality reduction, and feature extraction.

[0132] For example, principal component analysis (PCA) can be used to reduce the dimensionality of the data while retaining the most important information. Alternatively, deep learning methods such as autoencoders can be used to learn implicit feature representations of the data. In the bearing fault example, features such as peak frequency, RMS value, kurtosis, etc. may be extracted and combined into a feature vector. In addition, it is necessary to consider how to integrate credibility information into the feature vector. One approach is to use credibility as an additional dimension of the feature vector. Another approach is to use credibility to weight the importance of each feature. For example, if the credibility of a feature is high, its weight in the vector may be increased. The final feature vector to be matched should be a compact representation that can fully reflect the fault characteristics and credibility information. This vector will be used in the subsequent fault pattern matching process and compared with the known fault modes in the knowledge base to determine the most likely fault type and severity.

[0133] A knowledge graph is constructed for the distributed fault pattern library to obtain a fault knowledge graph. A knowledge graph is a semantic network that uses nodes to represent entities (such as fault type, component name) and edges to represent relationships between entities (such as causal relationships, subordination relationships). The fault knowledge graph is embedded and learned through a graph neural network algorithm to obtain a fault feature vector. Graph neural network algorithms, such as graph convolutional networks (GCNs) or graph attention networks (GATs), can capture local and global information in the graph structure and map the nodes and edges in the graph to a low-dimensional vector space.

[0134] The similarity between the feature vector to be matched and the fault feature vector is calculated to obtain the fault matching result. This step uses measurement methods such as cosine similarity or Euclidean distance to calculate the similarity between the feature vector to be matched and the fault feature vector to find the most matching fault type. Finally, the fault matching results are comprehensively evaluated through the multi-criteria decision-making algorithm to obtain the target fault data. The multi-criteria decision-making algorithm considers multiple evaluation indicators, such as matching degree, fault severity, repair difficulty, etc., to comprehensively evaluate and select the optimal fault diagnosis result.

[0135] For example, the spindle system of a CNC machine tool has abnormal vibration. The system first generates a three-dimensional fault model and marks the abnormal vibration area in red on the 3D model of the spindle. Using a spatial registration algorithm based on SIFT features, this 3D model is accurately superimposed on the actual machine tool image seen by the operator through AR glasses. The interactive interface allows the operator to zoom in and rotate the model through gestures to view detailed information. When the operator makes a "zoom in" gesture, the gesture recognition algorithm (CNN using ResNet50 architecture) recognizes it as a "zoom_in" instruction. Semantic parsing interprets this instruction as "view spindle bearing details". After the operator confirms the fault location, the system uses a fuzzy comprehensive evaluation method to comprehensively consider the operator's 10 years of work experience and decisiveness of the operation, and gives a credibility score of 0.9.

[0136] At the same time, the system builds a knowledge graph from a distributed library containing 10,000 fault cases, in which the "bearing fault" node is connected to symptom nodes such as "abnormal vibration" and "temperature increase". The graph is embedded and learned using the Graph Attention Network (GAT) to obtain a 128-dimensional fault feature vector. The cosine similarity between the operator's confirmation data and these feature vectors is calculated to obtain the three most similar fault types: bearing wear (similarity 0.92), insufficient bearing lubrication (similarity 0.85), and cracks in the inner ring of the bearing (similarity 0.78). Finally, the analytic hierarchy process (AHP) is used as a multi-criteria decision-making algorithm, taking into account the similarity, fault frequency, and potential harm, and finally "bearing wear" is determined as the target fault, and a detailed fault report and repair suggestions are generated.

[0137] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0138] (1) Evaluate the fault severity of the target fault data to obtain the fault risk level;

[0139] (2) Classify the fault risk levels through the decision tree algorithm and obtain a preliminary set of response plans;

[0140] (3) Conduct historical data correlation analysis on the preliminary response plan set to obtain the plan effect evaluation indicators;

[0141] (4) Optimize the calculation of the scheme effect evaluation index through the simulated annealing algorithm to obtain the optimal response plan;

[0142] (5) Conduct resource demand analysis on the optimal response plan and obtain a resource allocation list;

[0143] (6) Optimize the resource allocation list through linear programming algorithm to obtain the optimal resource scheduling plan;

[0144] (7) Arrange the optimal resource scheduling plan in time series to obtain the execution schedule of the response strategy;

[0145] (8) Dynamically adjust the response strategy execution schedule through reinforcement learning algorithms to obtain an adaptive execution plan;

[0146] (9) Assign human-machine collaborative tasks to the adaptive execution plan and obtain collaborative work instructions;

[0147] (10) Perform semantic optimization on the collaborative work instructions to obtain the target response strategy.

[0148] It should be noted that the target fault data is evaluated for fault severity to obtain the fault risk level. This step comprehensively considers the impact of the fault on production efficiency, product quality, and equipment life, and uses the preset scoring criteria to assign a risk level, such as 1-5, to the fault. Next, the fault risk level is classified and processed by the decision tree algorithm to obtain a preliminary response plan set. The decision tree algorithm selects multiple candidate solutions suitable for the current situation from the preset response plan library based on characteristics such as fault type and risk level. The preliminary response plan set is analyzed for historical data association to obtain the solution effect evaluation index. This step queries the historical fault processing records and calculates the success rate, average repair time, cost and other indicators of each candidate solution under similar circumstances. The simulated annealing algorithm is used to optimize the solution effect evaluation index to obtain the preferred response plan. The simulated annealing algorithm simulates the physical annealing process, performs random search in the solution space, and gradually converges to the global optimal solution, thereby selecting the response plan with the best comprehensive effect.

[0149] Perform resource demand analysis on the preferred response plan to obtain a resource allocation list. This step lists in detail the resources such as manpower, equipment, and materials required to implement the plan, and estimates the demand for each resource. Optimize the allocation of the resource allocation list through a linear programming algorithm to obtain the optimal resource scheduling plan. The linear programming algorithm maximizes resource utilization efficiency and minimizes costs under the premise of meeting various constraints (such as total resources and time limits). Arrange the optimal resource scheduling plan in time series to obtain a response strategy execution schedule. This step sorts the tasks according to logical order and resource availability, and formulates a detailed execution plan. Dynamically adjust the response strategy execution schedule through a reinforcement learning algorithm to obtain an adaptive execution plan. Reinforcement learning algorithms, such as Q-learning or policy gradient methods, can continuously learn through interaction with the environment and adjust plans in real time according to actual execution conditions to respond to emergencies.

[0150] Human-machine collaborative tasks are assigned to the adaptive execution plan to obtain collaborative work instructions. This step reasonably allocates tasks for manual and automated operations according to the complexity and risk level of the tasks to ensure that the repair work is completed safely and efficiently. Finally, natural language processing (NLP) technology is used to perform text analysis on the collaborative work instructions. This includes basic NLP tasks such as word segmentation, part-of-speech tagging, and named entity recognition. For example, tools such as Stanford NLP or spaCy can be used to identify key terms, action words, and numerical information in the instructions. Next, semantic analysis is required to understand the meaning of sentences and paragraphs. This can be done by using word vector models such as Word2Vec or GloVe to capture the semantic relationship between words, or using more advanced language models such as BERT or GPT to understand more complex contexts. After understanding the semantics of the text, the content needs to be structured and standardized. This may involve converting unstructured text into structured data formats such as JSON or XML. For example, the fault handling steps can be converted into a standardized flowchart or decision tree structure. This structuring not only helps with subsequent processing, but also improves the readability and executability of the instructions.

[0151] Domain knowledge graphs are needed to enrich and optimize content. Knowledge graphs contain relationships between entities such as equipment, fault types, and processing methods, which can help supplement missing information, correct possible errors, and provide more contextual information. For example, if a specific tool is mentioned in the guide, the knowledge graph can provide detailed specifications of the tool, precautions for use, and even possible alternative tools. A key step in semantic optimization is to personalize the content based on the specific fault situation and operating environment. This may involve using rule engines or machine learning models to determine which information is most relevant and which steps can be simplified or expanded in detail. For example, for experienced technicians, some basic instructions may be omitted; for novices, more explanations and warning information may be added.

[0152] In the optimization process, readability and comprehensibility also need to be considered. This includes using concise and clear language, avoiding ambiguity, and organizing information structure reasonably. Natural language generation (NLG) technology can be used to rephrase complex instructions to make them easier to understand. At the same time, multimedia elements such as pictures, videos or interactive charts can be introduced to enhance the effect of guidance.

[0153] Finally, the optimized content needs to be quality checked and verified. This can be done through automated consistency checks, expert review, or even simulated execution. For example, a text similarity algorithm can be used to check whether the optimized content still retains the key information of the original guide, or a natural language inference model can be used to verify whether the optimized steps are logically consistent.

[0154] For example, suppose there is a collaborative work instruction for a spindle failure of a CNC machine tool. The original instruction may contain a lot of technical terms and lengthy instructions. Through semantic optimization processing, it can be transformed into a more structured and personalized target response strategy.

[0155] First, the text is analyzed using NLP technology to identify key fault symptoms, diagnostic steps, and repair methods. Then, the knowledge graph is used to supplement the detailed information of the spindle structure and the possible causes of the failure. Next, the content is personalized according to the specific fault situation at hand (for example, abnormal vibration and temperature increase). Inspection steps related to vibration and temperature may be highlighted, and some specific measurement value ranges are added.

[0156] In terms of language, professional terms such as "perform bearing clearance inspection" may be converted into more detailed and understandable instructions, such as "use a feeler gauge to measure the gap between the inner and outer rings of the bearing, the gap should be between 0.02-0.05mm." An interactive 3D model may also be added, showing the structure of the spindle and the location of the inspection points.

[0157] Finally, the optimized content was checked for consistency to ensure that all steps were logically coherent and no key information was omitted. Through these steps, the original collaborative work instructions were transformed into a more clear, precise and easy-to-implement target response strategy, which not only improved the efficiency of troubleshooting, but also reduced the risk of operational errors.

[0158] The embodiment of the present invention also provides a system for identifying operating faults of a numerically controlled machine tool, such as Figure 2 As shown, the operation fault identification system of a CNC machine tool specifically includes:

[0159] The encryption module 201 is used to perform distributed edge acquisition and homomorphic encryption preprocessing on multi-source heterogeneous sensor data of CNC machine tools to obtain a feature data stream;

[0160] A verification module 202 is used to perform federated learning modeling and virtual environment verification on the characteristic data stream to obtain a digital twin model of normal operation of the machine tool;

[0161] An extraction module 203 is used to extract data from the digital twin model of the normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data and simulated fault data, and to perform reinforcement learning on the operation data set to obtain a distributed fault mode library;

[0162] The detection module 204 is used to perform edge computing anomaly detection on the real-time feature data stream collected and analyzed in real time through the normal operation digital twin model of the machine tool and the distributed fault mode library to obtain hierarchical and collaborative preliminary fault diagnosis data;

[0163] The matching module 205 is used to perform augmented reality-assisted interactive confirmation on the hierarchical collaborative preliminary fault diagnosis data, obtain confirmation data, perform data feature mapping on the confirmation data, obtain a feature vector to be matched, and perform fault matching based on the feature vector to be matched and the knowledge graph of the distributed fault mode library to obtain target fault data;

[0164] The analysis module 206 is used to perform fault response strategy analysis on the target fault data to obtain a target response strategy corresponding to the target fault data.

[0165] Through the collaborative work of the above modules, distributed edge collection and homomorphic encryption preprocessing technology have significantly improved the security and privacy protection capabilities of data collection, effectively preventing the leakage of sensitive information; the introduction of federated learning modeling and virtual environment verification mechanisms not only protects the data privacy of all participants, but also improves the generalization and robustness of the model; the reinforcement learning method is used to build a distributed fault mode library, which enhances the system's adaptive learning ability for new faults; through edge computing anomaly detection and multi-model fusion diagnosis, fast and accurate fault identification is achieved, greatly improving the real-time and reliability of the system; the fault confirmation and matching method combining augmented reality technology and knowledge graphs greatly improves the intuitiveness and efficiency of human-computer interaction, enabling operators to understand and handle fault situations more quickly and accurately; the fault matching mechanism based on multi-criteria decision-making improves the accuracy of fault diagnosis and interpretability; by introducing a variety of optimization algorithms such as decision trees, simulated annealing, and linear programming, the intelligent generation and dynamic adjustment of fault response strategies are realized, greatly improving the efficiency and pertinence of fault handling; the reinforcement learning algorithm is used to adaptively adjust the execution plan, enhancing the system's adaptability to complex and dynamic environments; through human-machine collaborative task allocation and knowledge distillation technology, human experience and machine intelligence are effectively combined to improve the overall effect of fault handling; the distributed architecture of this method and the application of edge computing technology have significantly improved the scalability and processing capabilities of the system, and can better adapt to large-scale and complex industrial environments; by building a complete fault identification and processing process, from data collection, twin construction, fault diagnosis to strategy formulation and execution, a closed-loop intelligent fault management system is formed, which greatly improves the operation reliability and maintenance efficiency of CNC machine tools.

[0166] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the embodiments, a person skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for identifying operation faults of a CNC machine tool, characterized in that: include: Distributed edge acquisition and homomorphic encryption preprocessing of multi-source heterogeneous sensor data of CNC machine tools are performed to obtain feature data streams; The characteristic data stream is subjected to federated learning modeling and virtual environment verification to obtain a digital twin model of the normal operation of the machine tool, including: data segmentation of the characteristic data stream to obtain multiple local data subsets; privacy protection processing of the local data subsets through a differential privacy algorithm to obtain a privacy-protected data subset; local twin construction of the privacy-protected data subset to obtain an initial local twin; model parameter aggregation of the initial local twin through a secure aggregation protocol to obtain a global twin; model distillation of the global twin to obtain a compressed twin; iterative optimization of the compressed twin through a federated learning algorithm, An optimized global twin is obtained; a virtual environment is constructed for the optimized global twin to obtain a virtual operating environment for a machine tool; a multi-scenario simulation is performed on the virtual operating environment for the machine tool by a Monte Carlo simulation method to obtain a simulation data set; the simulation data set is homomorphically encrypted to obtain encrypted simulation data, and the encrypted simulation data is input into the optimized global twin to obtain global twin output data, and a data label comparison is performed on the global twin output data and the simulation data set to obtain label difference data; parameter tuning is performed on the label difference data by a Bayesian optimization algorithm to obtain a digital twin model for the normal operation of the machine tool; Extracting data from the digital twin model of the normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data and simulated fault data, and performing reinforcement learning on the operation data set to obtain a distributed fault mode library; The real-time feature data stream collected and analyzed in real time is subjected to edge computing anomaly detection through the digital twin model of normal operation of the machine tool and the distributed fault mode library to obtain preliminary hierarchical and collaborative fault diagnosis data, including: homomorphically encrypting the real-time feature data stream to obtain a real-time encrypted data stream, and performing real-time segmentation processing on the real-time feature data stream to obtain time series data fragments; extracting dynamic features from the time series data fragments through a sliding window algorithm to obtain a dynamic feature sequence; extracting the operation feature sequence from the digital twin model of normal operation of the machine tool to obtain an operation feature sequence, and comparing and analyzing the dynamic feature sequence with the operation feature sequence to obtain preliminary anomaly indicators; quantifying the preliminary anomaly indicators through a local anomaly factor algorithm to obtain an anomaly degree score; and The abnormality degree score is used for threshold judgment to obtain an abnormal trigger signal and an abnormal type label; the distributed fault mode library is quickly searched based on the abnormal trigger signal and the abnormal type label to obtain a candidate fault mode set; the candidate fault mode set is matched in multiple dimensions by a fuzzy inference algorithm to obtain a multidimensional feature vector, and the fault time domain feature and the fault frequency domain feature are obtained. The fault time domain feature and the fault frequency domain feature can be used for fault mode membership analysis to obtain a fault possibility distribution; a hierarchical clustering analysis is performed on the fault possibility distribution to obtain a hierarchical fault category; a causal relationship inference is performed on the hierarchical fault category through a Bayesian network to obtain a fault propagation path; a comprehensive analysis is performed on the fault propagation path and the hierarchical fault category to obtain the preliminary fault diagnosis data of the hierarchical collaboration; Performing augmented reality-assisted interactive confirmation on the hierarchical collaborative preliminary fault diagnosis data to obtain confirmation data and performing data feature mapping on the confirmation data to obtain a feature vector to be matched, and performing fault matching based on the feature vector to be matched and the knowledge graph of the distributed fault mode library to obtain target fault data, including: visualizing the hierarchical collaborative preliminary fault diagnosis data to obtain a three-dimensional fault model; matching the three-dimensional fault model with the actual machine tool through a spatial registration algorithm to obtain an augmented reality overlay view; designing a human-computer interaction interface for the augmented reality overlay view to obtain an interactive fault display interface; and performing gesture recognition on the operator's interactive operations. Identify and obtain an interactive instruction sequence; perform semantic analysis on the interactive instruction sequence to obtain fault confirmation information; perform credibility analysis on the fault confirmation information through a fuzzy comprehensive evaluation method to obtain the confirmation data, and perform data feature mapping on the confirmation data to obtain a feature vector to be matched; construct a knowledge graph for the distributed fault mode library to obtain a fault knowledge graph; perform embedding learning on the fault knowledge graph through a graph neural network algorithm to obtain a fault feature vector; perform similarity calculation on the feature vector to be matched and the fault feature vector to obtain a fault matching result; perform a comprehensive evaluation on the fault matching result through a multi-criteria decision-making algorithm to obtain the target fault data; A fault response strategy analysis is performed on the target fault data to obtain a target response strategy corresponding to the target fault data.

2. The method for identifying operation faults of a CNC machine tool according to claim 1, characterized in that: The step of performing distributed edge acquisition and homomorphic encryption preprocessing on multi-source heterogeneous sensor data of CNC machine tools to obtain feature data streams includes: Distributed deployment of multi-source heterogeneous sensors of the CNC machine tool is performed to obtain an original sensor network; Allocating edge computing nodes to the original sensor network to obtain an edge computing network topology structure; The sensor data is collected in real time through the edge computing network topology structure to obtain the original data stream; Performing data cleaning and outlier detection on the original data stream to obtain a preprocessed data stream; Performing noise reduction processing on the preprocessed data stream by using a wavelet transform algorithm to obtain a noise-reduced data stream; Extracting time-domain and frequency-domain features of the noise reduction data stream to obtain a multi-dimensional feature vector; Performing dimensionality reduction processing on the multidimensional feature vector by using a principal component analysis algorithm to obtain an optimized feature vector; Performing data standardization and normalization processing on the optimized feature vector to obtain a standardized feature vector; The standardized feature vector is encrypted by a homomorphic encryption algorithm to obtain encrypted feature data; The encrypted feature data is distributedly stored and indexed to obtain the feature data stream.

3. The method for identifying operation faults of a CNC machine tool according to claim 1, characterized in that: The step of extracting data from the digital twin model of the normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data and simulated fault data, and performing reinforcement learning on the operation data set to obtain a distributed fault mode library includes: Setting parameter thresholds for the normal operation digital twin model of the machine tool to obtain a normal operation parameter range; Randomly sampling the normal operating parameter range by Monte Carlo method to obtain a sampling data set, and performing statistical deviation analysis on the sampling data to obtain operation deviation data, wherein the operation deviation data includes: mean deviation data, variance deviation data and distribution deviation data; Performing data cleaning and standardization processing on the historical fault data to obtain standardized historical fault data; Performing data enhancement on the normalized historical fault data by generating an adversarial network algorithm to obtain an expanded historical fault data set; Conducting a fault injection experiment on the digital twin model of the normal operation of the machine tool to obtain simulated fault data; Labeling the operation deviation data, the expanded historical fault data set, and the simulated fault data to obtain a comprehensive operation data set; Performing reinforcement learning training on the comprehensive operation data set by a deep Q learning algorithm to obtain an initial fault mode library; Compressing the initial fault mode library to obtain a lightweight fault mode library; Distributed storage and consensus mechanism construction are performed on the lightweight fault mode library to obtain the distributed fault mode library.

4. The method for identifying operation faults of a CNC machine tool according to claim 1, characterized in that: The step of performing fault response strategy analysis on the target fault data to obtain a target response strategy corresponding to the target fault data includes: Performing fault severity assessment on the target fault data to obtain a fault risk level; Classify the fault risk levels using a decision tree algorithm to obtain a preliminary response plan set; Performing historical data correlation analysis on the preliminary response plan set to obtain plan effect evaluation indicators; The effect evaluation index of the scheme is optimized and calculated by a simulated annealing algorithm to obtain a response plan; Conduct resource demand analysis on the response plan to obtain a resource allocation list; Optimizing the allocation of the resource allocation list through a linear programming algorithm to obtain an optimal resource scheduling solution; Arrange the optimal resource scheduling scheme in time series to obtain a response strategy execution schedule; Dynamically adjusting the execution schedule of the response strategy through a reinforcement learning algorithm to obtain an adaptive execution plan; Performing human-machine collaborative task assignment on the adaptive execution plan to obtain collaborative operation instructions; The collaborative work instruction is semantically optimized to obtain the target response strategy.

5. A system for identifying an operation fault of a numerically controlled machine tool, used to execute the method for identifying an operation fault of a numerically controlled machine tool according to any one of claims 1 to 4, characterized in that: include: The encryption module is used to perform distributed edge acquisition and homomorphic encryption preprocessing on multi-source heterogeneous sensor data of CNC machine tools to obtain feature data streams; A verification module is used to perform federated learning modeling and virtual environment verification on the feature data stream to obtain a digital twin model of the normal operation of the machine tool, including: data segmentation of the feature data stream to obtain multiple local data subsets; privacy protection processing of the local data subsets through a differential privacy algorithm to obtain a privacy-protected data subset; local twin construction of the privacy-protected data subset to obtain an initial local twin; model parameter aggregation of the initial local twin through a secure aggregation protocol to obtain a global twin; model distillation of the global twin to obtain a compressed twin; iterative analysis of the compressed twin through a federated learning algorithm Generation optimization to obtain an optimized global twin; construct a virtual environment for the optimized global twin to obtain a virtual operating environment for the machine tool; perform multi-scenario simulation on the virtual operating environment of the machine tool through a Monte Carlo simulation method to obtain a simulation data set; perform homomorphic encryption on the simulation data set to obtain encrypted simulation data, and input the encrypted simulation data into the optimized global twin to obtain global twin output data, and perform data label comparison on the global twin output data and the simulation data set to obtain label difference data; perform parameter tuning on the label difference data through a Bayesian optimization algorithm to obtain a digital twin model of the normal operation of the machine tool; An extraction module is used to extract data from the digital twin model of the normal operation of the machine tool to obtain an operation data set, wherein the operation data set includes: operation deviation data, historical fault data and simulated fault data, and to perform reinforcement learning on the operation data set to obtain a distributed fault mode library; A detection module is used to perform edge computing anomaly detection on the real-time feature data stream collected and analyzed in real time through the normal operation digital twin model of the machine tool and the distributed fault mode library to obtain hierarchical collaborative preliminary fault diagnosis data, including: homomorphically encrypting the real-time feature data stream to obtain a real-time encrypted data stream, and performing real-time segmentation processing on the real-time feature data stream to obtain time series data fragments; extracting dynamic features from the time series data fragments through a sliding window algorithm to obtain a dynamic feature sequence; extracting an operation feature sequence from the normal operation digital twin model of the machine tool to obtain an operation feature sequence, and comparing and analyzing the dynamic feature sequence with the operation feature sequence to obtain a preliminary anomaly indicator; quantifying the preliminary anomaly indicator through a local anomaly factor algorithm to obtain an anomaly degree score; Perform threshold judgment on the abnormality degree score to obtain an abnormal trigger signal and an abnormal type label; perform rapid retrieval on the distributed fault mode library based on the abnormal trigger signal and the abnormal type label to obtain a candidate fault mode set; perform multi-dimensional matching on the candidate fault mode set through a fuzzy inference algorithm to obtain a multi-dimensional feature vector, obtain fault time domain features and fault frequency domain features, and perform fault mode membership analysis on the fault time domain features and fault frequency domain features to obtain a fault possibility distribution; perform hierarchical clustering analysis on the fault possibility distribution to obtain hierarchical fault categories; perform causal relationship inference on the hierarchical fault categories through a Bayesian network to obtain a fault propagation path; perform a comprehensive analysis on the fault propagation path and the hierarchical fault categories to obtain the hierarchical collaborative preliminary fault diagnosis data; A matching module is used to perform augmented reality-assisted interactive confirmation on the preliminary fault diagnosis data of the hierarchical collaboration, obtain confirmation data, perform data feature mapping on the confirmation data, obtain feature vectors to be matched, and perform fault matching based on the feature vectors to be matched and the knowledge graph of the distributed fault mode library to obtain target fault data, including: visualizing the preliminary fault diagnosis data of the hierarchical collaboration to obtain a three-dimensional fault model; matching the three-dimensional fault model with the actual machine tool through a spatial registration algorithm to obtain an augmented reality overlay view; designing a human-computer interaction interface for the augmented reality overlay view to obtain an interactive fault display interface; and performing gesture recognition on the operator's interaction The operation is identified to obtain an interactive instruction sequence; the interactive instruction sequence is semantically parsed to obtain fault confirmation information; the fault confirmation information is credibility analyzed by a fuzzy comprehensive evaluation method to obtain the confirmation data, and the confirmation data is mapped to data features to obtain a feature vector to be matched; a knowledge graph is constructed for the distributed fault mode library to obtain a fault knowledge graph; the fault knowledge graph is embedded and learned by a graph neural network algorithm to obtain a fault feature vector; similarity is calculated between the feature vector to be matched and the fault feature vector to obtain a fault matching result; a comprehensive evaluation is performed on the fault matching result by a multi-criteria decision-making algorithm to obtain the target fault data; The analysis module is used to perform fault response strategy analysis on the target fault data to obtain a target response strategy corresponding to the target fault data.

Images