State monitoring method and system for multi-axis linkage CNC machining

Abstract

The present invention relates to the technical field of multi-axis linkage CNC machining, and discloses a state monitoring method and system for multi-axis linkage CNC machining, wherein the method comprises: collecting and preprocessing vibration, acoustic emission and cutting force signals in the multi-axis linkage CNC machining process to obtain first multi-source sensor data; performing feature extraction and compressed sensing to obtain a low-dimensional feature subset; performing fusion judgment to obtain abnormal state detection results; training a deep learning model including a backbone network and multiple branch networks to obtain a state evaluation model; inputting the second multi-source sensor data collected in real time into the state evaluation model for state evaluation to obtain a machining state classification result and parameter estimation value. The method realizes the collaborative state evaluation and prediction simulation of multiple machine tools.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-axis linkage numerical control machining, and in particular to a state monitoring method and system for multi-axis linkage numerical control machining. Background Art

[0002] Multi-axis linkage CNC machining technology is widely used in modern manufacturing due to its high efficiency, high precision and flexibility. However, with the increase in machining complexity and the improvement of quality requirements, traditional machining state monitoring methods face many challenges. The multi-axis linkage machining process involves the coordinated control of multiple motion axes, and the changes in machining state are more complex and difficult to predict. A single sensor or simple signal processing method is difficult to fully capture the dynamic characteristics of the machining process.

[0003] Secondly, there is a lot of noise and interference in the processing, such as machine tool vibration, environmental factors, etc. These factors will affect the quality of sensor data and lead to inaccurate condition monitoring results. In addition, traditional condition monitoring methods often rely on empirical thresholds or simple statistical models, which are difficult to adapt to the diverse needs of different workpieces and different processing technologies, and lack adaptability and robustness. Summary of the invention

[0004] The present invention provides a state monitoring method and system for multi-axis linkage numerical control machining, which are used to realize collaborative state evaluation and prediction simulation of multiple machine tools.

[0005] In a first aspect, the present invention provides a state monitoring method for multi-axis linkage CNC machining, the state monitoring method for multi-axis linkage CNC machining comprising:

[0006] The vibration, acoustic emission and cutting force signals in the multi-axis linkage CNC machining process are collected and preprocessed to obtain the first multi-source sensor data;

[0007] Inputting the first multi-source sensor data into a state-space model builder for processing to obtain a state-space model;

[0008] According to the state space model, feature extraction and compressed sensing are performed on the first multi-source sensor data to obtain a low-dimensional feature subset;

[0009] Calculating the reconstruction error and Euclidean distance of the low-dimensional feature subset, and performing fusion judgment to obtain an abnormal state detection result;

[0010] A deep learning model including a backbone network and a plurality of branch networks is trained according to the abnormal state detection result and the low-dimensional feature subset to obtain a state assessment model;

[0011] The second multi-source sensor data collected in real time is input into the state assessment model for state assessment to obtain processing state classification results and parameter estimation values, and the processing state classification results and parameter estimation values ​​are mapped to the digital twin model to obtain distributed collaborative monitoring results.

[0012] In a second aspect, the present invention provides a state monitoring system for multi-axis linkage CNC machining, the state monitoring system for multi-axis linkage CNC machining comprising:

[0013] An acquisition module is used to acquire and pre-process the vibration, acoustic emission and cutting force signals in the multi-axis linkage CNC machining process to obtain first multi-source sensor data;

[0014] A processing module, configured to input the first multi-source sensor data into a state space model builder for processing to obtain a state space model;

[0015] A compression module, configured to perform feature extraction and compressive sensing on the first multi-source sensor data according to the state space model to obtain a low-dimensional feature subset;

[0016] A judgment module is used to calculate the reconstruction error and Euclidean distance of the low-dimensional feature subset, and perform fusion judgment to obtain an abnormal state detection result;

[0017] A training module, used for training a deep learning model including a trunk network and a plurality of branch networks according to the abnormal state detection result and the low-dimensional feature subset to obtain a state assessment model;

[0018] An evaluation module is used to input the second multi-source sensor data collected in real time into the state evaluation model for state evaluation, obtain processing state classification results and parameter estimation values, and map the processing state classification results and the parameter estimation values ​​into the digital twin model to obtain distributed collaborative monitoring results.

[0019] The third aspect of the present invention provides a state monitoring device for multi-axis linkage CNC machining, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the state monitoring device for multi-axis linkage CNC machining executes the above-mentioned state monitoring method for multi-axis linkage CNC machining.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned state monitoring method for multi-axis linkage CNC machining.

[0021] In the technical solution provided by the present invention, the ability to capture the dynamic characteristics of the machining process is improved through multi-source sensor data fusion and state space model construction, making the state monitoring more comprehensive and accurate. By adopting feature extraction and compressed sensing processing technology, a low-dimensional feature subset with sparsity constraints is obtained, which effectively reduces the data dimension and improves the calculation efficiency. The abnormal detection method combined with the autoencoder network and Euclidean distance calculation enhances the recognition ability of abnormal states and reduces false positives and false negatives. By using a deep learning model including a backbone network and multiple branch networks, adaptive processing of different types of sensor signals is achieved, and the accuracy and generalization ability of state assessment are improved. The introduction of attention mechanism and multi-task learning enhances the model's perception of key features, and at the same time realizes the collaborative optimization of state classification and parameter estimation. Through the construction and updating of the digital twin model, virtual-real fusion state monitoring is realized, which provides a basis for visualization and predictive analysis of the machining process. Based on the distributed collaborative monitoring results, the collaborative state evaluation and predictive simulation of multiple machine tools are realized, which provides support for the overall optimization and intelligent decision-making of the production line.

[0022] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of an embodiment of a state monitoring method for multi-axis linkage CNC machining in an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of an embodiment of a state monitoring system for multi-axis linkage CNC machining in an embodiment of the present invention;

[0026] Figure 3 It is a schematic diagram of an embodiment of a state monitoring device for multi-axis linkage CNC machining in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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.

[0028] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0029] To facilitate understanding of this embodiment, a state monitoring method for multi-axis linkage CNC machining disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps:

[0030] 101. Collect and preprocess vibration, acoustic emission and cutting force signals in a multi-axis linkage CNC machining process to obtain first multi-source sensor data;

[0031] It is understandable that the execution subject of the present invention may be a state monitoring system for multi-axis CNC machining, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject.

[0032] Specifically, the vibration signal generated during the multi-axis linkage CNC machining process is collected by a vibration sensor, and the collected vibration signal is subjected to bandpass filtering to effectively filter the low-frequency and high-frequency noise in the vibration signal, retain the intermediate frequency information that is most critical for state monitoring, so as to improve the signal quality and the accuracy of subsequent feature extraction, and obtain the filtered vibration signal. Similarly, the acoustic emission signal during the multi-axis linkage CNC machining process is collected by an acoustic emission sensor, and is subjected to bandpass filtering to obtain the filtered acoustic emission signal. The acoustic emission signal reflects physical phenomena such as microcracks, friction, and chip breaking during the machining process. Bandpass filtering can effectively extract useful information related to the machining, while reducing the interference of environmental noise on the monitoring results. The cutting force signal during the machining process is collected by a cutting force sensor, and the cutting force signal is subjected to bandpass filtering to obtain the filtered cutting force signal. The cutting force signal is an important parameter that directly reflects the interaction force between the tool and the workpiece, and can provide important information about the stability of the machining process, the tool state, and the workpiece quality. The filtered vibration, acoustic emission, and cutting force signals are analyzed in the time domain to extract their time domain features. The time domain characteristics reflect the amplitude, change trend and fluctuation of the signal. By performing time domain analysis on the filtered vibration signal, the time domain characteristics of the vibration signal, such as mean, root mean square value, peak value, kurtosis and skewness, are extracted. Similarly, the filtered acoustic emission signal is analyzed in the time domain to obtain the time domain characteristics of the acoustic emission signal, such as energy, number of pulses and duration, which effectively reflect the changes in tool wear and workpiece surface quality. The filtered cutting force signal is analyzed in the time domain to extract the time domain characteristics of the cutting force signal, such as average cutting force, cutting force fluctuation amplitude, etc., so as to judge the degree of tool wear and whether the cutting conditions are normal. The filtered vibration, acoustic emission and cutting force signals are fast Fourier transformed respectively to obtain the frequency domain feature set of these signals. Fast Fourier transform is an effective method to convert signals from time domain to frequency domain, which can reveal the energy distribution of signals at different frequencies. By performing fast Fourier transform on the vibration signal, a spectrum diagram is obtained, and the main frequency components of the signal are analyzed to identify the source of vibration and abnormal frequency; the acoustic emission signal is fast Fourier transformed to obtain the frequency domain feature set of the acoustic emission signal, identify the characteristic frequency band of the acoustic emission, and determine the characteristics such as cracks and friction; the cutting force signal is fast Fourier transformed to obtain the characteristics of the cutting force signal in the frequency domain. These frequency domain characteristics help to judge the resonance phenomenon and tool status during the cutting process. The filtered vibration signal time domain characteristics, acoustic emission signal time domain characteristics, cutting force signal time domain characteristics and all signal frequency domain feature sets are combined to obtain the complete first multi-source sensor data. This data contains important characteristics of multiple physical quantities in the time domain and frequency domain, and can comprehensively and accurately describe the state changes in the multi-axis linkage CNC machining process.

[0033] 102. Input the first multi-source sensor data into a state space model builder for processing to obtain a state space model;

[0034] Specifically, autocorrelation analysis and cross-correlation analysis are performed on the first multi-source sensor data. The time correlation characteristics of each sensor signal are obtained through autocorrelation analysis, so as to understand the temporal variation pattern and internal law of each signal. The correlation characteristics between different sensor signals are obtained through cross-correlation analysis, revealing the mutual influence and coupling relationship between vibration, acoustic emission and cutting force signals in the processing process. Based on the time correlation characteristics and correlation characteristics, dynamic time warping is performed on the first multi-source sensor data. Dynamic time warping is a technology for aligning time series data, which can time align multi-source signals with asynchronous acquisition time, thereby eliminating errors caused by sampling time difference or sensor working inconsistency. Through dynamic time warping, an aligned multi-source sensor data sequence is obtained, so that each sensor signal is in the same reference system on the time axis. The aligned multi-source sensor data sequence is segmented by sliding windows to obtain multiple overlapping data segments. The sliding window technology is used to segment a long data sequence into smaller subsequences, each of which has a certain overlap with each other, to ensure that important information is not missed during the segmentation process, and to improve the stability and robustness of feature extraction. For each overlapping data segment, singular value decomposition is performed to decompose the data segment into a set of eigenvectors and corresponding singular values. The singular value reflects the information energy contained in the data, while the eigenvector describes the change of the data in all directions. By extracting the main eigenvectors and the corresponding singular values, the dimensionality reduction of the data can be effectively achieved, the noise interference can be removed, the most representative state information can be retained, and the state representation after dimensionality reduction can be obtained. Based on the state representation after dimensionality reduction, the initial state transfer matrix is ​​constructed. The state transfer matrix is ​​used to describe the change relationship between the system state at adjacent moments. By dynamically analyzing the state representation after dimensionality reduction, the transfer law of the system between different time steps is obtained, and the initial state transfer matrix is ​​constructed based on this. At the same time, the initial observation matrix is ​​estimated using the least squares method. The least squares method is an effective parameter estimation method that can give the optimal observation matrix estimation result on the basis of minimizing the error and obtain the initial state space model. The initial state space model and the aligned multi-source sensor data sequence are input into the unscented Kalman filter for iterative parameter optimization. The unscented Kalman filter is a tool for state estimation of nonlinear systems. It can introduce random noise in the state transfer and observation process, and optimize the state transfer matrix and observation matrix through iterative updates to improve the accuracy of the model. Through the iterative optimization of the unscented Kalman filter, the optimized state transfer matrix and observation matrix are obtained. Based on the optimized state transfer matrix and observation matrix, the nonlinear state equation and observation equation are constructed, and process noise and observation noise are introduced, so that the model can more realistically reflect the complex dynamic characteristics of the multi-axis linkage CNC machining process.Process noise is used to describe the random interference of the system during the state transition process, while observation noise reflects the error in sensor measurement. By introducing these noises, the robustness and generalization ability of the model can be enhanced. A state space model representing the relationship between multi-source data at adjacent moments is obtained, which can effectively describe the dynamic evolution law in the multi-axis linkage CNC machining process.

[0035] 103. Perform feature extraction and compressive sensing on the first multi-source sensor data according to the state space model to obtain a low-dimensional feature subset;

[0036] Specifically, the first multi-source sensor data is input into the state space model, and the state estimation is performed by operating the state space model to obtain a state estimation sequence reflecting the dynamic characteristics of the machining process. The state estimation sequence is multi-scale decomposed to construct a hierarchical feature structure. Through multi-scale decomposition, the state estimation sequence is decomposed into multiple subsequences containing information of different time scales, and a multi-level feature set is obtained, which contains a variety of features from short-term changes to long-term trends, and can fully reflect the dynamic behavior in the machining process, and is conducive to capturing potential abnormalities or state changes at different time scales. Based on the multi-level feature set, an adaptive feature selection strategy is created to screen key features with discriminative ability for state monitoring. The adaptive feature selection strategy dynamically adjusts the weight of each feature according to the discriminative ability and information gain of the feature, and obtains a preliminarily screened feature subset. A feature correlation network is constructed to perform correlation analysis on the features in the preliminarily screened feature subset. Through correlation analysis, highly correlated redundant features and mutually complementary features in the subset are identified. Redundant features will lead to information duplication and computational burden, while complementary features can provide more dimensional information about the machining state. Based on the analysis results of the feature correlation network, feature optimization is performed to obtain an optimized feature combination. Based on the machining process knowledge, a feature enhancement mechanism is introduced to enhance the optimized feature combination. The machining process knowledge can provide prior information about the physical mechanism in the machining process. By combining this prior knowledge, the features are enhanced in a targeted manner to better reflect the key states in the machining process. The sparse constraint regularization term is introduced to perform sparse optimization on the enhanced feature set, remove irrelevant and weakly related features, and finally obtain a sparse feature representation. Sparse optimization can effectively compress the dimension of the feature set by regularizing the feature weights, making the retained features more sparse while retaining key information, thereby improving the computational efficiency and generalization ability of the condition monitoring model. Based on the feature importance evaluation mechanism, and combined with the machining quality indicators and condition monitoring objectives, the features in the sparse feature representation are ranked by importance to obtain a feature importance sequence. The feature importance evaluation mechanism determines the relative importance of each feature in the monitoring task by analyzing the contribution of the feature to the machining state classification and anomaly detection. Based on the feature importance sequence, the top N features whose cumulative importance reaches the preset threshold are selected to ensure that the total contribution of the retained features to condition monitoring reaches a certain level. Combined with dimensionality reduction techniques, the selected features are processed to obtain a low-dimensional feature subset with sparsity constraints.

[0037] 104. Calculate the reconstruction error and Euclidean distance of the low-dimensional feature subset, and perform fusion judgment to obtain the abnormal state detection result;

[0038] Specifically, the low-dimensional feature subset is normalized to eliminate the differences between different feature dimensions and obtain a normalized feature vector. The normalized feature vector is input into a pre-trained autoencoder network. The structure of the autoencoder network consists of an encoder and a decoder, where the encoder maps the input features to a latent space and extracts its inherent implicit features, while the decoder reconstructs the representation in the latent space back to the original feature space. The autoencoder can learn the feature distribution of the input data and reconstruct it during the decoding process to capture the main features of the input data. After being processed by the encoder and decoder, the reconstructed feature vector is obtained. In order to evaluate the effect of reconstruction, the mean square error between the normalized feature vector and the reconstructed feature vector is calculated. The mean square error reflects the degree of deviation in the reconstruction process. The larger the reconstruction error, the greater the difference between the input feature and the feature pattern learned by the model. The reconstruction error is used as the first indicator for abnormal state detection. By analyzing the reconstruction error, those states that deviate significantly from the normal state pattern are effectively identified. At the same time, the changes in the low-dimensional feature subset in the time series are analyzed to capture potential dynamic anomalies. The low-dimensional feature subsets of adjacent sampling periods are processed by sliding windows, with the window size set to N and the sliding step size set to 1, to obtain a feature sequence matrix. Through sliding window processing, the dynamic changes of the features are tracked in a fine-grained manner to ensure that no abnormal signals are missed. After obtaining the feature sequence matrix, the Euclidean distance between two adjacent feature vectors in the matrix is ​​calculated. The Euclidean distance reflects the magnitude of feature changes between adjacent sampling moments. If the Euclidean distance is large, it means that the feature has changed significantly in a short period of time. The Euclidean distance is used as the second indicator for abnormal state detection to measure the magnitude of changes in the system state over time. Based on the multi-index fusion rule of fuzzy logic, the reconstruction error and Euclidean distance are used as inputs, and the fused abnormality index is obtained through fuzzy reasoning. Fuzzy logic can handle uncertainty and ambiguity problems, map the specific values ​​of the reconstruction error and Euclidean distance into fuzzy sets, and reason them through fuzzy rules to obtain a comprehensive abnormality index that reflects the health of the system state. A dynamic threshold adjustment mechanism is constructed to generate a detection threshold that is updated in real time. The dynamic threshold adjustment mechanism can dynamically adjust the threshold according to the state changes in the current processing process, adapt to the characteristic changes under different processing conditions and working conditions, and avoid missed detection or false detection problems caused by fixed thresholds. Through this mechanism, the threshold is automatically adjusted according to the changing trend of real-time data, making the detection result more reliable. The fused abnormality index is compared with the dynamically adjusted detection threshold. If the abnormality index exceeds the detection threshold updated in real time, the current state is judged to be an abnormal state, indicating that there is equipment failure, severe tool wear or other abnormal processing phenomena in the processing process; if the abnormality index does not exceed the threshold, it is judged to be a normal state, and the final abnormal state detection result is obtained.

[0039] 105. According to the abnormal state detection results and the low-dimensional feature subset, a deep learning model including a trunk network and multiple branch networks is trained to obtain a state evaluation model;

[0040] Specifically, the abnormal state detection results are divided into a normal state sample set and an abnormal state sample set. By performing data enhancement processing on these two sample sets, the scale of the training data set is effectively expanded to obtain the expanded training data set. The purpose of data enhancement is to increase the diversity of samples, thereby improving the generalization ability of the model and avoiding model overfitting. The backbone network of the deep learning model is constructed, and the backbone network adopts the DenseNet structure. DenseNet is a convolutional neural network structure that enhances feature transfer and reuse through dense connections. It contains 4 dense blocks, each of which contains 6 convolutional layers. The size of the convolution kernel is 3×3, the step size is 1, and the padding is 1. Each convolution layer in the dense block is connected by dense connections, so that each layer directly receives the input of all previous layers, realizing the maximum reuse of features and efficient propagation of gradients. A transition layer is added between each dense block. The transition layer contains a 1×1 convolution layer and a 2×2 average pooling layer. The 1×1 convolution layer is used to compress the number of channels, while the 2×2 average pooling layer is used to reduce the spatial resolution of the feature map, thereby reducing the number of model parameters, improving computational efficiency, and effectively preventing overfitting. Multiple branch networks are constructed, each corresponding to a sensor signal type, such as vibration signal, acoustic emission signal and cutting force signal. The branch network adopts the ResNet structure, which contains three residual blocks. Each residual block contains two 3×3 convolutional layers and a short-circuit connection. The introduction of residual blocks effectively solves the gradient vanishing problem in deep networks and enhances the feature learning ability of the network. The short-circuit connection enables the network to directly transfer the input jump to the output, making the network more stable during the training process. In order to improve the accuracy of feature fusion, an attention mechanism module is added between the backbone network and multiple branch networks. The attention mechanism module includes two sub-modules: channel attention and spatial attention. The channel attention is used to adaptively adjust the weights of different feature channels, so that the network can pay more attention to the feature channels that contribute most to state evaluation; the spatial attention is used to adjust the weights of the feature map in the spatial dimension, so that the network can pay attention to the key areas in the feature map. The attention mechanism helps the network to adaptively adjust the importance of different features, improve the effectiveness of feature fusion and the network's ability to identify complex processing states. The low-dimensional feature subset in the expanded training data set is input into the backbone network, and the first backbone feature map is obtained after being processed by DenseNet. At the same time, the corresponding sensor signal features are input into the corresponding branch network, and the first branch feature map is obtained after being processed by ResNet. Then, the first backbone feature map and the first branch feature map are fused through the attention mechanism module to obtain the first fused feature map. A multi-task learning head is constructed, which includes the state classification task and the parameter regression task. The state classification task uses a fully connected layer and a Softmax layer. The probability distribution of different states is obtained through the Softmax layer to realize the classification of the processing state.The parameter regression task uses a fully connected layer and a linear activation function to estimate the key parameters related to the state during the processing. After the first fused feature map is input into the multi-task learning head, the state classification result and the parameter regression result are calculated. Through multi-task learning, the network learns classification and regression tasks at the same time, and uses the complementary information between the two to improve the learning effect of the overall model. During the model training process, a weighted loss function is used to optimize the network parameters. The weighted loss function contains the cross entropy loss of state classification and the mean square error loss of parameter regression. The cross entropy loss is used to measure the difference between the classification result and the true label, while the mean square error loss is used to measure the error between the regression result and the true parameter. By assigning weights to the losses of different tasks, the training objectives between state classification and parameter regression are balanced, so that the model can achieve good performance in both aspects at the same time. During the training process, the network parameters are updated through the back propagation algorithm and the Adam optimizer. The Adam optimizer is an adaptive learning rate optimization method that can dynamically adjust the learning rate according to the first-order and second-order moment estimates of the gradient, thereby accelerating convergence and improving the training efficiency of the model. Through the deep learning model training process, a state assessment model is obtained, which can effectively classify the states in the multi-axis linkage CNC machining process and estimate the parameters related to the machining state.

[0041] 106. Input the second multi-source sensor data collected in real time into the state assessment model for state assessment to obtain processing state classification results and parameter estimation values, and map the processing state classification results and parameter estimation values ​​into the digital twin model to obtain distributed collaborative monitoring results.

[0042] Specifically, the second multi-source sensor data collected in real time is collected and preprocessed to obtain preprocessed real-time sensor data. The sensor data is subjected to preprocessing steps such as denoising and normalization to eliminate noise and dimensional differences to ensure data consistency and reliability. The preprocessed real-time sensor data is input into the state space model to perform state estimation to obtain a real-time state sequence. The state estimation process can infer the dynamic state in the processing process in real time and provide the latest description of the system state. The real-time state sequence is subjected to feature extraction and compressed sensing processing to obtain a real-time feature subset. Through compressed sensing, the data dimension is reduced while ensuring data accuracy, the calculation efficiency is improved and the processing complexity is reduced. The real-time feature subset is input into the backbone network of the state assessment model, and the second backbone feature map is generated after being processed by the DenseNet structure. At the same time, various sensor signals in the preprocessed real-time sensor data are respectively input into the corresponding branch networks, and multiple second branch feature maps are obtained through the calculation of the ResNet structure. The second backbone feature map and multiple second branch feature maps are feature fused by the attention mechanism module to obtain a second fused feature map. The attention mechanism module adaptively adjusts the weight of the features so that different features receive different attention according to their importance during the fusion process, thereby improving the effectiveness of feature fusion and the accuracy of state monitoring. The second fusion feature map is input into the multi-task learning head of the state assessment model, and the processing state classification result is obtained through the state classification task, and the parameter estimation value is obtained through the parameter regression task. The state classification task classifies the processing state through the Softmax layer, determines whether the processing process is in a normal state or an abnormal state, and specifically distinguishes the abnormal type; while the parameter regression task obtains the estimated values ​​of key parameters in the processing process through the linear activation function, such as cutting force, vibration amplitude, etc. These parameters play an important role in judging processing quality and process control. Construct a digital twin model for virtual mapping and simulation of the physical processing system. The digital twin model includes geometric model, physical model and behavioral model. The geometric model represents the three-dimensional structure of the machine tool and workpiece, and describes the appearance and size of the processing object. The physical model is used to describe the mechanical and thermodynamic characteristics of the processing process, such as cutting force, stress strain, temperature distribution, etc. The physical model is used to simulate the change law of physical quantities in the processing process. The behavioral model is used to simulate the movement and control logic of each component of the machine tool, including the linkage control of each axis, tool path planning, etc. By constructing these three sub-models, a comprehensive digital simulation of the CNC machining process is achieved. The machining state classification results and parameter estimation values ​​are mapped to the digital twin model, and the state and parameters of the digital twin model are updated. The actual state and parameters in the physical system are transferred to the virtual model in real time, so that the digital twin model can reflect the latest operating status of the machining system.In the digital twin model, the geometric model is updated in real time according to the processing status, the parameters in the physical model are modified according to the actual processing conditions, and the behavioral model simulates the behavioral changes of the processing process according to the current control status and parameters. Through mapping, the digital twin model can keep pace with the actual physical system and realize real-time monitoring and virtual simulation of the processing process. Based on the updated digital twin model, multi-machine tool collaborative state evaluation and predictive simulation are carried out to obtain distributed collaborative monitoring results. By integrating the digital twin models of multiple machine tools, collaborative state monitoring of the entire processing workshop is realized. In the collaborative evaluation, the current state of each machine tool is obtained, and the future state trend is inferred through predictive simulation, such as predicting the tool wear trend of a certain machine tool, the error accumulation of a certain processing path, etc. By analyzing the operating status and predicted trends of multiple machine tools, collaborative optimization suggestions are put forward, such as how to adjust the workload of each machine tool to improve the overall processing efficiency, or how to reduce the processing error through coordinated control.

[0043] In the embodiment of the present invention, the ability to capture the dynamic characteristics of the machining process is improved through multi-source sensor data fusion and state space model construction, making the state monitoring more comprehensive and accurate. By adopting feature extraction and compressed sensing processing technology, a low-dimensional feature subset with sparsity constraints is obtained, which effectively reduces the data dimension and improves the calculation efficiency. The abnormal detection method combined with the autoencoder network and Euclidean distance calculation enhances the recognition ability of abnormal states and reduces false positives and false negatives. By using a deep learning model including a backbone network and multiple branch networks, adaptive processing of different types of sensor signals is achieved, and the accuracy and generalization ability of state assessment are improved. The introduction of attention mechanism and multi-task learning enhances the model's perception of key features, and at the same time realizes the collaborative optimization of state classification and parameter estimation. Through the construction and update of the digital twin model, virtual-real fusion state monitoring is realized, which provides a basis for visualization and predictive analysis of the machining process. Based on the distributed collaborative monitoring results, the collaborative state assessment and predictive simulation of multiple machine tools are realized, which provides support for the overall optimization and intelligent decision-making of the production line.

[0044] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0045] The vibration signal in the multi-axis linkage CNC machining process is collected by a vibration sensor, and band-pass filtering is performed to obtain a filtered vibration signal; the acoustic emission signal in the multi-axis linkage CNC machining process is collected by an acoustic emission sensor, and band-pass filtering is performed to obtain a filtered acoustic emission signal; the cutting force signal in the multi-axis linkage CNC machining process is collected by a cutting force sensor, and band-pass filtering is performed to obtain a filtered cutting force signal;

[0046] Performing time domain analysis on the filtered vibration signal to obtain the time domain characteristics of the vibration signal; performing time domain analysis on the filtered acoustic emission signal to obtain the time domain characteristics of the acoustic emission signal; performing time domain analysis on the filtered cutting force signal to obtain the time domain characteristics of the cutting force signal;

[0047] Performing fast Fourier transform on the filtered vibration signal, the filtered acoustic emission signal and the filtered cutting force signal respectively to obtain a frequency domain feature set;

[0048] The time domain features of the vibration signal, the time domain features of the acoustic emission signal, the time domain features of the cutting force signal and the frequency domain features are combined to obtain the first multi-source sensor data.

[0049] Specifically, vibration sensors are used to collect vibration signals during multi-axis linkage CNC machining. During machining, vibration signals can reflect the dynamic characteristics of the interaction between the tool and the workpiece. Changes in vibration usually indicate abnormalities during machining, such as tool wear, unstable workpiece clamping, etc. In order to extract vibration information closely related to the machining state, the collected vibration signals are bandpass filtered to remove low-frequency and high-frequency noise in the signals, and only those characteristic frequency bands related to the machining process are retained to obtain filtered vibration signals. Acoustic emission signals during multi-axis linkage CNC machining are collected by acoustic emission sensors. Acoustic emission signals are high-frequency elastic waves during material deformation, crack generation and expansion, and can reflect abnormal changes during machining. The collected acoustic emission signals are bandpass filtered to filter out environmental noise and the interference noise of the sensor itself, and obtain filtered acoustic emission signals. For cutting force signals, cutting force sensors are used to collect them. The cutting force signals reflect the interaction force between the tool and the workpiece, and are important parameters for judging tool status and machining stability. The cutting force signal is bandpass filtered to effectively remove low-frequency machine tool vibration and high-frequency electrical noise to obtain a filtered cutting force signal. The filtered vibration signal, acoustic emission signal and cutting force signal are analyzed in the time domain to extract the time domain characteristics of the signal. Time domain analysis is a basic step in signal analysis, which is mainly used to extract the amplitude, change trend and fluctuation of the signal. Taking the vibration signal as an example, common time domain characteristics include mean, root mean square value, peak value, kurtosis and skewness. The mean can reflect the average level of the signal and is used to evaluate the overall vibration of the system. The root mean square value is used to describe the energy of the signal, and its formula is:

[0050] ;

[0051] The root mean square value reflects the average amplitude of the vibration signal within a certain period of time and is an important indicator for measuring vibration intensity. The peak value and kurtosis are used to evaluate the fluctuation characteristics and sharpness of the signal, so as to determine whether there are violent fluctuations or abnormal impacts during the processing. The filtered acoustic emission signal and cutting force signal are analyzed in the time domain to obtain the time domain characteristics reflecting the signal energy, fluctuation amplitude and other characteristics. After completing the time domain analysis, the filtered vibration signal, acoustic emission signal and cutting force signal are respectively subjected to fast Fourier transform to extract the frequency domain characteristics of the signal. Fast Fourier transform is a method of converting time domain signals into frequency domain representation, revealing the energy distribution of the signal at different frequencies. Frequency domain characteristics play an important role in analyzing the periodicity and spectral characteristics of the signal. For example, after fast Fourier transforming the vibration signal, the amplitude distribution of the signal at different frequencies is obtained, and then the main vibration frequency in the processing process is identified. If abnormal vibration occurs during the processing, it is usually manifested in the spectrum as a significant increase in the amplitude of a specific frequency component, which provides a basis for fault diagnosis. The basic formula of fast Fourier transform is as follows:

[0052] ;

[0053] in, Represents the first frequency components, Indicates the time domain sampling points, is the total number of sampling points, is an imaginary unit. Through fast Fourier transform, the amplitude and phase of each frequency component are obtained to form the spectrum characteristics of the signal. The time domain characteristics of the vibration signal, the time domain characteristics of the acoustic emission signal, and the time domain characteristics of the cutting force signal are combined with the frequency domain feature set to form a complete feature set to obtain the first multi-source sensor data.

[0054] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0055] Performing autocorrelation analysis and cross-correlation analysis on the first multi-source sensor data to obtain time correlation characteristics of each sensor signal and correlation characteristics between different sensor signals;

[0056] Based on the time correlation feature and the correlation feature, dynamically time warping the first multi-source sensor data to obtain an aligned multi-source sensor data sequence;

[0057] The aligned multi-source sensor data sequence is segmented using a sliding window to obtain multiple overlapping data segments, and each overlapping data segment is subjected to singular value decomposition to extract the main eigenvectors and corresponding singular values ​​to obtain the state representation after dimensionality reduction.

[0058] Based on the state representation after dimensionality reduction, the initial state transfer matrix is ​​constructed, and the initial observation matrix is ​​estimated using the least squares method to obtain the initial state space model;

[0059] The initial state space model and the aligned multi-source sensor data sequence are input into the unscented Kalman filter to perform iterative parameter optimization to obtain the optimized state transfer matrix and observation matrix;

[0060] Based on the optimized state transfer matrix and observation matrix, nonlinear state equations and observation equations are constructed, process noise and observation noise are introduced, and a state space model representing the relationship between multi-source data at adjacent moments is obtained.

[0061] Specifically, autocorrelation analysis and cross-correlation analysis are performed on the first multi-source sensor data to extract the time correlation of sensor signals and the correlation characteristics between different sensor signals. Autocorrelation analysis is used to measure the similarity of the same signal under different time delays, reflecting the periodicity and stability of the signal. For a given signal , and its autocorrelation function is calculated by the following formula:

[0062] ;

[0063] in, Indicates signal Delay in time The autocorrelation value at is the number of sampling points, is the time delay. Autocorrelation function The larger the value, the stronger the similarity of the signals under this delay. Through autocorrelation analysis, the periodic characteristics in vibration, acoustic emission and cutting force signals are identified to determine whether the operating status of the system is normal. At the same time, cross-correlation analysis is performed to measure the relationship between different sensor signals, such as the similarity between vibration signals and cutting force signals. The cross-correlation function is expressed by the following formula:

[0064] ;

[0065] in, Indicates signal and signal Delay in time The cross-correlation value at and Represent signals from different sensors respectively. Cross-correlation analysis reveals the coupling relationship between different signals and helps identify the mutual influence between multi-source signals. Dynamic time warping is performed on the first multi-source sensor data based on the time correlation feature and the correlation feature to align the multi-source sensor data sequence. Dynamic time warping is a technology used to align time series data. The optimal matching path between two sequences is found through a dynamic programming algorithm to minimize the difference between the time series. The aligned multi-source sensor data sequence is segmented by a sliding window to obtain multiple overlapping data segments. The sliding window technology is used to divide a long data series into several smaller subsequences, and the length of each window is fixed. , there is a certain overlap between the windows, and the sliding window is used to capture the dynamic change characteristics in the data. Singular value decomposition is performed on each segmented data segment to extract the main eigenvectors and corresponding singular values. Singular value decomposition is a technique for matrix decomposition that can decompose a matrix into a set of products of singular values ​​and eigenvectors to extract the main features in the data. For a data matrix , obtained through singular value decomposition:

[0066] ;

[0067] in, is the original data matrix, and are the left singular vector matrix and the right singular vector matrix respectively, is a diagonal matrix whose diagonal elements are singular values. Through singular value decomposition, the main eigenvectors and corresponding singular values ​​are extracted to achieve data dimensionality reduction and obtain the state representation after dimensionality reduction. Based on the state representation after dimensionality reduction, the initial state transfer matrix is ​​constructed, and the initial observation matrix is ​​estimated using the least squares method to obtain the initial state space model. The state space model is a mathematical model used to describe dynamic systems. The state change and observation relationship of the system are described by the state transfer matrix and the observation matrix. Assume that the state vector is , the observation vector is , then the state space model is expressed as:

[0068] ;

[0069] ;

[0070] in, Indicates that the system is at time The state vector at time, Indicates at time The observation vector at time, is the state transfer matrix, is the observation matrix, and are process noise and observation noise, respectively, and they are usually assumed to be Gaussian noise with zero mean. Through the least squares method, the initial state transfer matrix and observation matrix are estimated under the premise of minimizing the observation error, thereby establishing the initial state space model. The initial state space model and the aligned multi-source sensor data sequence are input into the unscented Kalman filter for iterative parameter optimization. The unscented Kalman filter is a tool for state estimation of nonlinear systems. By introducing a set of unscented sampling points to approximate the state distribution, it can introduce nonlinear effects in the state transfer and observation process, and update the state transfer matrix and observation matrix through iterative optimization to improve the accuracy of the model. Based on the optimized state transfer matrix and observation matrix, nonlinear state equations and observation equations are constructed, process noise and observation noise are introduced, and a state space model representing the relationship between multi-source data at adjacent moments is obtained. This model can effectively describe the dynamic changes of each sensor signal and their mutual relationship in the multi-axis linkage CNC machining process, and provides a mathematical basis for subsequent state evaluation, anomaly detection and control optimization. For example, during the machining process, if the system state changes suddenly (such as severe tool wear or abnormal workpiece vibration), the state space model can capture this change through sensor data and make corresponding estimates through state transitions and observation equations, thereby identifying abnormal situations in a timely manner.

[0071] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0072] Inputting the first multi-source sensor data into the state space model to perform state estimation to obtain a state estimation sequence reflecting the dynamic characteristics of the machining process;

[0073] Perform multi-scale decomposition on the state estimation sequence, construct a hierarchical feature structure, and obtain a multi-level feature set containing information at different time scales;

[0074] Based on the multi-level feature set, an adaptive feature selection strategy is created. According to the feature's discriminative ability and information gain, the feature weight is dynamically adjusted to obtain a preliminary screened feature subset.

[0075] Construct a feature correlation network, perform correlation analysis on the features in the initially screened feature subset, identify redundant features and complementary features, and obtain an optimized feature combination;

[0076] Based on the feature enhancement mechanism of machining process knowledge, the optimized feature combination is enhanced to obtain an enhanced feature set, and a sparse constraint regularization term is introduced to perform sparse optimization on the enhanced feature set, remove irrelevant and weakly related features, and obtain a sparse feature representation;

[0077] Based on the feature importance evaluation mechanism and combined with the processing quality index and condition monitoring objectives, the features in the sparse feature representation are ranked by importance to obtain the feature importance sequence;

[0078] Based on the feature importance sequence, the top N features whose cumulative importance reaches the preset threshold are selected, and combined with dimensionality reduction technology to obtain a low-dimensional feature subset with sparsity constraints.

[0079] Specifically, the first multi-source sensor data is input into the state space model, and the dynamic characteristics of the machining process are captured through the state estimation process to obtain a state estimation sequence. The state space model is a mathematical tool that describes the dynamic behavior of the system. The input data is calculated through the state transfer and observation matrix of the model to obtain the state vector at each time point, thereby forming a state estimation sequence that reflects the changes in the machining process. For each time point in the state estimation sequence , the state vector It is represented by the following state equation:

[0080] ;

[0081] in, Indicates at time The state vector at time, is the state transfer matrix, which is used to describe the change relationship of the system state. represents process noise, which is a noise with zero mean and covariance of Gaussian noise term. In order to extract richer information from the state estimation sequence, the state estimation sequence is decomposed at multiple scales, and a hierarchical feature structure is constructed to obtain a multi-level feature set containing information at different time scales. Multi-scale decomposition is to decompose the signal at different time scales, extract the characteristics of the signal at different scales, and capture the short-term mutations and long-term trends in the processing process. For example, the state estimation sequence is decomposed at multiple scales using the wavelet decomposition method. Wavelet decomposition can decompose the signal into low-frequency and high-frequency parts, while obtaining global and local information. The decomposed multi-level features fully reflect the operating status of the system. An adaptive feature selection strategy is created based on the multi-level feature set. The feature weights are dynamically adjusted according to the feature's discriminative ability and information gain to obtain a preliminarily screened feature subset. The feature's discriminative ability is measured by the ratio of the between-class variance to the within-class variance, and the information gain is used to measure the amount of information provided by the feature for state classification. For a certain feature , its information gain is expressed as:

[0082] ;

[0083] in, Representation characteristics The information gain of Represents the target variable The entropy of describes the uncertainty of the system. Indicates that given a feature The conditional entropy of the target variable under certain conditions. The larger the information gain, the better the feature. The greater the contribution to reducing the uncertainty of the system. Based on the comprehensive evaluation of discriminative ability and information gain, the adaptive feature selection strategy can assign weights to each feature and perform preliminary screening according to the weight size to ensure that the most valuable features for condition monitoring are retained. Construct a feature correlation network and perform correlation analysis on the features in the preliminary screened feature subset to identify redundant features and complementary features. The feature correlation network is an analysis tool based on a graph structure, in which each node represents a feature and the edge represents the strength of the correlation between features. By calculating the Pearson correlation coefficient between features, feature pairs with strong correlation are identified:

[0084] ;

[0085] in, Representation characteristics and The correlation between is the sample size, and Characteristics and The mean of , when the correlation coefficient is high, it means that the two features are redundant, and one of them should be discarded appropriately; when the correlation coefficient is low, the features are complementary, and retaining them can provide more information. Through the analysis of the feature correlation network, the optimized feature combination is obtained, thereby removing redundant features and retaining features that are complementary to state monitoring. Based on the optimized feature combination, the feature enhancement mechanism is introduced in combination with the processing technology knowledge to enhance the features and obtain an enhanced feature set. Processing technology knowledge includes the understanding of the physical characteristics and process parameters of the processing process. By incorporating this prior knowledge into the features, the representativeness and discriminability of the features for the processing state are improved. The enhanced feature set then introduces a sparse constraint regularization term for sparse optimization to eliminate irrelevant and weakly correlated features. The goal of sparse optimization is to reduce the number of features and retain the most important features by imposing sparsity constraints on the features, thereby obtaining a sparse feature representation. Sparse optimization is achieved by introducing L1 regularization, and the optimization objective function is as follows:

[0086] ;

[0087] in, is the loss function, Representation sample The true value of represents the predicted value of the model, is the sample size, is the regularization parameter, controlling the degree of sparsity, For the The weight of each feature. Through sparse optimization, the number of features is effectively reduced, thereby simplifying the model and improving computational efficiency. After obtaining the sparse feature representation, the features in the sparse feature representation are ranked by importance based on the feature importance evaluation mechanism. The feature importance evaluation mechanism combines the processing quality indicators and the status monitoring objectives, and judges the importance of the feature by analyzing the contribution of each feature to the monitoring model. The importance of a feature is measured by the degree of influence of the feature on the improvement of model performance. The more important the feature, the more important it is for accurately classifying the processing status and evaluating the processing quality. According to the importance of the feature, the feature importance sequence is obtained, and based on the feature importance sequence, the top features whose cumulative importance reaches the preset threshold are selected. features, thus ensuring that the retained features contribute a certain proportion to the total information of the system. Combined with dimensionality reduction techniques, the selected features are reduced in dimensionality to obtain a low-dimensional feature subset with sparsity constraints. Dimensionality reduction techniques, such as principal component analysis, transform the high-dimensional feature space into a low-dimensional space while retaining most of the information.

[0088] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0089] Normalize the low-dimensional feature subset to obtain a normalized feature vector;

[0090] Input the normalized feature vector into a pre-trained autoencoder network to obtain a reconstructed feature vector, where the autoencoder network includes an encoder and a decoder, the encoder maps the input feature to a latent space, and the decoder reconstructs the latent space representation into the original feature space;

[0091] Calculate the mean square error between the normalized feature vector and the reconstructed feature vector to obtain the reconstruction error, which is used as the first indicator for abnormal state detection;

[0092] The low-dimensional feature subsets of adjacent sampling periods are processed by sliding windows, with a window size of N and a sliding step of 1, to obtain a feature sequence matrix, and the Euclidean distance between two adjacent feature vectors in the feature sequence matrix is ​​calculated as the second indicator for abnormal state detection;

[0093] Based on the multi-index fusion rule of fuzzy logic, the reconstruction error and Euclidean distance are used as input, the fused abnormality index is obtained through fuzzy reasoning, and a dynamic threshold adjustment mechanism is constructed to generate a real-time updated detection threshold;

[0094] The fused abnormality index is compared with the detection threshold updated in real time. If it exceeds the threshold, it is judged as an abnormal state, otherwise it is judged as a normal state, thereby obtaining the abnormal state detection result.

[0095] Specifically, the low-dimensional feature subset is normalized to eliminate the differences between different feature dimensions so that all feature values ​​are in the same range, thereby avoiding excessive influence of certain features on the results due to their large numerical range during the calculation process. Through normalization, a normalized feature vector is obtained. The normalized feature vector is input into a pre-trained autoencoder network. The autoencoder network consists of two parts: an encoder and a decoder. The encoder maps the input feature vector to a latent space, while the decoder reconstructs the representation in the latent space into the original feature space. The structures of the encoder and decoder are symmetrical. The encoder gradually compresses the high-dimensional input features into a low-dimensional latent space through a series of linear transformations and nonlinear activation functions, while the decoder maps the low-dimensional latent representation back to the original space through a reverse process. Suppose the input feature vector is , the encoder maps it to a latent space representation z:

[0096] ;

[0097] in, is the latent space representation, and are the weight matrix and bias vector of the encoder respectively, represents the activation function. The decoder converts the potential representation Reconstruct the feature vector of the original feature space

[0098] ;

[0099] in, is the reconstructed feature vector, and are the weight matrix and bias vector of the decoder respectively. Through the reconstruction process of the autoencoder network, the main feature patterns in the input features can be learned and reconstructed back to the original space. Calculate the normalized feature vector and reconstructed eigenvector The mean square error is used to measure the deviation between the reconstruction result and the original input. Its formula is:

[0100] ;

[0101] Among them, MSE is the mean square error, and denote the original eigenvector and the reconstructed eigenvector respectively. Quantity, is the dimension of the feature vector. The larger the mean square error, the greater the deviation between the input feature and the reconstructed feature, which means that the input data contains anomalies. Therefore, the mean square error is used as the first indicator for abnormal state detection. In order to identify the dynamic changes of the system, the low-dimensional feature subsets of adjacent sampling periods are processed by sliding windows, and the window size is set to , the sliding step is 1, and the feature sequence matrix is ​​obtained. The sliding window technology tracks the time series changes of the feature data to ensure continuity in time and fine-grained state monitoring. In the feature sequence matrix obtained by sliding window segmentation, each row represents a feature vector in a window. The Euclidean distance between two adjacent feature vectors in the feature sequence matrix is ​​calculated to measure the magnitude of the feature change in time. For two feature vectors and , the Euclidean distance is calculated by the following formula:

[0102] ;

[0103] in, Indicates at time and The Euclidean distance of the feature vectors between time points, and The eigenvectors are The value of each component. The Euclidean distance can reflect the change amplitude of the feature vector in time. If the distance is large, it means that the system state has changed significantly during this period of time. Therefore, the Euclidean distance is used as the second indicator for abnormal state detection. In order to comprehensively consider the two indicators of mean square error and Euclidean distance, the multi-indicator fusion rule based on fuzzy logic takes these two indicators as input, and obtains the fused abnormality index through fuzzy reasoning. Fuzzy logic can handle the uncertainty and ambiguity of the system, map the numerical values ​​of mean square error and Euclidean distance to fuzzy sets, and obtain a comprehensive abnormality index through reasoning through pre-set fuzzy rules. In order to improve the real-time and adaptability of anomaly detection, a dynamic threshold adjustment mechanism is constructed to generate real-time updated detection thresholds. The dynamic threshold adjustment mechanism dynamically adjusts the threshold according to the current system operation status and the statistical characteristics of historical data, so as to adapt to the feature changes under different working conditions and avoid misjudgment and missed judgment caused by fixed thresholds. Through dynamic adjustment, the detection standard is automatically updated according to the current processing conditions to ensure the accuracy and timeliness of the anomaly detection results. The fused abnormality index is compared with the dynamically updated detection threshold. If the abnormality index exceeds the threshold, the current state is judged to be an abnormal state, indicating that there are abnormal phenomena such as equipment failure, tool wear, and loose clamping in the processing process; if the abnormality index does not exceed the threshold, it is judged to be a normal state, and the final abnormal state detection result is obtained. For example, in the CNC processing process, if the tool wear increases, the reconstruction error of the vibration signal will gradually increase, and the time variation amplitude of the feature vector will also increase significantly. These changes will cause the fused abnormality index to exceed the normal threshold, so that the system judges it as an abnormal state and prompts the operator to check and handle it.

[0104] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0105] The abnormal state detection results are divided into a normal state sample set and an abnormal state sample set, and data enhancement processing is performed on the normal state sample set and the abnormal state sample set to obtain an expanded training data set;

[0106] Construct a backbone network. The backbone network adopts the DenseNet structure, which contains 4 dense blocks. Each dense block contains 6 convolutional layers. Each convolutional layer is connected by dense connections. The convolution kernel size is 3×3, the step size is 1, and the padding is 1. Add a transition layer between each dense block. The transition layer contains a 1×1 convolutional layer and a 2×2 average pooling layer to reduce the spatial resolution and number of channels of the feature map.

[0107] Construct multiple branch networks, each corresponding to a sensor signal type, using the ResNet structure, including three residual blocks, each residual block contains two 3×3 convolutional layers and a short-circuit connection; add an attention mechanism module between the backbone network and multiple branch networks, including two sub-modules: channel attention and spatial attention, for adaptively adjusting feature weights;

[0108] Input a low-dimensional feature subset in the expanded training data set into the backbone network to obtain a first backbone feature map, and input the corresponding sensor signal features into the corresponding branch network to obtain a first branch feature map, and perform feature fusion on the first backbone feature map and the first branch feature map through the attention mechanism module to obtain a first fused feature map;

[0109] Construct a multi-task learning head, including state classification task and parameter regression task. The state classification task uses a fully connected layer and a Softmax layer, and the parameter regression task uses a fully connected layer and a linear activation function. Input the first fused feature map into the multi-task learning head to obtain the state classification result and parameter regression result.

[0110] A weighted loss function is used for model training. The weighted loss function includes the cross entropy loss of state classification and the mean square error loss of parameter regression. The network parameters are updated through the back propagation algorithm and the Adam optimizer to obtain the state evaluation model.

[0111] Specifically, the abnormal state detection results are divided into a normal state sample set and an abnormal state sample set. The normal state sample set contains all normal operation records in the processing process, while the abnormal state sample set contains various abnormal conditions, such as severe tool wear and excessive workpiece vibration. Data enhancement processing is performed on the normal and abnormal state sample sets. The training data set is expanded by increasing the diversity of samples to improve the generalization ability of the model. For example, by adding random noise, data flipping, scale transformation and other methods to simulate different processing environments, the expanded training data set is richer and the model can cope with more actual situations. The backbone network is constructed. The backbone network adopts the DenseNet structure. DenseNet is a convolutional neural network that realizes feature reuse and efficient gradient transfer through dense connections. When constructing the backbone network, it contains 4 dense blocks, each of which contains 6 convolutional layers. Each convolutional layer is connected by dense connections, and each convolutional layer receives the output of all previous layers as input. This structure can effectively alleviate the gradient vanishing problem and improve the efficiency of feature transfer. The convolution kernel size is 3×3, the stride is 1, and the padding is 1, ensuring that the convolution operation can maintain the spatial size of the feature map. A transition layer is added between each dense block, which contains a 1×1 convolution layer and a 2×2 average pooling layer. The 1×1 convolution layer is used to reduce the number of channels, while the 2×2 average pooling layer is used to reduce the spatial resolution of the feature map, reducing the amount of calculation and the complexity of the network. At the same time, in order to effectively utilize the characteristics of different sensor signals, multiple branch networks are constructed, each branch network corresponds to a sensor signal type, such as vibration, acoustic emission, and cutting force signals. The branch network adopts the ResNet structure, which is a convolutional neural network that solves the gradient vanishing problem of deep networks through residual connections. In each branch network, there are 3 residual blocks, each of which contains two 3×3 convolutional layers and a short-circuit connection. The short-circuit connection allows the input to be directly passed to the output layer, avoiding the loss of gradient information due to too many layers, making the deep network easier to train. In order to improve the effect of feature fusion, an attention mechanism module is added between the backbone network and multiple branch networks. The attention mechanism module includes two submodules: channel attention and spatial attention. Channel attention is used to adaptively adjust the weights of different feature channels, so that the network can pay more attention to the feature channels that contribute more to state monitoring; while spatial attention is used to adjust the weights of feature maps in the spatial dimension, so that the network can focus on the key areas in the feature map. Through the introduction of the attention mechanism, the network can automatically give more weight to important features, improve the expressiveness of features and the accuracy of state monitoring. The low-dimensional feature subset in the expanded training data set is input into the backbone network, and after processing by DenseNet, the first backbone feature map is obtained. At the same time, the corresponding sensor signal features are input into the corresponding branch network, and the first branch feature map is obtained through processing by ResNet.The first trunk feature map and the first branch feature map are fused through the attention mechanism module to obtain the first fused feature map. In order to realize multi-task learning of state classification and parameter regression, a multi-task learning head is constructed, which includes state classification task and parameter regression task. The state classification task uses a fully connected layer and a Softmax layer to classify the processing state and obtain the probability distribution of different states. Assuming there are C state categories, the Softmax output is:

[0112] ;

[0113] in, Represents the input feature vector Classified into categories The probability of For and category The Softmax layer can convert the network output into a probability distribution to achieve state classification. The parameter regression task uses a fully connected layer and a linear activation function to estimate key parameters related to the state during the machining process, such as cutting force, vibration amplitude, etc. Its output is expressed as:

[0114] ;

[0115] in, is the regression output, and are weight vector and bias respectively, is the input feature. Through the linear activation function, the model can estimate the continuous parameters in the processing process. During the model training process, in order to balance the losses of the two tasks of state classification and parameter regression, a weighted loss function is used. The weighted loss function contains the cross entropy loss of state classification and the mean square error loss of parameter regression. The cross entropy loss is used to measure the difference between the classification result and the true label, and its formula is:

[0116] ;

[0117] in, is the cross entropy loss, is the sample size, Indicates samples belong to the category The real label, For the samples belong to the category The predicted probability. The mean square error loss is used to measure the difference between the regression result and the true value, and its formula is:

[0118] ;

[0119] in, is the mean square error loss, For the The true value of the samples, For the The predicted value of samples. The final weighted loss function is:

[0120] ;

[0121] in, and is the weight of the two tasks, used to balance the loss contribution of classification and regression. The network parameters are updated through the back propagation algorithm and the Adam optimizer. The Adam optimizer is an adaptive learning rate optimization method. Combining the advantages of momentum and RMSProp, it can dynamically adjust the learning rate according to the first-order and second-order moment estimates of the gradient, thereby accelerating the convergence process of the model. Through the above model construction and training process, a state evaluation model is finally obtained, which can simultaneously classify the state of the multi-axis linkage CNC machining process and estimate the key parameters related to the machining state.

[0122] In a specific embodiment, the process of executing step 106 may specifically include the following steps:

[0123] Performing real-time acquisition and preprocessing of the second multi-source sensor data acquired in real time to obtain preprocessed real-time sensor data, and inputting the preprocessed real-time sensor data into a state space model to perform state estimation to obtain a real-time state sequence;

[0124] Perform feature extraction and compressed sensing processing on the real-time state sequence to obtain a real-time feature subset, and input the real-time feature subset into the backbone network of the state assessment model to obtain a second backbone feature map;

[0125] Input various sensor signals in the preprocessed real-time sensor data into the corresponding branch networks respectively to obtain multiple second branch feature maps, and perform feature fusion on the second trunk feature map and the multiple second branch feature maps through the attention mechanism module to obtain a second fused feature map;

[0126] The second fused feature map is input into the multi-task learning head, the processing state classification result is obtained through the state classification task, and the parameter estimation value is obtained through the parameter regression task;

[0127] Build a digital twin model, including geometric model, physical model and behavioral model. The geometric model represents the three-dimensional structure of the machine tool and workpiece, the physical model describes the mechanical and thermodynamic characteristics during the processing, and the behavioral model simulates the movement and control logic of each component of the machine tool;

[0128] The processing status classification results and parameter estimation values ​​are mapped to the digital twin model, the status and parameters of the digital twin model are updated, and based on the updated digital twin model, multi-machine tool collaborative status evaluation and predictive simulation are performed to obtain distributed collaborative monitoring results. The distributed collaborative monitoring results include the current status of each machine tool, predicted trends and collaborative optimization suggestions.

[0129] Specifically, the second multi-source sensor data collected in real time is collected and preprocessed. The sensor data includes data from sensors such as vibration, acoustic emission, and cutting force. After preprocessing steps such as denoising, normalization, and filtering, preprocessed real-time sensor data is obtained. The preprocessed real-time sensor data is input into the state space model, and the state of the machining process is estimated through the state space model to obtain a real-time state sequence. The state estimation process of the state space model is described by the following formula:

[0130] ;

[0131] ;

[0132] in, Indicates that the system is at time The state vector at time, is the observation vector, is the state transfer matrix, which describes the change relationship between the system state at adjacent moments. is the observation matrix, which is used to map the system state to the observation space. and They are process noise and observation noise respectively, and are usually assumed to have zero mean and variance and Gaussian noise. The real-time state sequence is subjected to feature extraction and compressed sensing processing to obtain a real-time feature subset. Feature extraction extracts representative features in the state sequence, such as root mean square value, peak value, etc., which are used to describe the operating state of the system, while compressed sensing is a data compression technology that can reduce data dimensions while retaining key features, thereby improving computational efficiency. The real-time feature subset is input into the backbone network of the state assessment model, and the second backbone feature map is obtained through calculation of the backbone network. The backbone network adopts the DenseNet structure, and dense connections are used to achieve efficient transmission and reuse of features, so that the obtained feature map has richer representation capabilities. The various types of sensor signals in the preprocessed real-time sensor data are respectively input into the corresponding branch networks to obtain multiple second branch feature maps. The branch network adopts the ResNet structure, and the gradient disappearance is avoided by short-circuiting the residual block, so that the deep network can be trained more effectively. The second trunk feature map and multiple second branch feature maps are feature fused through the attention mechanism module. The attention mechanism module contains two submodules: channel attention and spatial attention. Channel attention is used to adaptively adjust the weights of different feature channels so that important feature channels receive more attention, while spatial attention is used to adjust the weights of feature maps in the spatial dimension so that the model can pay more attention to the key areas in the feature map. This fusion method can adaptively adjust the importance of various features to obtain a more representative second fused feature map. The second fused feature map is input into the multi-task learning head, and the processing state classification results and parameter estimates are obtained through the state classification task and parameter regression task. The state classification task uses a fully connected layer and a Softmax layer. The probability that the processing state belongs to different categories is obtained through the Softmax layer. The formula is as follows:

[0133] ;

[0134] in, Represents the input feature vector Classified into categories The probability of For and category The associated weight vector is the total number of state categories. The parameter regression task estimates the key parameters of the machining process, such as cutting force, vibration amplitude, etc., through the fully connected layer and linear activation function. Its output is expressed as:

[0135] ;

[0136] in, is the regression output, and are the weight vector and bias of the regression model respectively. Through these two tasks, the model can simultaneously classify the state of the machining process and estimate the relevant parameters. After obtaining the machining state classification results and parameter estimation values, a digital twin model is constructed for digital mapping and real-time simulation of the physical machining system. The digital twin model includes three sub-models: geometric model, physical model and behavioral model. The geometric model is used to represent the three-dimensional structure of the machine tool and the workpiece, reflecting the appearance and spatial layout of the machining object; the physical model is used to describe the mechanical and thermodynamic characteristics of the machining process, such as the distribution of cutting force, the stress and strain of the workpiece, etc., so as to simulate the physical phenomena generated during the machining process; the behavioral model is used to simulate the movement and control logic of each component of the machine tool, such as the path planning of the tool, the linkage control of the machine tool axis, etc. By constructing these three sub-models, the digital twin model can perform a comprehensive virtual representation of the machining process. The machining state classification results and parameter estimation values ​​are mapped to the digital twin model, and the state and parameters of the digital twin model are updated. The updated digital twin model can reflect the operating status of the physical system in real time and simulate the machining process in the virtual space. For example, when the system detects the state of tool wear, the digital twin model updates the geometry of the tool according to the estimated degree of wear, simulates its impact, and predicts processing errors and quality problems. Based on the digital twin model, collaborative state evaluation and predictive simulation of multiple machine tools are carried out to obtain distributed collaborative monitoring results. By integrating the digital twin models of multiple machine tools, collaborative state monitoring and optimization of the entire processing workshop are achieved. During the collaborative evaluation process, the digital twin model analyzes the mutual influence between multiple machine tools, such as evaluating the load balance between machine tools to determine how to schedule the tasks of each machine tool to improve processing efficiency and reduce energy consumption. Predictive simulation is used to infer the state trend of each machine tool in the future, such as predicting the service life of the tool and the processing accuracy of the workpiece. These predictive information helps operators to perform maintenance and adjustments in advance to avoid sudden failures and quality problems in the production process. For example, in an actual multi-axis processing workshop, multiple machine tools perform different processing tasks at the same time. If the tool of a machine tool is severely worn, the state evaluation model will detect the anomaly and map it to the digital twin model to update the wear state and geometry of the tool. Then, through distributed collaborative monitoring, the digital twin system will analyze the impact of the machine tool status on other machine tool processing tasks. For example, if the processing efficiency decreases due to tool wear, the tasks of other machine tools need to be reallocated to ensure production progress. Through predictive simulation, the remaining service life of the current tool is estimated, and replacement suggestions are given to avoid tool failure during processing and affect workpiece quality.

[0137] The above describes the state monitoring method of multi-axis linkage CNC machining in the embodiment of the present invention. The following describes the state monitoring system of multi-axis linkage CNC machining in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a state monitoring system for multi-axis linkage CNC machining includes:

[0138] The acquisition module 201 is used to collect and pre-process the vibration, acoustic emission and cutting force signals in the multi-axis linkage CNC machining process to obtain the first multi-source sensor data;

[0139] The processing module 202 is used to input the first multi-source sensor data into the state space model builder for processing to obtain a state space model;

[0140] A compression module 203 is used to perform feature extraction and compressive sensing on the first multi-source sensor data according to the state space model to obtain a low-dimensional feature subset;

[0141] The judgment module 204 is used to calculate the reconstruction error and Euclidean distance of the low-dimensional feature subset, and perform fusion judgment to obtain an abnormal state detection result;

[0142] A training module 205 is used to train a deep learning model including a trunk network and multiple branch networks according to the abnormal state detection result and the low-dimensional feature subset to obtain a state assessment model;

[0143] The evaluation module 206 is used to input the second multi-source sensor data collected in real time into the state evaluation model for state evaluation, obtain the processing state classification results and parameter estimation values, and map the processing state classification results and parameter estimation values ​​to the digital twin model to obtain distributed collaborative monitoring results.

[0144] Through the collaboration of the above components, the fusion of multi-source sensor data and the construction of state space model, the ability to capture the dynamic characteristics of the machining process is improved, making the state monitoring more comprehensive and accurate. By adopting feature extraction and compressed sensing processing technology, a low-dimensional feature subset with sparsity constraints is obtained, which effectively reduces the data dimension and improves the computational efficiency. The abnormality detection method combining the autoencoder network and Euclidean distance calculation enhances the recognition ability of abnormal states and reduces false positives and false negatives. By using a deep learning model containing a backbone network and multiple branch networks, adaptive processing of different types of sensor signals is achieved, improving the accuracy and generalization ability of state assessment. The introduction of attention mechanism and multi-task learning enhances the model's perception of key features, and realizes the collaborative optimization of state classification and parameter estimation. Through the construction and update of the digital twin model, virtual-real fusion state monitoring is realized, which provides a basis for the visualization and predictive analysis of the machining process. Based on the distributed collaborative monitoring results, the collaborative state assessment and predictive simulation of multiple machine tools are realized, providing support for the overall optimization and intelligent decision-making of the production line.

[0145] above Figure 2 The state monitoring system of multi-axis linkage CNC machining in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The state monitoring device of multi-axis linkage CNC machining in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0146] Figure 3 3 is a schematic diagram of the structure of a state monitoring device for multi-axis linkage numerical control machining provided by an embodiment of the present invention. The state monitoring device 300 for multi-axis linkage numerical control machining may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 may be temporary storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the state monitoring device 300 for multi-axis linkage numerical control machining. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the state monitoring device 300 for multi-axis linkage numerical control machining to implement the steps of the state monitoring method for multi-axis linkage numerical control machining.

[0147] The multi-axis linkage CNC machining state monitoring device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will appreciate that Figure 3 The structure of the state monitoring device for multi-axis CNC machining shown does not constitute a limitation on the state monitoring device for multi-axis CNC machining provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0148] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the state monitoring method for multi-axis linkage CNC machining.

[0149] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0150] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0151] As described above, 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 aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A state monitoring method for multi-axis linkage CNC machining, characterized in that: The method comprises: The vibration, acoustic emission and cutting force signals in the multi-axis linkage CNC machining process are collected and preprocessed to obtain the first multi-source sensor data; Input the first multi-source sensor data into a state space model builder for processing to obtain a state space model; specifically including: performing autocorrelation analysis and cross-correlation analysis on the first multi-source sensor data to obtain the time correlation characteristics of each sensor signal and the correlation characteristics between different sensor signals; based on the time correlation characteristics and the correlation characteristics, dynamically time warping the first multi-source sensor data to obtain an aligned multi-source sensor data sequence; performing sliding window segmentation on the aligned multi-source sensor data sequence to obtain multiple overlapping data segments, and performing singular value decomposition on each overlapping data segment to extract the main eigenvectors and corresponding singular values ​​to obtain a reduced-dimensional state representation; based on the reduced-dimensional state representation, constructing an initial state transfer matrix, and using the least squares method to estimate the initial observation matrix to obtain an initial state space model; inputting the initial state space model and the aligned multi-source sensor data sequence into an unscented Kalman filter, performing iterative parameter optimization, and obtaining an optimized state transfer matrix and observation matrix; based on the optimized state transfer matrix and observation matrix, constructing a nonlinear state equation and an observation equation, introducing process noise and observation noise, and obtaining a state space model representing the relationship between multi-source data at adjacent moments; According to the state space model, feature extraction and compressed sensing are performed on the first multi-source sensor data to obtain a low-dimensional feature subset; Calculating the reconstruction error and Euclidean distance of the low-dimensional feature subset, and performing fusion judgment to obtain an abnormal state detection result; A deep learning model including a backbone network and a plurality of branch networks is trained according to the abnormal state detection result and the low-dimensional feature subset to obtain a state assessment model; The second multi-source sensor data collected in real time is input into the state assessment model for state assessment to obtain processing state classification results and parameter estimation values, and the processing state classification results and parameter estimation values ​​are mapped to the digital twin model to obtain distributed collaborative monitoring results.

2. The state monitoring method for multi-axis linkage CNC machining according to claim 1 is characterized in that: The vibration, acoustic emission and cutting force signals in the multi-axis linkage CNC machining process are collected and preprocessed to obtain the first multi-source sensor data, including: The vibration signal in the multi-axis linkage CNC machining process is collected by a vibration sensor, and band-pass filtering is performed to obtain a filtered vibration signal; the acoustic emission signal in the multi-axis linkage CNC machining process is collected by an acoustic emission sensor, and band-pass filtering is performed to obtain a filtered acoustic emission signal; the cutting force signal in the multi-axis linkage CNC machining process is collected by a cutting force sensor, and band-pass filtering is performed to obtain a filtered cutting force signal; Performing a time domain analysis on the filtered vibration signal to obtain a time domain feature of the vibration signal; performing a time domain analysis on the filtered acoustic emission signal to obtain a time domain feature of the acoustic emission signal; performing a time domain analysis on the filtered cutting force signal to obtain a time domain feature of the cutting force signal; Performing fast Fourier transform on the filtered vibration signal, the filtered acoustic emission signal and the filtered cutting force signal respectively to obtain a frequency domain feature set; The vibration signal time domain features, the acoustic emission signal time domain features, the cutting force signal time domain features and the frequency domain feature set are combined to obtain the first multi-source sensor data.

3. The state monitoring method for multi-axis linkage CNC machining according to claim 1 is characterized in that: The step of performing feature extraction and compressed sensing on the first multi-source sensor data according to the state space model to obtain a low-dimensional feature subset includes: Inputting the first multi-source sensor data into the state space model to perform state estimation to obtain a state estimation sequence reflecting the dynamic characteristics of the machining process; Performing multi-scale decomposition on the state estimation sequence, constructing a hierarchical feature structure, and obtaining a multi-level feature set containing information of different time scales; Based on the multi-level feature set, an adaptive feature selection strategy is created, and feature weights are dynamically adjusted according to the discriminative ability and information gain of the features to obtain a preliminarily screened feature subset; Constructing a feature correlation network, performing correlation analysis on the features in the initially screened feature subset, identifying redundant features and complementary features, and obtaining an optimized feature combination; Based on the feature enhancement mechanism of processing technology knowledge, the optimized feature combination is enhanced to obtain an enhanced feature set, and a sparse constraint regularization term is introduced to perform sparse optimization on the enhanced feature set, and irrelevant and weakly related features are eliminated to obtain a sparse feature representation; Based on the feature importance evaluation mechanism and in combination with the processing quality index and the state monitoring target, the features in the sparse feature representation are ranked by importance to obtain a feature importance sequence; Based on the feature importance sequence, the first N features whose cumulative importance reaches a preset threshold are selected, and combined with dimensionality reduction technology to obtain a low-dimensional feature subset with sparsity constraints.

4. The state monitoring method for multi-axis linkage CNC machining according to claim 1 is characterized in that: The calculation of the reconstruction error and the Euclidean distance of the low-dimensional feature subset and the fusion judgment to obtain the abnormal state detection result includes: Normalizing the low-dimensional feature subset to obtain a normalized feature vector; Inputting the normalized feature vector into a pre-trained autoencoder network to obtain a reconstructed feature vector, wherein the autoencoder network includes an encoder and a decoder, the encoder maps the input feature to a latent space, and the decoder reconstructs the latent space representation into the original feature space; Calculating a mean square error between the normalized feature vector and the reconstructed feature vector to obtain a reconstruction error as a first indicator for abnormal state detection; Perform sliding window processing on the low-dimensional feature subsets of adjacent sampling periods, with a window size of N and a sliding step of 1, to obtain a feature sequence matrix, and calculate the Euclidean distance between two adjacent feature vectors in the feature sequence matrix as a second indicator for abnormal state detection; Based on the multi-index fusion rule of fuzzy logic, the reconstruction error and the Euclidean distance are used as inputs, the fused abnormality index is obtained through fuzzy reasoning, and a dynamic threshold adjustment mechanism is constructed to generate a real-time updated detection threshold; The fused abnormality index is compared with the real-time updated detection threshold. If it exceeds the threshold, it is determined to be an abnormal state, otherwise it is determined to be a normal state, thereby obtaining an abnormal state detection result.

5. The state monitoring method for multi-axis linkage CNC machining according to claim 1 is characterized in that: The deep learning model including a trunk network and a plurality of branch networks is trained according to the abnormal state detection result and the low-dimensional feature subset to obtain a state assessment model, including: Dividing the abnormal state detection results into a normal state sample set and an abnormal state sample set, performing data enhancement processing on the normal state sample set and the abnormal state sample set to obtain an expanded training data set; Construct a backbone network, which adopts a DenseNet structure and includes 4 dense blocks. Each dense block includes 6 convolutional layers. Each convolutional layer is connected by dense connections. The convolution kernel size is 3×3, the step size is 1, and the padding is 1. Add a transition layer between each dense block. The transition layer includes a 1×1 convolutional layer and a 2×2 average pooling layer to reduce the spatial resolution and number of channels of the feature map. Construct multiple branch networks, each branch network corresponds to a sensor signal type, adopts a ResNet structure, includes 3 residual blocks, each residual block includes two 3×3 convolutional layers and a short-circuit connection; add an attention mechanism module between the trunk network and the multiple branch networks, including two sub-modules of channel attention and spatial attention, for adaptively adjusting feature weights; Inputting a low-dimensional feature subset in the expanded training data set into the backbone network to obtain a first backbone feature map, and inputting corresponding sensor signal features into corresponding branch networks to obtain a first branch feature map, and performing feature fusion on the first backbone feature map and the first branch feature map through the attention mechanism module to obtain a first fused feature map; Constructing a multi-task learning head, including a state classification task and a parameter regression task, wherein the state classification task uses a fully connected layer and a Softmax layer, and the parameter regression task uses a fully connected layer and a linear activation function, and inputting the first fused feature map into the multi-task learning head to obtain a state classification result and a parameter regression result; A weighted loss function is used for model training. The weighted loss function includes the cross entropy loss of state classification and the mean square error loss of parameter regression. The network parameters are updated through the back propagation algorithm and the Adam optimizer to obtain the state evaluation model.

6. The state monitoring method for multi-axis linkage CNC machining according to claim 1 is characterized in that: The second multi-source sensor data collected in real time is input into the state assessment model to perform state assessment, obtain a processing state classification result and a parameter estimation value, and map the processing state classification result and the parameter estimation value to the digital twin model to obtain a distributed collaborative monitoring result, including: Performing real-time acquisition and preprocessing on the second multi-source sensor data acquired in real time to obtain preprocessed real-time sensor data, and inputting the preprocessed real-time sensor data into the state space model to perform state estimation to obtain a real-time state sequence; Performing feature extraction and compressed sensing processing on the real-time state sequence to obtain a real-time feature subset, and inputting the real-time feature subset into the backbone network of the state assessment model to obtain a second backbone feature graph; Inputting various types of sensor signals in the preprocessed real-time sensor data into corresponding branch networks respectively to obtain multiple second branch feature maps, and performing feature fusion on the second trunk feature map and the multiple second branch feature maps through an attention mechanism module to obtain a second fused feature map; Inputting the second fused feature map into a multi-task learning head, obtaining a processing state classification result through a state classification task, and obtaining a parameter estimation value through a parameter regression task; Build a digital twin model, including geometric model, physical model and behavioral model. The geometric model represents the three-dimensional structure of the machine tool and workpiece, the physical model describes the mechanical and thermodynamic characteristics during the processing, and the behavioral model simulates the movement and control logic of each component of the machine tool; The processing state classification results and the parameter estimation values ​​are mapped to the digital twin model, the state and parameters of the digital twin model are updated, and based on the updated digital twin model, multi-machine tool collaborative state evaluation and predictive simulation are performed to obtain distributed collaborative monitoring results, which include the current state of each machine tool, predicted trends and collaborative optimization suggestions.

7. A state monitoring system for multi-axis CNC machining, characterized in that: A method for monitoring the state of multi-axis CNC machining according to any one of claims 1 to 6, the system comprising: An acquisition module is used to acquire and pre-process the vibration, acoustic emission and cutting force signals in the multi-axis linkage CNC machining process to obtain first multi-source sensor data; A processing module, used for inputting the first multi-source sensor data into a state space model builder for processing to obtain a state space model; A compression module, configured to perform feature extraction and compressive sensing on the first multi-source sensor data according to the state space model to obtain a low-dimensional feature subset; A judgment module is used to calculate the reconstruction error and Euclidean distance of the low-dimensional feature subset, and perform fusion judgment to obtain an abnormal state detection result; A training module, used for training a deep learning model including a trunk network and a plurality of branch networks according to the abnormal state detection result and the low-dimensional feature subset to obtain a state assessment model; An evaluation module is used to input the second multi-source sensor data collected in real time into the state evaluation model for state evaluation, obtain processing state classification results and parameter estimation values, and map the processing state classification results and the parameter estimation values ​​into the digital twin model to obtain distributed collaborative monitoring results.

8. A state monitoring device for multi-axis CNC machining, characterized in that: The state monitoring device for multi-axis linkage CNC machining comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the state monitoring device of the multi-axis linkage CNC machining executes the state monitoring method of the multi-axis linkage CNC machining as described in any one of claims 1-6.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, the state monitoring method of multi-axis linkage CNC machining as described in any one of claims 1 to 6 is implemented.

Images