Calibration method and device for precisely-manufactured multi-axis linkage machine tool

Abstract

The invention relates to the technical field of machine tool calibration, and discloses a calibration method and device for precisely manufacturing a multi-axis linkage machine tool, and the method comprises the steps: collecting machine tool operation data and multi-modal sensor data of the multi-axis linkage machine tool, and carrying out the error prediction, and obtaining the error prediction data; the positions of all shafts of the machine tool are measured, and nanoscale resolution ratio position data are obtained; real-time state representation of the machine tool is obtained through the multi-branch deep neural network; constructing a multi-level hypergraph based on the real-time state representation of the machine tool, and carrying out feature processing to obtain a fault feature vector; reinforcement learning is carried out through a deep Q network to obtain an adaptive compensation strategy; linkage compensation values of all axes are generated on the basis of the self-adaptive compensation strategy, real-time adjustment and fault prevention analysis are conducted on the multi-axis linkage machine tool, a fault prevention control scheme is generated, the precision maintaining capacity and operation reliability of the multi-axis linkage machine tool are improved, and the method can be suitable for different types of multi-axis linkage machine tools.

Description

Technical Field

[0001] The present invention relates to the technical field of machine tool calibration, and in particular to a calibration method and device for a precision manufacturing multi-axis linkage machine tool. Background Art

[0002] As modern manufacturing develops towards high precision and high efficiency, multi-axis linkage machine tools, as key equipment in high-end equipment manufacturing, have a direct impact on product quality due to their machining accuracy and operating stability. Traditional machine tool calibration methods mainly rely on manual experience and single sensor data, making it difficult to achieve accurate error compensation and real-time adjustment, resulting in a gradual decrease in machine tool accuracy and frequent failures during long-term operation.

[0003] At present, multi-axis linkage machine tools face technical difficulties in actual operation, such as the difficulty in effectively integrating multi-source heterogeneous data, the difficulty in maintaining nanometer-level position accuracy, and the difficulty in accurately predicting multi-axis collaborative errors. Especially under the influence of external factors such as temperature changes and vibration interference, the coupling errors between the axes of the machine tool will accumulate over time, and the traditional single compensation strategy cannot adapt to the accuracy requirements under complex working conditions. In addition, existing machine tool fault diagnosis methods are often passive, lacking active prevention and intelligent decision-making capabilities, and it is difficult to detect potential faults in a timely manner and take effective preventive measures. This leads to high machine tool maintenance costs and reduced equipment utilization. Summary of the invention

[0004] The present invention provides a calibration method and device for a precision-manufactured multi-axis linkage machine tool. The present invention improves the accuracy retention capability and operational reliability of the multi-axis linkage machine tool and can be applied to different types of multi-axis linkage machine tools.

[0005] In a first aspect, the present invention provides a precision manufacturing multi-axis linkage machine tool calibration method, the precision manufacturing multi-axis linkage machine tool calibration method comprising: Collecting machine tool operation data and multi-modal sensor data of a multi-axis linkage machine tool, and performing error prediction on the machine tool operation data through a multi-axis linkage error prediction model to obtain error prediction data; The position of each axis of the machine tool is measured using an atomic interferometer and a grating ruler, and the position data with nanometer resolution is obtained through Kalman filtering. Inputting the multimodal sensor data and the nanometer-level resolution position data into a multi-branch deep neural network, fusing them through an attention mechanism, and obtaining a real-time state representation of the machine tool; Constructing a multi-level hypergraph based on the real-time state representation of the machine tool, and inputting the multi-level hypergraph into a hierarchical hypergraph model for feature processing to obtain a fault feature vector; Taking the fault feature vector and the error prediction data as state input, performing reinforcement learning through a deep Q network to obtain an adaptive compensation strategy; Based on the adaptive compensation strategy, the linkage compensation value of each axis is generated, and the multi-axis linkage machine tool is adjusted in real time and fault prevention analysis is performed to generate a fault prevention control plan.

[0006] In a second aspect, the present invention provides a precision manufacturing multi-axis linkage machine tool calibration device, the precision manufacturing multi-axis linkage machine tool calibration device comprising: An acquisition module is used to acquire machine tool operation data and multi-modal sensor data of a multi-axis linkage machine tool, and to perform error prediction on the machine tool operation data through a multi-axis linkage error prediction model to obtain error prediction data; The measurement module is used to measure the position of each axis of the machine tool using an atomic interferometer and a grating ruler, and obtain nanometer-level resolution position data through Kalman filtering; A fusion module, used for inputting the multimodal sensor data and the nanometer-level resolution position data into a multi-branch deep neural network, and fusing them through an attention mechanism to obtain a real-time state representation of the machine tool; A feature processing module, used for constructing a multi-level hypergraph based on the real-time state representation of the machine tool, and inputting the multi-level hypergraph into a hierarchical hypergraph model for feature processing to obtain a fault feature vector; A reinforcement learning module, used for taking the fault feature vector and the error prediction data as state inputs, performing reinforcement learning through a deep Q network, and obtaining an adaptive compensation strategy; A generation module is used to generate the linkage compensation value of each axis based on the adaptive compensation strategy, and to perform real-time adjustment and fault prevention analysis on the multi-axis linkage machine tool to generate a fault prevention control plan.

[0007] The third aspect of the present invention provides a computer device, 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 computer device executes the above-mentioned precision manufacturing multi-axis linkage machine tool calibration method.

[0008] 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 precision manufacturing multi-axis linkage machine tool calibration method.

[0009] In the technical solution provided by the present invention, by fusing multimodal sensor data and nanometer-resolution position data, combined with a multi-head self-attention mechanism, high-precision perception and feature extraction of the machine tool state are achieved, thereby improving the accuracy and real-time performance of state monitoring. A multi-level hypergraph structure is used to model the physical layer, motion layer, and functional layer of the machine tool, and feature processing is performed through a layered hypergraph model, which effectively captures the correlation between different levels and improves the expression ability of fault features. The reinforcement learning method based on the deep Q network realizes the dynamic optimization of the adaptive compensation strategy, so that the compensation value can be adaptively adjusted according to the real-time state of the machine tool, significantly improving the compensation effect. A virtual model is constructed through digital twin technology, and the fault prevention plan is optimized in a virtual environment, achieving early warning and active prevention of faults, reducing equipment maintenance costs, and extending the service life of the machine tool. The present invention improves the accuracy retention capability and operational reliability of multi-axis linkage machine tools, and can be applied to different types of multi-axis linkage machine tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 A schematic diagram of the steps of a calibration method for a precision manufacturing multi-axis linkage machine tool according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a precision manufacturing multi-axis linkage machine tool calibration device in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] An embodiment of the present invention provides a method and device for calibrating a precision manufacturing multi-axis linkage machine tool. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 An embodiment of the calibration method of a precision manufacturing multi-axis linkage machine tool in an embodiment of the present invention includes: Step S1, collecting machine tool operation data and multi-modal sensor data of a multi-axis linkage machine tool, and performing error prediction on the machine tool operation data through a multi-axis linkage error prediction model to obtain error prediction data; It is understandable that the execution subject of the present invention may be a precision manufacturing multi-axis linkage machine tool calibration device, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0014] Specifically, the real-time position, velocity and acceleration of each axis of the multi-axis linkage machine tool are collected to obtain the machine tool operation data. At the same time, the temperature distribution, vibration characteristics and stress state of the machine tool are collected through the temperature sensor array, the three-axis acceleration sensor and the strain gauge to obtain the multimodal sensor data. The real-time data collection enables the system to comprehensively analyze and record the operation status of the machine tool, including the environmental factors affecting the machine tool, such as temperature changes, vibration and stress distribution. The machine tool operation data and the multimodal sensor data are synchronized in time to ensure that all data can be effectively used together to capture the real machine tool operation status and environmental changes. The sliding window method is used to process these data in segments. The size of the sliding window is set to 1 second, and the overlap rate is set to 50%. The overlapping processing enhances the capture of the continuity of the data and improves the sensitivity of the prediction model to short-term changes. The data in each window is normalized by using the standardization processing to ensure that the characteristics of various types of data are on the same scale, so that the subsequent error prediction model can effectively process these data to obtain the first standardized input data. The first standardized input data is input into the first input layer of the multi-axis linkage error prediction model for one-dimensional convolution processing. The input layer contains N parallel one-dimensional convolutional neural networks, where N represents the number of axes of the machine tool, so that each axis of the machine tool has a corresponding one-dimensional convolutional network to specifically process its data to ensure the independence and specificity of the data of each axis. Each one-dimensional convolutional neural network contains 3 convolutional layers, and the convolution kernel sizes of the convolutional layers are 5, 3, and 3, respectively, and the number of convolution kernels is 32, 64, and 128, respectively. The input data is convoluted with convolution kernels of different sizes and numbers to extract features of different scales in the running data. BatchNormalization and ReLU activation functions are connected to each convolutional layer. BatchNormalization is used to accelerate the training process and stabilize the network performance, while the ReLU activation function introduces nonlinearity, so that the network can learn complex features and obtain the first convolution feature vector. The first convolution feature vector is fused through two three-dimensional convolution layers to fuse cross-axis features. The convolution kernel size of the three-dimensional convolution layer is 3x3x3, and the number of convolution kernels is 256 and 512 respectively. The three-dimensional convolution can perform convolution operations on data of multiple axes at the same time, capturing the mutual influence and linkage features between different axes. After each three-dimensional convolution layer, the MaxPooling and Dropout operations are connected. MaxPooling is used to reduce the dimension of the feature, reduce the amount of calculation, and extract important feature information. Dropout is used to prevent overfitting and enhance the generalization ability of the model. After being processed by two three-dimensional convolution layers, the second convolution feature vector is obtained. The second convolution feature vector is input into the multi-head self-attention mechanism layer, and the correlation analysis of the features between different axes is performed through the self-attention mechanism.The number of heads of the multi-head self-attention mechanism is set to 8, and the dimension of each head is 64. The correlation weights between different axes are obtained by scaling the dot product attention calculation. The attention mechanism can give different weights to each feature, thereby strengthening the focus on the key features in the overall state, reorganizing the features, and obtaining the weighted convolution feature vector. The time series features of the weighted convolution feature vector are captured, and the long-term dependencies in the feature vector are effectively captured to obtain a feature vector with time series characteristics. The time series feature vector is input into the first fully connected layer, which contains 3 hidden layers. The number of neurons in the hidden layers is 256, 128, and 64, respectively, and the feature dimensionality is gradually reduced and abstracted. A linear activation function is used in the last hidden layer to map the features to the output space. The output dimension is N, corresponding to the predicted error values ​​of N machine tool axes, and the error prediction data is obtained.

[0015] Step S2, using an atomic interferometer and a grating ruler to measure the position of each axis of the machine tool, and obtaining nanometer-level resolution position data through Kalman filtering; Specifically, a high-precision grating ruler is installed on each axis of the multi-axis linkage machine tool. The resolution of the grating ruler is 1 nanometer and the sampling frequency is 10 kHz. The position data of the grating ruler is obtained by performing preliminary measurement on the position of each axis. As a high-precision displacement measurement tool, the grating ruler can provide accurate position information for each axis of the machine tool. An atomic interferometer system is fixed on the reference surface of the machine tool to accurately measure the position of each axis of the machine tool relative to the reference surface. The atomic interferometer system includes a frequency-stable laser light source, a spectrometer, multiple reflectors and photodetectors. Through these optical components, the atomic interferometer system can use the interference effect to accurately measure the change of position, and its measurement resolution can reach 0.1 nanometers and the sampling frequency is 100 kHz. This makes the atomic interferometer position data have a higher resolution than the grating ruler position data, and can provide an accurate description of the relative position of each axis of the machine tool. The grating ruler position data and the atomic interferometer position data are time synchronized to ensure that data from different sources have a consistent time reference to form synchronized position data. Through synchronous processing, the measurement errors and inconsistencies caused by data time offset are effectively eliminated. The synchronous position data is subjected to wavelet transform to extract the multi-scale position features of the data. Wavelet transform can analyze data in both time and frequency domains and extract the position change features at different scales. Multi-scale features help to more comprehensively characterize the dynamic characteristics of machine tool position changes, including smooth changes and transient characteristics of position at different time scales. Based on the multi-scale position features, a state space model is constructed to describe the dynamic behavior of the system. In the state space model, the state variables include position, velocity and acceleration, while the observed variable is the measured value of the position. In this way, the position change of the machine tool is described by a linear state space representation, which can well capture the dynamic relationship between position, velocity and acceleration and convert it into a form that can be used for filter processing. Based on this linear state space representation, a Kalman filter is created to estimate the state. The Kalman filter is a recursive algorithm for state estimation of dynamic systems, including a state prediction equation and a measurement update equation. When initializing the Kalman filter, the covariance matrix P of the state estimate and the process noise covariance matrix Q are set to ensure that the Kalman filter can reasonably balance the relationship between prediction and measurement to achieve the optimal state estimation. The synchronized position data is input into the Kalman filter, and the system state is estimated through recursive calculation. In the recursive process, the Kalman filter uses the state prediction equation to predict the state at the next moment based on the current state, and then uses the measurement update equation combined with the measurement data to correct the prediction result, thereby gradually updating the state estimate and error covariance. This process can effectively integrate the data of the grating scale and the atomic interferometer, make full use of the complementary advantages of the two, and improve the accuracy and resolution of the position estimation. After the Kalman filter is processed, the filtered position estimate is obtained.The filtered position estimate is subjected to an inverse wavelet transform to reconstruct the position data with nanometer resolution. The multi-scale features are restored to the position information at the original scale through the inverse wavelet transform.

[0016] Step S3, input the multimodal sensor data and nanometer-resolution position data into a multi-branch deep neural network, and fuse them through an attention mechanism to obtain a real-time state representation of the machine tool; Specifically, the multimodal sensor data and nanometer-resolution position data are time-aligned and normalized to obtain the second standardized input data. Time alignment can ensure that all data from different sources are synchronized under the same time base to avoid information loss or inconsistency caused by data time offset, while normalization is to map various types of data to the same dimension for efficient processing by the subsequent feature extraction network. The second standardized input data is divided into four branches: temperature data, vibration data, stress data, and position data, and input into four parallel feature extraction networks for feature extraction. Each feature extraction network contains three one-dimensional convolutional layers, in which the sizes of the convolution kernels are 5, 3, and 3, respectively, and the number of convolution kernels is 32, 64, and 128, respectively. Through these convolutional layers, the model can gradually perform convolution operations on the input data and extract features of different scales in the data. BatchNormalization and ReLU activation functions are connected to each convolutional layer, in which BatchNormalization is used to stabilize the training process of the network to avoid gradient disappearance or gradient explosion, and ReLU activation function introduces nonlinearity, so that the network has the ability to learn complex features and obtain the primary feature vector of each branch. The primary feature vectors of each branch are subjected to global average pooling and global maximum pooling operations, and global feature information is extracted from each feature vector, reflecting the average characteristics and extreme value characteristics of the data as a whole. Through these two pooling operations, the feature dimension is effectively reduced, and information at different levels is captured, thereby improving the robustness and expressiveness of the feature. The pooled feature vectors of each branch are concatenated to form a global feature vector. Based on the global feature vector, the correlation weights between different modalities are calculated through the scaled dot product attention mechanism to obtain the attention-weighted feature. The role of the scaled dot product attention mechanism is to perform correlation analysis on different modalities in the global feature, thereby giving appropriate weights to each modal data to better represent the importance of each modal data. In this way, the attention mechanism can enhance the focus on key features, suppress redundant or unimportant features, and make the overall feature representation more focused on information that plays a key role in state judgment. Temporal feature extraction is performed on the attention-weighted feature to capture the temporal contextual relationship between the features of each modality and obtain the temporal context vector. The temporal context vector is input into the cross-modal transformer encoder for fusion processing. The cross-modal transformer encoder contains 6 encoder layers, each of which contains a multi-head self-attention sublayer and a feed-forward neural network sublayer. The multi-head self-attention sublayer can perform multiple attention calculations on the input data to capture the complex relationship between different modal features, while the feed-forward neural network sublayer performs feature mapping and abstraction on the features after attention processing. By stacking multiple encoder layers, the transformer can deeply fuse and extract cross-modal features to obtain a fused feature representation.The fused feature representation is subjected to residual connection and layer normalization operations, where the residual connection is used to retain the information of the input features to prevent the gradient vanishing problem during deep network training, and the layer normalization is to further stabilize the training process. The fused feature representation is feature mapped by the nonlinear activation function ReLU to obtain the intermediate feature representation. The intermediate feature representation is input into the second fully connected layer, which contains 3 hidden layers, each of which contains 512, 256 and 128 neurons respectively. The input features are abstracted and reduced layer by layer to extract important information from the features. The Softmax activation function is used in the last hidden layer to map the features to the real-time state representation of the machine tool. The output representation can reflect the comprehensive state of the multi-axis linkage machine tool at the current moment, and contains information after data fusion from different modalities.

[0017] Step S4, constructing a multi-level hypergraph based on the real-time state representation of the machine tool, and inputting the multi-level hypergraph into the hierarchical hypergraph model for feature processing to obtain a fault feature vector; Specifically, the real-time state representation of the machine tool is divided into three levels: physical layer, motion layer and functional layer. Each level contains N nodes, where N is the number of axes of the machine tool, thereby obtaining a first node set. Each node in the first node set is assigned attributes. The attributes of the physical layer nodes include temperature, vibration and stress, the attributes of the motion layer nodes include position, velocity and acceleration, and the attributes of the functional layer nodes include processing accuracy and motion synchronization, thereby obtaining a second node set. Based on the second node set, a multi-level hypergraph structure is constructed. When constructing a hypergraph, the similarity between nodes is calculated by the K nearest neighbor algorithm, and the threshold of the similarity is set to 0.8. Hyperedge connections are established between nodes with similarities greater than the threshold to obtain the initial hypergraph structure. Through the K nearest neighbor algorithm, similar nodes in a specific attribute space are effectively found and connected through hyperedges. The threshold of similarity is set to 0.8 to ensure that the selected nodes have strong similarity, making the constructed hyperedge connection more reasonable and accurate. The initial hypergraph structure is subjected to intra-layer edge weight calculation, and the correlation between node attributes is calculated using cosine similarity, and the correlation is used as the edge weight to obtain a weighted intra-layer hypergraph. Cosine similarity is used to measure the cosine value of the angle between two vectors. The closer its value is to 1, the higher the similarity between the two node attributes. The intra-layer edge weight calculation using cosine similarity reflects the attribute relationship between nodes, so that the intra-layer hypergraph can more realistically represent the correlation between nodes. The inter-layer edge weight calculation is performed on the weighted intra-layer hypergraph. The Pearson correlation coefficient is used to calculate the correlation between node attributes in different layers, and the calculated correlation is used as the weight of the inter-layer edge to obtain the target multi-level hypergraph. The Pearson correlation coefficient is used to measure the linear correlation between two variables. Its value range is between -1 and 1, which can well reflect the positive or negative correlation between nodes at different levels. By weighting the inter-layer edges, the multi-level hypergraph can represent the internal relationship of nodes in each layer, capture the interactive relationship between different layers, and build a more comprehensive multi-level network structure. The target multi-level hypergraph is input into the hierarchical hypergraph model for feature processing. The hierarchical hypergraph model contains three hypergraph convolutional layers, each with 64, 128, and 256 output channels, and the activation function is LeakyReLU. The hypergraph convolutional layer is used to perform convolution operations on node features in the hypergraph structure, thereby extracting deeper features. The introduction of the LeakyReLU activation function effectively avoids the "death" of neurons, allowing the network to retain a portion of the activation output when facing negative inputs, thereby enhancing the network's nonlinear expression ability and overall feature extraction capabilities. By stacking three hypergraph convolutional layers, the model gradually abstracts and integrates the input features and extracts higher-level hierarchical features. Cross-layer feature fusion is performed on hierarchical features to obtain cross-layer fusion features. Information from different levels is integrated to capture the interaction and influence of machine tool states at different levels.Through cross-layer fusion, the feature representation includes the information within each layer and the interaction between the layers. The cross-layer fusion features are input into the third fully connected layer, which contains two hidden layers with 128 and 64 neurons respectively. The input features are gradually reduced and abstracted to further extract and retain the most critical information for fault prediction. The Tanh activation function is used in the last layer of the fully connected layer to map the features to the final output space with an output dimension of 32, thereby obtaining the fault feature vector. The output range of the Tanh activation function is -1 to 1, which can effectively perform nonlinear mapping on the features, ensure the balanced distribution of the output values, and better adapt to subsequent classification or regression tasks.

[0018] Step S5: taking the fault feature vector and the error prediction data as state input, performing reinforcement learning through a deep Q network, and obtaining an adaptive compensation strategy; Specifically, the fault feature vector is concatenated with the error prediction data to obtain the state vector. The state vector is normalized to obtain a standardized state input, which improves the stability and speed of learning and reduces the training difficulties caused by data range differences. The standardized state input is input into the second input layer of the deep Q network, which is a fully connected layer containing 256 neurons and uses the ReLU activation function. Through the processing of the fully connected layer, the high-dimensional state vector is converted into a feature representation, in which the ReLU activation function enables the model to have the ability to learn nonlinear relationships, so that the model can better extract the complex associations between different states and obtain the initial feature representation. The initial feature representation is input into a deep feature extraction module composed of three residual blocks. Each residual block contains two convolutional layers and a short-circuit connection. The size of the convolution kernel is 3x3, and the number of output channels is 128, 256, and 512 respectively. Through the residual block, the model can learn deep features on the basis of retaining the initial information. The short-circuit connection can effectively alleviate the problem of gradient disappearance in the deep network, so that information can be better transmitted between network layers and maintain the model's sensitivity to the initial features. After each convolutional layer, the model gradually captures more complex and abstract features in the state vector and obtains deep features. The deep features are input into the attention layer, and the correlation weights between features are calculated through the self-attention mechanism to obtain attention-weighted features. The self-attention mechanism analyzes the relationship between different features, gives higher weights to important features, and highlights the most critical features for current state evaluation and action selection, so that the model can focus on key feature information more accurately and enhance the ability to understand the machine tool state. The attention-weighted features are modeled in time series to obtain time series context features. Time series modeling captures the evolution relationship of features over time, thereby reflecting the dynamic changes of the machine tool state. The time series context features are input into the advantage function estimator, which contains two fully connected layers with 128 and 64 neurons respectively. The last layer uses a linear activation function to output the Q value corresponding to different compensatory actions to obtain the action value estimate. The Q value is the expected estimate of the rewards brought by different actions in the current state. The time series context features are abstracted through two fully connected layers to extract high-level features related to action selection. The introduction of the linear activation function allows the output Q value to vary within an arbitrary range, which can better distinguish the value of different actions. Based on the obtained action value estimation, the ε-greedy strategy is used to select the compensation action, where the ε value is initially set to 0.9 and gradually decreases according to the exponential decay rule. The ε-greedy strategy is a strategy that balances exploration and utilization. In the early stage, the model tends to randomly select compensation actions to explore the optimal strategy; as the training progresses, the ε value gradually decreases, and the model will select more currently known optimal actions, thereby achieving a transition from exploration to utilization and ensuring the convergence and optimality of the model.In this way, a compensation action sequence is obtained, which is used to adjust the operating state of the machine tool in real time. In order to improve the training effect of the deep Q network, the compensation action sequence is replayed. The role of experience replay is to save the compensation action and the corresponding state change. In the subsequent training process, the random sampling of batch experience samples is used to update the parameters of the Q network. This method can break the time correlation of the data, increase the independence of the data, and effectively improve the training efficiency and stability of the model. Through temporal difference learning, the parameters of the Q network are gradually updated, so that the model can more accurately evaluate the value of different actions, and finally obtain an adaptive compensation strategy.

[0019] Step S6: Generate linkage compensation values ​​for each axis based on the adaptive compensation strategy, perform real-time adjustment and fault prevention analysis on the multi-axis linkage machine tool, and generate a fault prevention control plan.

[0020] Specifically, the adaptive compensation strategy is input into the compensation value generator, and the initial compensation value of each axis is obtained by inverse kinematics calculation. The calculation process of inverse kinematics is to deduce the required displacement or rotation of each axis according to the target state and the geometric model of the machine tool to achieve the expected posture correction. The initial compensation value is nonlinearly optimized to overcome the errors and uncertainties in the kinematic model and obtain the linkage compensation value of each axis. The nonlinear optimization process uses a nonlinear programming algorithm to ensure the accuracy and linkage of the compensation value, thereby minimizing the error and improving the compensation effect. The multi-axis linkage machine tool is adjusted in real time based on the linkage compensation value of each axis, and the operating parameters of each axis are adjusted to ensure that the machine tool can maintain high-precision processing performance under different conditions. While completing the real-time adjustment, the adjusted machine tool operation data is continuously collected to monitor the operating status of the machine tool and ensure the effectiveness of the compensation strategy. The collected data is fast Fourier transformed to obtain the frequency domain characteristics of the machine tool in the adjusted operating state. Fast Fourier transform converts the operating signal of the machine tool from the time domain to the frequency domain, which is convenient for analyzing abnormal vibration or periodic characteristics in the operation process of the machine tool. The frequency domain features are input into a multi-class support vector machine for fault prediction. The multi-class support vector machine is an effective classification algorithm used to classify input features to determine whether there is a fault or abnormality in the current state of the machine tool. Among them, the kernel function uses the radial basis function, which can handle nonlinear data well and adapt to complex feature distribution, thereby improving the accuracy of fault prediction. Through the calculation of the support vector machine, the probability distribution of the fault type is obtained, and the faults of the machine tool are classified and predicted based on the distribution. The threshold judgment of the probability distribution of the fault type is performed, and the threshold of the fault warning is set to 0.8. When the predicted probability of a certain type of fault exceeds this threshold, the fault warning signal is triggered. After the fault warning signal is triggered, the fault tracing analysis program is immediately started. The main task of the program is to find out the specific cause of the warning signal. First, a fault causal relationship graph is constructed, and the root cause of the fault is located through probabilistic reasoning. The fault causal relationship graph is a model based on a graph structure, in which nodes represent different states or parameters of the system, and edges represent the causal relationship between these states or parameters. By performing probabilistic reasoning on the fault causal relationship graph, it is possible to find out which nodes have a strong causal relationship, so as to accurately locate the root cause of the fault and obtain the fault tracing result. Based on the fault tracing results, a virtual model of the machine tool is constructed using digital twin technology. Fault prevention solutions are executed and optimized in a virtual environment. Digital twin technology can realistically simulate the operating performance of machine tools in different states by building a mirror model of the machine tool in a virtual environment. In the virtual model, different control solutions are simulated and compared based on the fault tracing results to find the optimal fault prevention solution. Based on the verification results in the digital twin environment, a fault prevention control solution is formed and applied to the actual machine tool control, thereby achieving real-time adjustment and fault prevention analysis of multi-axis linkage machine tools.At the same time, real-time adjustment based on adaptive compensation strategy can effectively compensate for the errors of machine tools during operation and ensure high precision and consistency of the machining process.

[0021] In the embodiment of the present invention, by fusing multimodal sensor data and nanometer-resolution position data, combined with a multi-head self-attention mechanism, high-precision perception and feature extraction of the machine tool state are achieved, and the accuracy and real-time performance of state monitoring are improved. A multi-level hypergraph structure is used to model the physical layer, motion layer, and functional layer of the machine tool, and feature processing is performed through a layered hypergraph model, which effectively captures the correlation between different levels and improves the expression ability of fault features. The reinforcement learning method based on the deep Q network realizes the dynamic optimization of the adaptive compensation strategy, so that the compensation value can be adaptively adjusted according to the real-time state of the machine tool, significantly improving the compensation effect. A virtual model is constructed through digital twin technology, and the fault prevention plan is optimized in a virtual environment, which realizes early warning and active prevention of faults, reduces equipment maintenance costs, and extends the service life of the machine tool. The present invention improves the accuracy retention ability and operation reliability of multi-axis linkage machine tools, and can be applied to different types of multi-axis linkage machine tools.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: The real-time position, speed and acceleration of each axis of the multi-axis linkage machine tool are collected to obtain the machine tool operation data. The temperature distribution, vibration characteristics and stress state of the machine tool are collected through the temperature sensor array, three-axis acceleration sensor and strain gauge to obtain multi-modal sensor data. The machine tool operation data and the multimodal sensor data are synchronized in time, and the data are segmented by a sliding window method, with a window size of 1 second and an overlap rate of 50%. Then, the data in each window is standardized to obtain the first standardized input data. The first standardized input data is input into the first input layer of the multi-axis linkage error prediction model for one-dimensional convolution processing to obtain the first convolution feature vector. The first input layer contains N parallel one-dimensional convolutional neural networks, where N is the number of machine tool axes. Each one-dimensional convolutional neural network contains 3 convolutional layers, with convolution kernel sizes of 5, 3, and 3, respectively, and the number of convolution kernels of 32, 64, and 128, respectively. Each convolution layer is followed by BatchNormalization and ReLU activation functions. The first convolution feature vector is extracted and cross-axis feature fusion is performed through two three-dimensional convolution layers. The convolution kernel size is 3x3x3, and the number of convolution kernels is 256 and 512 respectively. Each layer is followed by MaxPooling and Dropout operations to obtain the second convolution feature vector. The second convolution feature vector is input into the multi-head self-attention mechanism layer, with 8 heads and 64 dimensions of each head. The correlation weights between different axes are calculated by scaling the dot product attention and feature reorganization to obtain the weighted convolution feature vector. The time series features and long-term dependencies of the weighted convolution feature vector are captured to obtain the time series feature vector, and the time series feature vector is input into the first fully connected layer. The first fully connected layer contains 3 hidden layers with 256, 128, and 64 neurons respectively. The last layer uses a linear activation function with an output dimension of N, corresponding to the prediction error values ​​of N axes, to obtain the error prediction data.

[0023] Specifically, the real-time position, velocity and acceleration of each axis of the multi-axis linkage machine tool are collected to obtain the operation data of the machine tool, reflecting the dynamic motion state of the machine tool at different time points. At the same time, the environmental status information of the machine tool is obtained through a variety of sensors. The temperature sensor array accurately measures the temperature distribution of different parts of the machine tool, the three-axis acceleration sensor is used to measure the vibration characteristics of the machine tool during operation, and the strain gauge is used to measure the stress state of the machine tool. These data together constitute the multimodal sensor data. The machine tool operation data and the multimodal sensor data are synchronized in time to ensure that all data are processed under the same time reference and eliminate information deviation caused by time inconsistency. The sliding window method is used to segment the data, with a window size of 1 second and an overlap rate of 50%. The role of sliding window segmentation is to ensure the continuity of the data and enhance the sensitivity to short-term changes by overlapping. The data in each window is standardized to eliminate the differences between different feature dimensions and obtain the first standardized input data. The first standardized input data is input into the first input layer of the multi-axis linkage error prediction model for one-dimensional convolution processing. The first input layer contains N parallel one-dimensional convolutional neural networks, where N is the number of axes of the machine tool. The data of each axis is extracted through an independent convolutional network to ensure that the features of each axis can be extracted and learned independently. Each one-dimensional convolutional neural network contains 3 convolutional layers, with convolution kernel sizes of 5, 3, and 3, respectively, and the number of convolution kernels is 32, 64, and 128, respectively. The purpose of the convolution operation is to extract local features in the input data. The convolution kernel size of 5 means that the data is subjected to a sliding convolution of 5 time steps, which can capture features in a larger range, while the subsequent 3, 3 convolution kernels can capture more subtle feature changes. Each convolution layer is followed by BatchNormalization and ReLU activation functions. BatchNormalization is used to normalize the convolution output, thereby accelerating the network training process and improving the stability of the network; the ReLU activation function is used to introduce nonlinearity, so that the network can learn more complex feature relationships. After convolution processing, the first convolution feature vector is obtained. The first convolution feature vector is input into two three-dimensional convolution layers for cross-axis feature fusion. The convolution kernel size of the three-dimensional convolution is 3x3x3, and the number of convolution kernels is 256 and 512 respectively. The three-dimensional convolution performs convolution operations on features in both spatial and temporal dimensions to capture the linkage relationship between different axes. Through three-dimensional convolution, the associated features are effectively extracted to enhance the understanding of the overall state of the system. MaxPooling and Dropout operations are connected to each three-dimensional convolution layer. MaxPooling is used to reduce the dimension of the feature and reduce the amount of calculation while retaining important feature information; Dropout is used to prevent overfitting and enhance the generalization ability of the model. After processing by these two three-dimensional convolution layers, the second convolution feature vector is obtained.The second convolution feature vector is input into the multi-head self-attention mechanism layer. The number of heads of the multi-head self-attention mechanism is 8, and the dimension of each head is 64. The role of the self-attention mechanism is to calculate the correlation weights between different axes by scaling the dot product attention, thereby realizing the reorganization of features. The self-attention mechanism calculates the attention weight by the following formula:. ; in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key vector. Through this formula, the correlation between different axes is obtained, and the correlation is used to weight the features to obtain a weighted convolution feature vector. Time series feature extraction is performed on the weighted convolution feature vector to capture the long-term dependencies in the features. The time series feature vector is input into the first fully connected layer. The fully connected layer contains 3 hidden layers with 256, 128 and 64 neurons respectively. The role of the fully connected layer is to further integrate and abstract the features to extract the most representative feature information. In the last layer, a linear activation function is used to map the features to the output space. The output dimension is N, corresponding to the prediction error values ​​of the N axes, and the error prediction data is obtained.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: A high-precision grating ruler is installed on each axis of the multi-axis linkage machine tool. The grating ruler resolution is 1nm and the sampling frequency is 10kHz. The position of each axis is preliminarily measured to obtain the grating ruler position data. An atomic interferometer system is fixed on the machine tool reference surface. The atomic interferometer system includes a frequency-stable laser light source, a beam splitter, a reflector, and a photodetector. The measurement resolution of the atomic interferometer is 0.1nm, and the sampling frequency is 100kHz. The position of each axis relative to the reference surface is accurately measured to obtain the atomic interferometer position data. Performing time synchronization on the grating ruler position data and the atomic interferometer position data to obtain synchronized position data, and performing wavelet transformation on the synchronized position data to obtain multi-scale position features; A state space model is constructed based on multi-scale position features. The state variables include position, velocity and acceleration. The observed variables are position measurements, and a linear state space representation is obtained. A Kalman filter is created based on the linear state space representation. The Kalman filter includes a state prediction equation and a measurement update equation. The state estimation covariance matrix P and the process noise covariance matrix Q are initialized to obtain the Kalman filter parameters. The synchronized position data is input into the Kalman filter, and state estimation and error covariance update are performed through recursive calculation to obtain a filtered position estimate value, and an inverse wavelet transform is performed on the filtered position estimate value to reconstruct the nanometer-level resolution position data.

[0025] Specifically, a high-precision grating ruler is installed on each axis of the machine tool for preliminary position measurement. The resolution of the grating ruler is 1 nanometer and the sampling frequency is 10 kHz. The position of each axis of the machine tool is accurately measured in real time to obtain the grating ruler position data. As a high-precision position measurement tool, the grating ruler can provide position feedback for each axis, so that the operation status of the machine tool can be accurately described. In order to obtain higher-precision data, an atom interferometer system is fixed on the reference surface of the machine tool to accurately measure the position of each axis relative to the reference surface. The system includes a frequency-stable laser light source, a spectrometer, a reflector and a photodetector. Its measurement resolution reaches 0.1 nanometers and the sampling frequency is 100 kHz. High-precision position data is obtained through the interferometric measurement principle of the atom interferometer. The grating ruler position data and the atom interferometer position data are time-synchronized to ensure that all position data are processed under the same time reference, and synchronized position data is obtained, which contains the position information of the machine tool at each moment. Wavelet transform is performed on the synchronized position data to extract multi-scale position features. Wavelet transform is a tool for analyzing the changes of signals in the time and frequency domains, which can decompose the signal into detail and trend components at different scales. Assume that the synchronous position data is , wavelet transform is expressed as: ; in, are the wavelet coefficients, is the scale parameter, which indicates the resolution of the signal. is the translation parameter, which indicates the time position of the signal, is the mother wavelet function. Through wavelet transform, the position data is decomposed into multi-scale position features, which effectively represents the dynamic changes of the machine tool at different time scales. Based on the multi-scale position features, a state space model is constructed to describe the dynamic behavior of the system. In the state space model, the state variables include position, velocity and acceleration, while the observed variables are the position measurements. The purpose of the state space model is to describe the motion of the machine tool as a dynamic system, in which the relationship between position, velocity and acceleration is represented by a mathematical model. The state equation is expressed as: ; ; in, It is at the moment The state vector of , including position, velocity and acceleration, is the state transition matrix, describing the change of state over time, is the control matrix, which represents the influence of the system input on the state. is the control input, is the process noise, representing the uncertainty of the model, is the observation vector, is the observation matrix, is the observation noise. Through the state space model, the motion process of the machine tool is described by a linear system. A Kalman filter is created based on the linear state space model to optimally estimate the system state. The Kalman filter is a recursive algorithm that includes a state prediction equation and a measurement update equation. The state prediction equation is used to estimate the state at the next moment based on the current state, and the measurement update equation is used to correct the prediction result in combination with the measurement data. The state prediction and measurement update process of the Kalman filter is expressed as the following two equations: ; ; in, It is the prediction of the state at the next moment. is the updated state estimate, is the Kalman gain, which indicates the degree of correction of the prediction by the measurement. Initialize the state estimation covariance matrix and the process noise covariance matrix , so as to make a reasonable trade-off for the uncertainty of state estimation in recursive calculation. The synchronized position data is input into the Kalman filter, and the state estimation and error covariance are updated through recursive calculation to obtain the filtered position estimate. Kalman filtering combines the advantages of multi-source data, removes noise and reduces measurement errors at the same time, and obtains a more accurate position estimate. The filtered position estimate is subjected to inverse wavelet transform to reconstruct the position data with nanometer resolution. The inverse wavelet transform restores the features in the wavelet domain back to the time domain to obtain the original position data with high precision.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Performing time alignment and normalization processing on the multimodal sensor data and the nanometer-resolution position data to obtain second standardized input data; The second standardized input data is divided into four branches: temperature data, vibration data, stress data, and position data. They are input into four parallel feature extraction networks respectively. Each feature extraction network contains three one-dimensional convolutional layers. The convolution kernel sizes are 5, 3, and 3 respectively. The number of convolution kernels is 32, 64, and 128 respectively. Each layer is followed by BatchNormalization and ReLU activation functions to obtain the primary feature vector of each branch. Perform global average pooling and global maximum pooling operations on the primary feature vector of each branch to obtain the pooled feature vector of each branch, and concatenate the pooled feature vectors of each branch to obtain the global feature vector; Based on the global feature vector, the correlation weights between different modalities are calculated through the scaled dot product attention mechanism to obtain the attention weighted features, and the time series features are extracted from the attention weighted features to obtain the time series context vector; The temporal context vector is input into the cross-modal transformer encoder. The cross-modal transformer encoder contains 6 encoder layers, each of which contains a multi-head self-attention sublayer and a feed-forward neural network sublayer. The fused feature representation is obtained, and the fused feature representation is residually connected and layer normalized. The feature mapping is performed through the nonlinear activation function ReLU to obtain the intermediate feature representation. The intermediate feature representation is input into the second fully connected layer. The second fully connected layer contains three hidden layers with 512, 256, and 128 neurons respectively. The last layer uses the Softmax activation function to output the real-time status representation of the machine tool.

[0027] Specifically, time alignment and normalization are performed on multimodal sensor data and nanometer-resolution position data. Multimodal sensor data includes information such as temperature, vibration and stress, and nanometer-resolution position data provides accurate spatial position information. Time alignment enables all data from different sources to have the same time reference to ensure the accuracy of subsequent analysis and eliminate the deviation caused by different sampling times. Normalization is to convert all features to the same dimension so that they have the same mean and variance, reducing the impact of differences in numerical ranges of different features on model training. After completing time alignment and normalization, the second standardized input data is obtained. The second standardized input data is divided into four branches: temperature data, vibration data, stress data and position data, and input into four parallel feature extraction networks respectively. Each feature extraction network contains three one-dimensional convolution layers, with convolution kernel sizes of 5, 3, and 3, respectively, and the number of convolution kernels is 32, 64 and 128, respectively. The main purpose of the convolution operation is to extract local features in the data, such as extracting the temperature change pattern over time in temperature data, or capturing specific vibration frequencies in vibration data. The first convolution kernel size is 5, which means that the convolution operation extracts features of a relatively long time span, while the subsequent convolution kernel size is 3 to capture more subtle temporal features. Each convolution layer is followed by BatchNormalization and ReLU activation functions. BatchNormalization is used to normalize the output of the convolution to ensure that the model training process is more stable and efficient, and the ReLU activation function introduces nonlinearity so that the network can learn more complex features. Through the processing of the convolution layer, the primary feature vector of each branch is obtained. Global average pooling and global maximum pooling operations are performed on the primary feature vector of each branch. Global average pooling compresses features by calculating the average value of each feature channel, while global maximum pooling takes the maximum value of each feature channel. The features are summarized at different scales, the most representative information is retained, and the dimension of the features is reduced to obtain the pooled feature vector of each branch. The pooled feature vectors of each branch are concatenated to obtain the global feature vector. Based on the global feature vector, the correlation weights between different modalities are calculated by the scaled dot product attention mechanism to obtain the attention-weighted features. The scaled dot product attention mechanism calculates the similarity between features by the following formula: ; in, , and Represent query, key and value matrices respectively, which are obtained by linear transformation of input features; represents the dimension of the key vector, by dividing the dot product result by Scaling is performed to prevent the dot product value from becoming too large as the dimension increases. The dot product result is converted into a weight through the softmax function and then combined with the value matrix Multiply them together to obtain weighted features, so that the model pays attention to the intrinsic relationship between different modal features and enhances the focus on key features. Perform time series feature extraction on the attention weighted features to obtain the time series context vector, which captures the evolution of each modal feature in the time dimension, especially the long-term dependencies of dynamic changes in machine tool operation. Input the time series context vector into the cross-modal transformer encoder. The cross-modal transformer encoder consists of 6 encoder layers, each of which contains a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention mechanism processes the input in parallel through multiple heads, and each head pays attention to different aspects of the features, thereby enhancing the model's ability to express the input features. For each head, the self-attention is calculated using the following formula: ; in, represents the number of attention heads, The linear transformation matrix is ​​the output. The outputs of all heads are concatenated and then linearly transformed to obtain the fused feature representation. Each encoder layer contains a feedforward neural network sublayer, which is used to perform linear transformation and nonlinear mapping on the features output by the self-attention sublayer. The encoder layer includes residual connection and layer normalization operations. The residual connection alleviates the problem of gradient vanishing in deep networks and ensures that the model can still retain the original feature information at a deeper level, while the layer normalization can improve the stability and convergence speed of training. Through multiple processing of the encoder layer, the features from different modalities are effectively fused to obtain a more comprehensive fused feature representation. After the fused feature representation is processed, the feature mapping is performed through the nonlinear activation function ReLU to obtain the intermediate feature representation. The intermediate feature representation contains the feature information of temperature, vibration, stress and position data under multimodal interaction, which is a high-level description of the current operating status of the machine tool. The intermediate feature representation is input into the second fully connected layer, which contains 3 hidden layers. The number of neurons in each hidden layer is 512, 256 and 128 respectively. The input features are further abstracted layer by layer to extract the information most relevant to the machine tool status. The gradual dimensionality reduction of the hidden layer can reduce the amount of calculation while retaining the most meaningful features. The Softmax activation function is used in the last layer of the fully connected layer to map the intermediate features to the output space to obtain the real-time state representation of the machine tool. The Softmax activation function converts the feature value into a probability distribution so that the output represents the possibility of different states, thereby helping the operator or control system to make real-time judgments and adjustments to the machine tool state.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: The real-time state representation of the machine tool is divided into three levels: physical layer, motion layer and function layer. Each level contains N nodes, where N is the number of machine tool axes, to obtain a first node set. Assigning attributes to the nodes in the first node set, where the physical layer node attributes include temperature, vibration and stress, the motion layer node attributes include position, velocity and acceleration, and the functional layer node attributes include processing accuracy and motion synchronization, to obtain a second node set; A multi-level hypergraph structure is constructed based on the second node set, the similarity between nodes is calculated by the K nearest neighbor algorithm, the similarity threshold is set to 0.8, and hyperedge connections are established between nodes with similarity greater than the threshold to obtain an initial hypergraph structure; The initial hypergraph structure is subjected to intra-layer edge weight calculation, and the correlation between node attributes is calculated using cosine similarity, and the correlation is used as the edge weight to obtain an intra-layer hypergraph with weights. The inter-layer edge weights of the weighted intra-layer hypergraph are calculated, the Pearson correlation coefficient is used to calculate the correlation between the attributes of nodes in different layers, and the correlation is used as the inter-layer edge weight to obtain the target multi-level hypergraph; The target multi-level hypergraph is input into the hierarchical hypergraph model. The hierarchical hypergraph model contains three hypergraph convolutional layers. The number of output channels of each hypergraph convolutional layer is 64, 128 and 256 respectively. The activation function is LeakyReLU to obtain hierarchical features. The hierarchical features are fused across layers to obtain cross-layer fusion features, and the cross-layer fusion features are input into the third fully connected layer. The third fully connected layer contains two hidden layers with 128 and 64 neurons respectively. The last layer uses the Tanh activation function with an output dimension of 32 to obtain the fault feature vector.

[0029] Specifically, the real-time state representation of the machine tool is divided into three levels: physical layer, motion layer and functional layer. Each level contains N nodes, where N is the number of axes of the machine tool. These three levels reflect the operating status of the machine tool at different levels: the physical layer describes physical information related to the environment, such as temperature, vibration and stress; the motion layer is related to the motion parameters of the machine tool, including position, speed and acceleration; the functional layer represents the processing performance of the machine tool, such as processing accuracy and motion synchronization between axes. Through layering, the first node set is obtained, and each node corresponds to the state of a certain axis at a certain level. The nodes in the first node set are assigned attributes. The node attributes of the physical layer include temperature, vibration and stress, which directly affect the mechanical properties and stability of the machine tool; the node attributes of the motion layer include position, speed and acceleration, which describe the dynamic behavior of each axis of the machine tool; and the node attributes of the functional layer include processing accuracy and motion synchronization, which are related to the actual processing effect of the machine tool. By assigning attributes to the nodes, the second node set is obtained, which describes the state characteristics of the machine tool at the three levels of physical, motion and function. Based on the second node set, a multi-level hypergraph structure is constructed. A hypergraph is a graph structure that can connect multiple nodes. Unlike traditional simple edge connections, a hyperedge connection is used to describe the relationship between multiple nodes. In order to determine which nodes need to be connected by hyperedges, the K nearest neighbor algorithm is used to calculate the similarity between nodes. The K nearest neighbor algorithm can effectively find similar nodes in the attribute space. By setting the similarity threshold to 0.8, hyperedge connections are established between nodes with similarity greater than the threshold to obtain the initial hypergraph structure. The degree of similarity reflects the degree of similarity between node attributes. Through the hyperedge connection between high-similarity nodes, the linkage relationship between similar attributes in the machine tool status is captured. The initial hypergraph structure is processed, the weights of the edges within the layer are calculated, and the cosine similarity is used to calculate the correlation between node attributes. The formula for cosine similarity is as follows: CosineSimilarity ; in, and are the attribute vectors of the two nodes, represents the dot product of vectors, and are the moduli of the vectors respectively. The similarity between each pair of node attributes is calculated by cosine similarity, and these similarities are used as the weights of the intra-layer edges to obtain a weighted intra-layer hypergraph. The value range of cosine similarity is between 0 and 1. The closer the value is to 1, the more similar the attributes between the two nodes are. The weighted intra-layer hypergraph can more accurately describe the correlation between the machine tool status at the same level. The weights of the inter-layer edges are calculated for the weighted intra-layer hypergraph. The Pearson correlation coefficient is used to calculate the correlation between the attributes of nodes in different layers. The formula for the Pearson correlation coefficient is: ; in, and are the attribute vectors of the two nodes respectively, and are the means of the vectors, is the dimension of the vector. The Pearson correlation coefficient is used to measure the linear correlation between nodes in different layers. The value of the correlation coefficient is between -1 and 1. The closer its absolute value is to 1, the closer the relationship between the two nodes is. The correlation is used as the weight of the inter-layer edge to obtain the target multi-level hypergraph. The target multi-level hypergraph is input into the hierarchical hypergraph model for feature extraction. The hierarchical hypergraph model contains 3 hypergraph convolution layers. The number of output channels of each hypergraph convolution layer is 64, 128 and 256 respectively, and the activation function of each layer is LeakyReLU. The function of the hypergraph convolution layer is to perform convolution operations on the node features in the hypergraph and extract high-order features between nodes. Through hypergraph convolution, the information between nodes and hyperedges is effectively aggregated to capture complex multi-node relationships. The LeakyReLU activation function is used to introduce nonlinearity. The formula of LeakyReLU is: ; in, is a positive number less than 1, usually 0.01. Used to handle situations where the input is a negative number. Through the activation function, the "death" of neurons is effectively avoided, and the nonlinear expression ability of the network and the robustness of feature extraction are enhanced. Cross-layer feature fusion is performed on the hierarchical features to integrate information from different levels to capture the interaction and influence of the machine tool status at different levels. Through the fusion operation, cross-layer fusion features are obtained. The cross-layer fusion features are input into the third fully connected layer. The fully connected layer contains two hidden layers with 128 and 64 neurons in the hidden layers respectively. The input features are abstracted and reduced layer by layer to extract the information most relevant to the fault features. The Tanh activation function is used in the last layer of the fully connected layer to map the features to the output space with an output dimension of 32 to obtain the fault feature vector. The output range of the Tanh activation function is -1 to 1, so that the eigenvalues ​​are evenly distributed within a certain range. The formula is: ; Through Tanh activation, the feature value is ensured to remain within a stable range, which is helpful for subsequent feature analysis and fault diagnosis.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: The fault feature vector and the error prediction data are concatenated to obtain a state vector, and the state vector is normalized to obtain a standardized state input; The standardized state is input into the second input layer of the deep Q network. The second input layer is a fully connected layer with 256 neurons and uses the ReLU activation function to obtain the initial feature representation. Feature extraction is performed on the initial feature representation, and deep feature learning is performed through three residual blocks. Each residual block contains two convolutional layers and a short-circuit connection. The convolution kernel size is 3x3, and the number of output channels is 128, 256, and 512 respectively to obtain deep features; The deep features are input into the attention layer, and the correlation weights between features are calculated using the self-attention mechanism to obtain the attention-weighted features. The attention-weighted features are then temporally modeled to obtain the temporal context features. The temporal context features are input into the advantage function estimator, which contains two fully connected layers with 128 and 64 neurons respectively. The last layer uses a linear activation function and outputs the Q value to obtain the action value estimate. Based on action value estimation, the ε-greedy strategy is adopted to select compensation actions. The ε value is initially set to 0.9 and gradually reduced according to the exponential decay rule to obtain the compensation action sequence. The compensation action sequence is then replayed empirically, batch experience samples are randomly sampled, and the Q network parameters are updated through temporal difference learning to obtain an adaptive compensation strategy.

[0031] Specifically, the fault feature vector and the error prediction data are concatenated to obtain the state vector. The state vector integrates the system's fault information and error prediction, and can fully reflect the current operating status and potential deviations of the machine tool. The state vector is normalized to standardize all data and eliminate the dimensional differences between different features, so that each feature is within the same scale range, thereby improving the subsequent training stability and efficiency of the deep neural network and obtaining a standardized state input. The standardized state input is sent to the second input layer of the deep Q network, which is a fully connected layer containing 256 neurons and using the ReLU activation function. The function of the fully connected layer is to perform linear transformation and nonlinear activation on the input state vector to obtain the initial feature representation. Through the ReLU activation function, nonlinear features are introduced, so that the model can learn the relationship between complex states and actions. The definition of the ReLU activation function is as follows: ; in, are input features, is the output feature. When the input is greater than 0, the output is equal to the input itself, and when the input is less than or equal to 0, the output is 0. The use of ReLU effectively solves the problem of gradient vanishing in traditional activation functions and enhances the learning ability of deep neural networks for complex features. Deep feature extraction is performed on the initial feature representation, and three residual blocks are used to achieve deep feature learning. Each residual block contains two convolutional layers and a short-circuit connection, which maintains information flow in the deep network and prevents the problem of gradient vanishing due to excessive depth. The convolution kernel size of the convolutional layer is 3x3, and the number of output channels is 128, 256, and 512 respectively. The gradually increasing number of output channels can capture more feature details. The residual block adds the input directly to the output through a short-circuit connection. The formula is as follows: ; in, represents the input features, represents the weight of the convolutional layer, represents the output features after two convolutional layers, is the output of the residual block. Through short-circuit connection, the model retains the input feature information and reduces the risk of gradient disappearance, so that deep features can be effectively learned. The deep features are input into the attention layer, and the self-attention mechanism is used to calculate the correlation weights between features. The self-attention mechanism performs global correlation calculations on the input features, allowing the model to focus on the important relationships between features. The calculation formula of attention weight is as follows: ; in, , and They represent query, key, and value matrices respectively, which are obtained by linear transformation of input features. Represents the dimension of the key vector, which is used to scale the dot product result to prevent the dot product value from becoming too large as the dimension increases. The softmax function converts the dot product result into a probability distribution to obtain the correlation weights between features, and reorganizes the features according to the weights to finally obtain the attention-weighted features. The attention-weighted features are modeled in time series to capture the time dependency of the machine tool state and obtain the time series context features, so that the model can take into account the dynamic changes of the machine tool state over time, especially the long-term trend changes, and better predict and compensate for the machine tool errors. The time series context features are input into the advantage function estimator, which contains two fully connected layers with 128 and 64 neurons respectively. The fully connected layer is used to abstract the input features to extract the most important information related to the compensation action. A linear activation function is used in the last layer to output the Q value corresponding to different actions. The formula is as follows: ; in, Indicates status Next action The value of represents the state value function, represents the advantage function, which is used to estimate the advantage of a certain action over other actions in a specific state. In this way, the model evaluates the value of different compensatory actions and obtains the action value estimate. Based on the action value estimate, - Greedy strategy to select compensating actions. -The core idea of ​​the greedy strategy is to balance exploration and exploitation, where the ε value is initially set to 0.9 and gradually decreases according to the exponential decay rule. The larger ε value at the beginning makes the model tend to explore more different actions to fully understand the dynamic characteristics of the system; as the training progresses, As the value decreases, the model will choose more currently known optimal actions, thus achieving a transition from exploration to utilization. Specifically, -The selection rule of the greedy strategy is expressed as: ; in, Indicates at time Select the action, Is the current state The value of different actions is estimated under this strategy. Through this strategy, a sequence of compensating actions is obtained. Experience replay is performed on the sequence of compensating actions. The compensating actions and the corresponding state changes are saved, and batches of experience samples are randomly sampled for training during the training process. In this way, the time correlation of the data is broken, the independence of the data is increased, and the generalization ability of the model and the stability of training are improved. During the experience replay, the parameters of the deep Q network are updated through temporal difference learning. Temporal difference learning combines the advantages of dynamic programming and Monte Carlo methods, and gradually adjusts the parameters during the state estimation process, so that the model can more accurately predict future states and action benefits. The specific temporal difference update formula is as follows: ; in, is the learning rate, which controls the step size of parameter update; For at the moment Immediate rewards after performing an action; is the discount factor, which measures the importance of future earnings; is the next state reached after executing the action; Represents the value estimate of the optimal action in the next state. Through this recursive calculation, the strategy of the deep Q network is gradually improved to obtain a more accurate adaptive compensation strategy.

[0032] In a specific embodiment, the process of executing step S6 may specifically include the following steps: The adaptive compensation strategy is input into the compensation value generator, and the initial compensation value of each axis is obtained by inverse kinematics calculation. The initial compensation value of each axis is nonlinearly optimized to obtain the linkage compensation value of each axis. The multi-axis linkage machine tool is adjusted in real time based on the linkage compensation value of each axis, and the operation data of the adjusted machine tool is collected, and the adjusted machine tool operation data is subjected to fast Fourier transform to obtain frequency domain characteristics; The frequency domain features are input into the multi-class support vector machine for fault prediction. The radial basis function is selected as the kernel function to obtain the probability distribution of the fault type. The fault type probability distribution is judged by the threshold. The fault warning threshold is set to 0.8. When the probability of a certain type of fault exceeds the threshold, the fault warning signal is triggered. Based on the fault warning signal, the fault tracing analysis program is started, a fault cause-effect relationship diagram is constructed, the root cause of the fault is located through probabilistic reasoning, and the fault tracing results are obtained. Based on the fault tracing results, a virtual model of the machine tool is constructed using digital twin technology, and the fault prevention plan is executed and optimized in the virtual environment to obtain a fault prevention control plan.

[0033] Specifically, the adaptive compensation strategy is input into the compensation value generator, and the initial compensation value of each axis is obtained by inverse kinematics calculation. Inverse kinematics refers to calculating the motion parameters of each axis according to the desired end state to obtain the appropriate compensation value of each axis. Assume that the end state of the machine tool is determined by the target position Given, the calculation process of inverse kinematics is described as: ; in, is the angle or displacement compensation value vector of each axis, is the inverse kinematics solver function, is the desired end position information. The initial compensation value of each axis is obtained by inverse kinematics solution. The initial compensation value is nonlinearly optimized. Under the consideration of system constraints, the initial compensation value is further adjusted so that the final compensation value can achieve the optimal correction effect. Nonlinear optimization is achieved by minimizing a certain error function, which is defined as the sum of squares of the error between the actual position and the desired position: ; in, is the error function, For the The actual position of the axis, is the expected position, is the number of axes. By optimizing the error function, the compensation value of each axis that minimizes the error is found, and finally the linkage compensation value of each axis is obtained. These compensation values ​​are used to adjust the movement of each axis so that the end state of the machine tool is closer to the desired position and the machining accuracy is improved. Based on the linkage compensation value of each axis, the multi-axis linkage machine tool is adjusted in real time to ensure the accuracy and stability of the machine tool during operation. During the real-time adjustment process, the adjusted machine tool operation data is collected at the same time to monitor the compensation effect and the operation status of the machine tool. The collected data includes the displacement, speed, vibration signal, etc. of each axis. In order to analyze the frequency characteristics of the machine tool during operation, the adjusted operation data is subjected to fast Fourier transform, the data is converted from the time domain to the frequency domain, and the frequency domain features are extracted. The formula for fast Fourier transform is: ; in, Indicates the frequency The frequency domain signal, For the The time domain signal of sampling points is is the total number of sampling points. Through fast Fourier transform, the spectrum of the signal is obtained to analyze the vibration frequency and other frequency domain characteristics of the machine tool during operation. The frequency domain features are input into the multi-class support vector machine (SVM) for fault prediction. The multi-class support vector machine is a classification tool that can classify the input frequency domain feature data to determine whether the machine tool has a fault and the possible fault type. The kernel function uses the radial basis function, and the expression of the radial basis function is: ; in, The input feature vector and The similarity between is the parameter of the kernel function, which is used to control the range of similarity. Through the radial basis function, the support vector machine can map the input features to a high-dimensional space, find the best hyperplane in this space, separate different fault types, and obtain the probability distribution of the fault type. The probability distribution of the fault type is judged by the threshold, and the threshold of the fault warning is set to 0.8. When the probability of a certain type of fault exceeds this threshold, the system will trigger a fault warning signal. The fault warning signal can promptly remind the operator or the control system that the machine tool has a certain type of serious fault and needs to take immediate measures to deal with it. The fault source tracing analysis program is triggered by the fault warning signal to find the root cause of the fault. Construct a fault causal relationship graph. The fault causal relationship graph is a graph structure used to describe the causal relationship between various variables in the system, where nodes represent variables in the system and edges represent the causal relationship between variables. Through this graph structure, the propagation path of the fault is described. Based on the fault causal relationship graph, the probabilistic reasoning method is used to locate the root cause of the fault. Probabilistic reasoning is a reasoning method based on Bayes' theorem. By calculating the probabilities of different fault paths, the most likely cause of the fault is found. Based on the fault tracing results, a virtual model of the machine tool is constructed using digital twin technology. Digital twin is a simulation technology that constructs a digital model corresponding to the actual machine tool in a virtual environment to realistically simulate the operating behavior of the machine tool in different states. In the virtual environment, based on the results of fault tracing, the fault prevention plan is executed and optimized. This process involves simulating different preventive measures in the virtual model, analyzing their effects, and selecting the best plan. In this way, the fault prevention plan is fully verified and optimized without affecting the actual operation of the machine tool, and a fault prevention control plan is obtained.

[0034] The above describes the calibration method of the precision manufacturing multi-axis linkage machine tool in the embodiment of the present invention. The following describes the calibration device of the precision manufacturing multi-axis linkage machine tool in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a precision manufacturing multi-axis linkage machine tool calibration device includes: An acquisition module is used to acquire machine tool operation data and multi-modal sensor data of a multi-axis linkage machine tool, and to perform error prediction on the machine tool operation data through a multi-axis linkage error prediction model to obtain error prediction data; The measurement module is used to measure the position of each axis of the machine tool using an atomic interferometer and a grating ruler, and obtain nanometer-level resolution position data through Kalman filtering; The fusion module is used to input multimodal sensor data and nanometer-resolution position data into a multi-branch deep neural network and fuse them through an attention mechanism to obtain a real-time status representation of the machine tool; A feature processing module is used to construct a multi-level hypergraph based on the real-time state representation of the machine tool, and input the multi-level hypergraph into a hierarchical hypergraph model for feature processing to obtain a fault feature vector; A reinforcement learning module is used to use the fault feature vector and error prediction data as state inputs, perform reinforcement learning through a deep Q network, and obtain an adaptive compensation strategy; The generation module is used to generate the linkage compensation value of each axis based on the adaptive compensation strategy, and to perform real-time adjustment and fault prevention analysis on the multi-axis linkage machine tool to generate a fault prevention control plan.

[0035] Through the collaborative cooperation of the above-mentioned components, through the fusion of multimodal sensor data and nanometer-level resolution position data, combined with the multi-head self-attention mechanism, high-precision perception and feature extraction of machine tool status are achieved, and the accuracy and real-time performance of status monitoring are improved. The multi-level hypergraph structure is used to model the physical layer, motion layer and functional layer of the machine tool, and the feature processing is performed through the layered hypergraph model, which effectively captures the correlation between different levels and improves the expression ability of fault characteristics. The reinforcement learning method based on the deep Q network realizes the dynamic optimization of the adaptive compensation strategy, so that the compensation value can be adaptively adjusted according to the real-time status of the machine tool, which significantly improves the compensation effect. Through the digital twin technology to build a virtual model, the fault prevention plan is optimized in the virtual environment, and the early warning and active prevention of faults are realized, which reduces the equipment maintenance cost and prolongs the service life of the machine tool. The present invention improves the precision maintenance ability and operation reliability of multi-axis linkage machine tools, and can be applied to different types of multi-axis linkage machine tools.

[0036] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0037] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0038] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0039] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0040] 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.

[0041] 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.

[0042] 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 precision manufacturing multi-axis linkage machine tool calibration method, characterized in that: The method comprises: Collecting machine tool operation data and multi-modal sensor data of a multi-axis linkage machine tool, and performing error prediction on the machine tool operation data through a multi-axis linkage error prediction model to obtain error prediction data; The position of each axis of the machine tool is measured using an atomic interferometer and a grating ruler, and the position data with nanometer resolution is obtained through Kalman filtering. Inputting the multimodal sensor data and the nanometer-level resolution position data into a multi-branch deep neural network, fusing them through an attention mechanism, and obtaining a real-time state representation of the machine tool; Constructing a multi-level hypergraph based on the real-time state representation of the machine tool, and inputting the multi-level hypergraph into a hierarchical hypergraph model for feature processing to obtain a fault feature vector; Taking the fault feature vector and the error prediction data as state input, performing reinforcement learning through a deep Q network to obtain an adaptive compensation strategy; Based on the adaptive compensation strategy, the linkage compensation value of each axis is generated, and the multi-axis linkage machine tool is adjusted in real time and fault prevention analysis is performed to generate a fault prevention control plan.

2. The precision manufacturing multi-axis linkage machine tool calibration method according to claim 1, characterized in that: The collecting of machine tool operation data and multi-modal sensor data of the multi-axis linkage machine tool, and performing error prediction on the machine tool operation data through a multi-axis linkage error prediction model to obtain error prediction data, includes: The real-time position, speed and acceleration of each axis of the multi-axis linkage machine tool are collected to obtain the machine tool operation data. The temperature distribution, vibration characteristics and stress state of the machine tool are collected through the temperature sensor array, three-axis acceleration sensor and strain gauge to obtain multi-modal sensor data. Performing time synchronization on the machine tool operation data and the multimodal sensor data, segmenting the data using a sliding window method, with a window size of 1 second and an overlap rate of 50%, and then standardizing the data in each window to obtain first standardized input data; Input the first standardized input data into the first input layer of the multi-axis linkage error prediction model for one-dimensional convolution processing to obtain a first convolution feature vector, wherein the first input layer includes N parallel one-dimensional convolutional neural networks, where N is the number of machine tool axes, and each one-dimensional convolutional neural network includes 3 convolutional layers, with convolution kernel sizes of 5, 3, and 3, respectively, and the number of convolution kernels of 32, 64, and 128, respectively, and each convolution layer is followed by BatchNormalization and ReLU activation functions; Extract features from the first convolution feature vector, perform cross-axis feature fusion through two three-dimensional convolution layers, the convolution kernel size is 3x3x3, the number of convolution kernels is 256 and 512 respectively, and each layer is followed by MaxPooling and Dropout operations to obtain a second convolution feature vector; The second convolution feature vector is input into a multi-head self-attention mechanism layer, with 8 heads and 64 dimensions of each head. The correlation weights between different axes are calculated by scaling the dot product attention and the features are reorganized to obtain a weighted convolution feature vector. The time series features and long-term dependencies of the weighted convolution feature vector are captured to obtain a time series feature vector, and the time series feature vector is input into the first fully connected layer. The first fully connected layer includes 3 hidden layers with 256, 128, and 64 neurons respectively. The last layer uses a linear activation function with an output dimension of N, corresponding to the prediction error values ​​of N axes, to obtain error prediction data.

3. The precision manufacturing multi-axis linkage machine tool calibration method according to claim 2, characterized in that: The atomic interferometer and grating ruler are used to measure the position of each axis of the machine tool, and the nanometer-level resolution position data is obtained through Kalman filtering, including: A high-precision grating ruler is installed on each axis of the multi-axis linkage machine tool. The grating ruler resolution is 1nm and the sampling frequency is 10kHz. The position of each axis is preliminarily measured to obtain the grating ruler position data. An atom interferometer system is fixed on the machine tool reference surface, wherein the atom interferometer system includes a frequency-stable laser light source, a beam splitter, a reflector and a photodetector. The measurement resolution of the atom interferometer is 0.1 nm, and the sampling frequency is 100 kHz. The position of each axis relative to the reference surface is accurately measured to obtain the position data of the atom interferometer; Performing time synchronization on the grating ruler position data and the atomic interferometer position data to obtain synchronized position data, and performing wavelet transformation on the synchronized position data to obtain multi-scale position features; A state space model is constructed based on the multi-scale position feature, where the state variables include position, velocity and acceleration, and the observation variable is the position measurement value, to obtain a linear state space representation; Creating a Kalman filter according to the linear state space representation, wherein the Kalman filter includes a state prediction equation and a measurement update equation, initializing a state estimation covariance matrix P and a process noise covariance matrix Q, and obtaining Kalman filter parameters; The synchronized position data is input into the Kalman filter, and state estimation and error covariance update are performed through recursive calculation to obtain a filtered position estimate value, and an inverse wavelet transform is performed on the filtered position estimate value to reconstruct the nanometer-level resolution position data.

4. The precision manufacturing multi-axis linkage machine tool calibration method according to claim 3 is characterized in that: The step of inputting the multimodal sensor data and the nanometer-level resolution position data into a multi-branch deep neural network and fusing them through an attention mechanism to obtain a real-time state representation of the machine tool includes: Performing time alignment and normalization processing on the multimodal sensor data and the nanometer-level resolution position data to obtain second standardized input data; The second standardized input data is divided into four branches: temperature data, vibration data, stress data, and position data, and input into four parallel feature extraction networks respectively. Each feature extraction network contains three one-dimensional convolutional layers, and the convolution kernel sizes are 5, 3, and 3 respectively. The number of convolution kernels is 32, 64, and 128 respectively. Each layer is followed by BatchNormalization and ReLU activation functions to obtain the primary feature vector of each branch. Perform global average pooling and global maximum pooling operations on the primary feature vector of each branch to obtain the pooled feature vector of each branch, and concatenate the pooled feature vectors of each branch to obtain the global feature vector; Based on the global feature vector, the correlation weights between different modalities are calculated by a scaled dot product attention mechanism to obtain an attention weighted feature, and temporal feature extraction is performed on the attention weighted feature to obtain a temporal context vector; Inputting the temporal context vector into a cross-modal transformer encoder, the cross-modal transformer encoder comprising 6 encoder layers, each layer comprising a multi-head self-attention sublayer and a feed-forward neural network sublayer, obtaining a fused feature representation, performing residual connection and layer normalization on the fused feature representation, and performing feature mapping through a nonlinear activation function ReLU to obtain an intermediate feature representation; The intermediate feature representation is input into the second fully connected layer, which includes three hidden layers with 512, 256 and 128 neurons respectively. The last layer uses the Softmax activation function to output the real-time status representation of the machine tool.

5. The precision manufacturing multi-axis linkage machine tool calibration method according to claim 4, characterized in that: The method of constructing a multi-level hypergraph based on the real-time state representation of the machine tool and inputting the multi-level hypergraph into a hierarchical hypergraph model for feature processing to obtain a fault feature vector includes: The real-time state representation of the machine tool is divided into three levels: a physical layer, a motion layer, and a functional layer, each level includes N nodes, where N is the number of axes of the machine tool, to obtain a first node set; Assigning attributes to the nodes in the first node set, where the physical layer node attributes include temperature, vibration and stress, the motion layer node attributes include position, velocity and acceleration, and the functional layer node attributes include processing accuracy and motion synchronization, to obtain a second node set; Building a multi-level hypergraph structure based on the second node set, calculating the similarity between nodes by using a K-nearest neighbor algorithm, setting a similarity threshold to 0.8, establishing hyperedge connections between nodes with similarities greater than the threshold, and obtaining an initial hypergraph structure; Calculating the intra-layer edge weights of the initial hypergraph structure, using cosine similarity to calculate the correlation between node attributes, and using the correlation as the edge weight to obtain a weighted intra-layer hypergraph; Calculating the inter-layer edge weights of the weighted intra-layer hypergraph, using the Pearson correlation coefficient to calculate the correlation between the attributes of nodes in different layers, and using the correlation as the inter-layer edge weight to obtain the target multi-level hypergraph; Input the target multi-level hypergraph into a hierarchical hypergraph model, wherein the hierarchical hypergraph model comprises three hypergraph convolutional layers, the number of output channels of each hypergraph convolutional layer is 64, 128 and 256 respectively, and the activation function is LeakyReLU, to obtain hierarchical features; The hierarchical features are fused across layers to obtain cross-layer fusion features, and the cross-layer fusion features are input into the third fully connected layer, wherein the third fully connected layer includes two hidden layers with 128 and 64 neurons respectively. The last layer uses a Tanh activation function with an output dimension of 32 to obtain a fault feature vector.

6. The precision manufacturing multi-axis linkage machine tool calibration method according to claim 5, characterized in that: The method uses the fault feature vector and the error prediction data as state inputs, performs reinforcement learning through a deep Q network, and obtains an adaptive compensation strategy, including: The fault feature vector and the error prediction data are concatenated to obtain a state vector, and the state vector is normalized to obtain a standardized state input; Input the standardized state into the second input layer of the deep Q network, where the second input layer is a fully connected layer with 256 neurons and a ReLU activation function to obtain an initial feature representation; Extracting features from the initial feature representation, performing deep feature learning through three residual blocks, each residual block comprising two convolutional layers and a short-circuit connection, a convolution kernel size of 3x3, and output channel numbers of 128, 256, and 512, respectively, to obtain deep features; Input the deep features into the attention layer, use the self-attention mechanism to calculate the correlation weights between the features to obtain attention-weighted features, and perform time series modeling on the attention-weighted features to obtain time series context features; The temporal context features are input into the advantage function estimator, which includes two fully connected layers with 128 and 64 neurons respectively. The last layer uses a linear activation function and outputs a Q value to obtain an action value estimate; Based on the action value estimation, the ε-greedy strategy is adopted to select the compensation action. The ε value is initially set to 0.9 and gradually reduced according to the exponential decay rule to obtain a compensation action sequence. The compensation action sequence is then empirically replayed, batch experience samples are randomly sampled, and the Q network parameters are updated through temporal difference learning to obtain an adaptive compensation strategy.

7. The precision manufacturing multi-axis linkage machine tool calibration method according to claim 6, characterized in that: The method of generating the linkage compensation value of each axis based on the adaptive compensation strategy, performing real-time adjustment and fault prevention analysis on the multi-axis linkage machine tool, and generating a fault prevention control plan includes: The adaptive compensation strategy is input into the compensation value generator, the initial compensation value of each axis is obtained by inverse kinematics calculation, and the initial compensation value of each axis is nonlinearly optimized to obtain the linkage compensation value of each axis; Based on the linkage compensation values ​​of each axis, the multi-axis linkage machine tool is adjusted in real time, and the adjusted machine tool operation data is collected, and the adjusted machine tool operation data is fast Fourier transformed to obtain frequency domain features; The frequency domain features are input into a multi-class support vector machine for fault prediction. The radial basis function is selected as the kernel function to obtain the probability distribution of the fault type, and a threshold judgment is performed on the probability distribution of the fault type. The fault warning threshold is set to 0.

8. When the probability of a certain type of fault exceeds the threshold, a fault warning signal is triggered; Based on the fault warning signal, a fault tracing analysis program is started, a fault cause-effect relationship diagram is constructed, the root cause of the fault is located through probabilistic reasoning, and a fault tracing result is obtained. Based on the fault tracing result, a machine tool virtual model is constructed using digital twin technology, and a fault prevention plan is executed and optimized in a virtual environment to obtain a fault prevention control plan.

8. A precision manufacturing multi-axis linkage machine tool calibration device, characterized in that: The device is used to perform the precision manufacturing multi-axis linkage machine tool calibration method according to any one of claims 1 to 7, and comprises: An acquisition module is used to acquire machine tool operation data and multi-modal sensor data of a multi-axis linkage machine tool, and to perform error prediction on the machine tool operation data through a multi-axis linkage error prediction model to obtain error prediction data; The measurement module is used to measure the position of each axis of the machine tool using an atomic interferometer and a grating ruler, and obtain nanometer-level resolution position data through Kalman filtering; A fusion module, used for inputting the multimodal sensor data and the nanometer-level resolution position data into a multi-branch deep neural network, and fusing them through an attention mechanism to obtain a real-time state representation of the machine tool; A feature processing module, used for constructing a multi-level hypergraph based on the real-time state representation of the machine tool, and inputting the multi-level hypergraph into a hierarchical hypergraph model for feature processing to obtain a fault feature vector; A reinforcement learning module, used for taking the fault feature vector and the error prediction data as state inputs, performing reinforcement learning through a deep Q network, and obtaining an adaptive compensation strategy; A generation module is used to generate the linkage compensation value of each axis based on the adaptive compensation strategy, and to perform real-time adjustment and fault prevention analysis on the multi-axis linkage machine tool to generate a fault prevention control plan.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the precision manufacturing multi-axis linkage machine tool calibration method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the precision manufacturing multi-axis linkage machine tool calibration method according to any one of claims 1 to 7.

Images