Intelligent CNC machining system and machining method thereof

Abstract

The invention relates to the technical field of CNC machining control, and discloses an intelligent CNC machining system and method, and the method comprises the steps: carrying out the comprehensive modeling of the machining quality, efficiency and energy consumption of CNC, and obtaining a machining quality-efficiency-energy consumption function model; obtaining the multi-dimensional time sequence state data set for feature extraction to obtain a feature data set; constructing a processing state real-time identification model; carrying out five-axis linkage CNC machining simulation to obtain a machining simulation state data set, and carrying out machining state classification identification through the machining state real-time identification model to obtain a machining state simulation identification result; and performing deep reinforcement learning, generating a machining parameter dynamic adjustment strategy and performing global parameter optimization to obtain a target CNC machining control parameter set, and the method can adaptively optimize the machining parameters to cope with different machining tasks and environment changes, thereby improving the machining precision, efficiency and energy utilization rate.

Description

Technical Field

[0001] The present invention relates to the technical field of CNC machining control, and in particular to an intelligent CNC machining system and a machining method thereof. Background Art

[0002] With the rapid development of intelligent manufacturing technology, CNC machining plays an increasingly important role in modern industrial production. Traditional CNC machining methods mainly rely on manual experience and fixed machining parameters, which are difficult to adapt to complex and changing machining environments and ever-increasing machining requirements. In actual production, there is often a contradiction between machining quality, efficiency and energy consumption. How to improve machining efficiency and reduce energy consumption while ensuring machining quality has become the main challenge facing the current CNC machining field.

[0003] In addition, various conditions during CNC machining, such as tool wear, chipping and chatter, can have a significant impact on machining results. However, traditional machining methods lack the ability to monitor and identify these conditions in real time, and cannot adjust machining parameters in time to cope with state changes, resulting in unstable machining quality, low efficiency and energy waste. Summary of the invention

[0004] The present invention provides an intelligent CNC machining system and a machining method thereof, which can adaptively optimize machining parameters to cope with different machining tasks and environmental changes, thereby improving machining accuracy, efficiency and energy utilization.

[0005] In a first aspect, the present invention provides an intelligent CNC machining method, the intelligent CNC machining method comprising: Comprehensively model the machining quality, efficiency and energy consumption of CNC to obtain the machining quality-efficiency-energy consumption function model; Based on the processing quality-efficiency-energy consumption function model, a multi-sensor data acquisition scheme is created to obtain a multi-dimensional time series state data set in the CNC processing process, and time-frequency domain feature extraction and noise reduction processing are performed on the multi-dimensional time series state data set to obtain a feature data set; Training a multi-branch hybrid deep neural network based on the feature data set to build a real-time recognition model for processing status; Based on a preset initial CNC machining control parameter set, a five-axis linkage CNC machining simulation is performed to obtain a machining simulation state data set, and the machining simulation state data set is input into the machining state real-time recognition model to perform machining state classification recognition to obtain a machining state simulation recognition result; Based on the machining state simulation recognition result, deep reinforcement learning is performed to generate a machining parameter dynamic adjustment strategy, and the initial CNC machining control parameter set is globally optimized according to the machining parameter dynamic adjustment strategy to obtain a target CNC machining control parameter set.

[0006] In a second aspect, the present invention provides an intelligent CNC machining system, the intelligent CNC machining system comprising: Modeling module, used to comprehensively model the processing quality, efficiency and energy consumption of CNC, and obtain the processing quality-efficiency-energy consumption function model; An extraction module is used to create a multi-sensor data acquisition scheme based on the processing quality-efficiency-energy consumption function model and obtain a multi-dimensional time series state data set in the CNC processing process, perform time-frequency domain feature extraction and noise reduction processing on the multi-dimensional time series state data set to obtain a feature data set; A training module, used for training a multi-branch hybrid deep neural network according to the feature data set to construct a real-time recognition model for processing status; An identification module is used to perform five-axis linkage CNC machining simulation based on a preset initial CNC machining control parameter set to obtain a machining simulation state data set, and input the machining simulation state data set into the machining state real-time identification model to perform machining state classification identification to obtain a machining state simulation identification result; The optimization module is used to perform deep reinforcement learning based on the machining state simulation recognition result, generate a dynamic adjustment strategy for machining parameters, and perform global parameter optimization on the initial CNC machining control parameter set according to the dynamic adjustment strategy for machining parameters to obtain a target CNC machining control parameter set.

[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 intelligent CNC machining 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 intelligent CNC machining method.

[0009] In the technical solution provided by the present invention, by comprehensively modeling the CNC processing quality, efficiency and energy consumption, a processing quality-efficiency-energy consumption function model is obtained, and the quantitative expression of the multi-objective optimization problem is realized. A multi-sensor data acquisition scheme is adopted to obtain the multi-dimensional time series state data in the CNC processing process, and the time-frequency domain feature extraction and noise reduction processing are performed. The perception ability of the processing state is improved, and a real-time recognition model of the processing state is constructed by training a multi-branch hybrid deep neural network. This network structure can make full use of the time domain, frequency domain and time-frequency domain features to improve the accuracy and robustness of the processing state recognition. Five-axis linkage CNC processing simulation is performed based on the initial processing parameters, and the simulation data is input into the state recognition model to realize the prediction and evaluation of the processing process. The risk of actual processing is reduced and the efficiency of parameter optimization is improved. A deep reinforcement learning method is used to generate a dynamic adjustment strategy for processing parameters, and adaptive optimization of processing parameters is realized. This method can adjust parameters according to the real-time processing state, which improves the intelligence level and adaptability of the processing process. The initial CNC processing control parameter set is globally optimized based on the dynamic adjustment strategy to obtain the target parameter set. This optimization method takes the entire machining process into consideration and can achieve the overall optimization of machining quality, efficiency and energy consumption. A dynamic weight adjustment mechanism is introduced, an adaptive QEE function is constructed, and sensitivity analysis and optimization are performed. This method improves the model's adaptability to different machining requirements. By constructing a multi-scale feature fusion matrix and performing feature dimensionality reduction, a more representative feature data set is extracted. The efficiency and accuracy of state recognition are improved. 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 the intelligent CNC machining method in an embodiment of the present invention; Figure 2 It is a structural schematic diagram of an intelligent CNC machining system in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The embodiment of the present invention provides an intelligent CNC machining system and a machining method thereof. 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 intelligent CNC processing method in the embodiment of the present invention includes: Step S1, comprehensively modeling the processing quality, efficiency and energy consumption of CNC to obtain a processing quality-efficiency-energy consumption function model; It is understandable that the execution subject of the present invention may be an intelligent CNC processing system, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0014] Specifically, the machining quality index is quantified to construct a machining quality sub-model. This sub-model covers three key quality indexes: surface roughness, dimensional accuracy, and shape accuracy. By assigning weights to each index and summing their numerical values, a comprehensive machining quality score is obtained. The machining efficiency index is quantified to construct a machining efficiency sub-model. This sub-model mainly includes two key factors: material removal rate and machining time. By establishing a nonlinear relationship function between them, these indexes are converted into a machining efficiency score to quantify the efficiency performance in the machining process. At the same time, the energy consumption index is quantified to obtain an energy consumption sub-model. The energy consumption sub-model mainly considers factors such as spindle power, feed power, and power consumption of auxiliary equipment, and the total energy consumption in the machining process is obtained by integrating the power consumption index to form an energy consumption score. Based on the machining quality sub-model, the machining efficiency sub-model, and the energy consumption sub-model, a multi-objective optimization problem is constructed to obtain the initial QEE function (i.e., quality-efficiency-energy consumption function). The initial QEE function is normalized to obtain a normalized QEE function to ensure that the results of each sub-model are comparable under a unified standard. In order to improve the adaptability and flexibility of the QEE function, a dynamic weight adjustment mechanism is introduced to obtain an adaptive QEE function. This mechanism can adjust the weights of each target according to the real-time processing situation to cope with the changes in demand under different processing conditions, so that the model can better adapt to the actual production environment. A sensitivity analysis is performed on the adaptive QEE function to determine the influence of each parameter on the QEE function. This analysis process helps to identify the parameters that have the greatest impact on the QEE function. Based on the influence of key parameters, the adaptive QEE function is optimized to obtain the optimized processing quality-efficiency-energy consumption function model.

[0015] Step S2, creating a multi-sensor data acquisition scheme based on the processing quality-efficiency-energy consumption function model and obtaining a multi-dimensional time series state data set in the CNC processing process, performing time-frequency domain feature extraction and noise reduction processing on the multi-dimensional time series state data set to obtain a feature data set; Specifically, the target parameters of the CNC machining process are analyzed based on the machining quality-efficiency-energy consumption function model to determine which parameters have the greatest impact on quality, efficiency and energy consumption during the machining process, and a sensor layout plan is formulated accordingly. The plan includes selecting appropriate sensor types and determining their optimal installation positions and sampling frequencies. Spindle current sensors, feed axis current sensors, vibration sensors, temperature sensors and acoustic emission sensors are arranged at key positions of CNC machining equipment to ensure that various state data during the machining process can be monitored in real time and accurately. After the sensor layout plan is determined, real-time data collection of the CNC machining process is carried out according to the plan to generate a multi-dimensional time series state data set. The data set covers a variety of important parameters such as spindle current, feed axis current, vibration, temperature and acoustic emission signals, reflecting the dynamic changes in the CNC machining process. The multi-dimensional time series state data set is analyzed in the time domain, and the time domain features are extracted and the time domain feature set is constructed. The time domain features include mean, variance, peak factor, skewness and kurtosis, which can effectively reflect the temporal change law and fluctuation of the signal. At the same time, based on the fast Fourier transform algorithm, the multidimensional time series state data set is analyzed in the frequency domain, and the frequency domain features are extracted to form a frequency domain feature set. The frequency domain features include spectrum energy, main frequency and frequency band energy ratio, etc. These features can reveal the distribution of the signal in the frequency domain and help identify specific frequency components and harmonic features in the processing process. The wavelet packet transform algorithm is used to perform time-frequency domain analysis on the original multidimensional time series state data set, extract time-frequency domain features from it, and generate a time-frequency domain feature set. Time-frequency domain features such as wavelet packet energy entropy and wavelet packet energy ratio can capture the energy distribution of the signal in different frequency bands and its changes over time. Based on the time domain feature set, frequency domain feature set and time-frequency domain feature set, a multi-scale feature fusion matrix is ​​constructed. The multi-scale feature fusion matrix is ​​subjected to feature dimensionality reduction processing to improve the computational efficiency of the features and reduce redundant information, thereby obtaining a streamlined and efficient feature data set.

[0016] Step S3: training a multi-branch hybrid deep neural network according to the feature data set to construct a real-time recognition model for processing status; Specifically, the feature data set is segmented to obtain time domain feature subsets, frequency domain feature subsets, and time-frequency domain feature subsets. These subsets represent information of different dimensions, and each subset is used to construct a specific neural network branch to maximize the extraction and utilization of information of different signal features. Based on the time domain feature subset, a time domain branch network is constructed. The branch network adopts a structure of three one-dimensional convolutional layers and two fully connected layers. The convolution layer is used to extract local features of the time domain signal. The convolution kernel size of the first one-dimensional convolutional layer is set to 5 and the stride is 2 to capture features in a larger range; the convolution kernel size of the second and third convolutional layers is 3 and the stride is 1 to capture finer-grained features. Each one-dimensional convolutional layer is followed by a batch normalization layer and a ReLU activation function to ensure the stability of training and nonlinear expression ability. The number of neurons in the two fully connected layers is 128 and 64 respectively. The fully connected layer further compresses and abstracts the features extracted by the convolution layer. At the same time, based on the frequency domain feature subset, a frequency domain branch network is constructed. The branch network adopts the ResNet structure to handle the complexity and diversity of frequency domain features. The ResNet network solves the problem of gradient vanishing in deep networks by introducing residual blocks. The frequency domain branch network includes an input convolution layer, three residual blocks, and a global average pooling layer. Each residual block contains two two-dimensional convolution layers and a short-circuit connection. This design allows the network to learn more representative features at a deeper level. The convolution kernel size of the convolution layer is 3x3, the stride is 1, and a batch normalization layer and a ReLU activation function are added after each convolution layer to ensure the stability and nonlinear expression ability of the network. For the time-frequency domain feature subset, an attention-enhanced LSTM structure is used to construct the time-frequency domain branch network. The time-frequency domain branch network processes the time-frequency domain features through a bidirectional LSTM layer and two self-attention layers. The number of hidden units in the bidirectional LSTM layer is set to 128, which can effectively capture the front-end relationship of the temporal information, and through the enhancement of the self-attention layer, the model can pay more attention to important feature information. The number of heads of the self-attention layer is set to 4 to enhance the selective attention to multi-dimensional features. The output features of the time domain branch network, the frequency domain branch network, and the time-frequency domain branch network are concatenated to obtain a fused feature vector. A feature selection algorithm is applied to select the most important features. The feature importance ranking method based on XGBoost is used to screen and retain the features that have the greatest impact on the classification results, and the optimized feature vector is obtained. A fully connected classification network is constructed based on the optimized feature vector. The fully connected classification network includes three fully connected layers with 256, 128, and 64 neurons, respectively. A Dropout layer and a ReLU activation function are added after each fully connected layer to enhance the generalization ability and nonlinear expression ability of the network. The last layer uses the Softmax function to output the classification probability of the processing state, ensuring that the model can accurately classify the input features into different processing states.The time domain branch network, frequency domain branch network, time-frequency domain branch network and fully connected classification network are connected in series to obtain a multi-branch hybrid deep neural network. The network realizes real-time recognition of processing status through parallel processing and comprehensive analysis of multiple feature spaces. The entire network uses the cross entropy loss function as the objective function and uses the Adam optimizer for end-to-end training. By continuously adjusting the network parameters, an accurate and robust real-time recognition model for processing status is obtained, which can monitor and identify various states in real time during CNC processing.

[0017] Step S4, performing five-axis linkage CNC machining simulation based on a preset initial CNC machining control parameter set to obtain a machining simulation state data set, and inputting the machining simulation state data set into a machining state real-time recognition model to perform machining state classification recognition to obtain a machining state simulation recognition result; Specifically, the preset initial CNC machining control parameter set is parameter-encoded, and each control parameter is converted into a unified encoding vector, including key factors such as spindle speed, feed speed, cutting depth, tool path, and coolant flow. Based on the parameter encoding vector, a simulation model for five-axis linkage CNC machining is constructed. The simulation model uses the finite element analysis method to accurately simulate key physical phenomena such as cutting force, heat distribution, and workpiece deformation during the machining process. In this way, various physical effects that may occur in real machining are reproduced in a virtual environment. The simulation model is time-discretized. By decomposing the entire machining process into a series of discrete time points, simulation calculations are performed at each time point to generate a machining simulation state data set. The data set covers a variety of key state variables including simulated spindle current, feed axis current, vibration, temperature, and acoustic emission signals, reflecting the dynamic changes of the machining process. Feature extraction is performed on the machining simulation state data set, and the original simulation data is converted into a simulation feature data set. The simulation feature data sets are sequentially input into each branch network of the machining state real-time recognition model for processing. The simulation feature dataset is input into the time domain branch network, which includes 3 one-dimensional convolutional layers and 2 fully connected layers. The time domain features are extracted through convolution operation, and these features are compressed and abstracted through the fully connected layer to obtain the time domain feature vector. The same simulation feature dataset is input into the frequency domain branch network. The network structure includes 1 input convolutional layer, 3 residual blocks and 1 global average pooling layer. Through these layer processing, the frequency domain features are extracted and integrated to generate the frequency domain feature vector. The simulation feature dataset is input into the time-frequency domain branch network, which consists of 1 bidirectional LSTM layer and 2 self-attention layers. The LSTM layer captures the timing information, and the self-attention layer is used to enhance the expression of key features to obtain the time-frequency domain feature vector. The time domain feature vector, frequency domain feature vector and time-frequency domain feature vector are concatenated and input into the fully connected classification network. The fully connected classification network consists of 3 fully connected layers, and each fully connected layer is connected to the Dropout layer and ReLU activation function to improve the generalization ability and nonlinear feature capture ability of the model. After being processed by the Softmax function, the machining state simulation identification results are obtained, which provide the probability distribution of the machining state, including the probabilities of normal machining, tool wear, chipping and chattering states.

[0018] Data preprocessing is performed on the frequency domain features in the simulation feature data set. This includes normalization and standardization operations on the data. Normalization scales the data to the same scale range, while standardization adjusts the mean and variance of the data so that the distribution of the feature data is more suitable for the processing of the neural network, and the preprocessed frequency domain feature matrix is ​​obtained. The preprocessed frequency domain feature matrix is ​​input into the input convolution layer, which uses 64 convolution kernels of size 3x3 for convolution operation. The convolution operation can effectively extract the local features of the data. The result of the convolution operation is 64 feature maps. These feature maps are further processed and batch normalized. The training process is accelerated by adjusting the distribution of the feature maps. Then the ReLU activation function is applied. The ReLU activation function introduces nonlinear characteristics and enhances the model's ability to capture complex features to obtain the initial feature map set. The initial feature map set is input into the first residual block. The first residual block contains two 3x3 convolution layers and a short-circuit connection. The convolutional layer extracts deeper features, and each convolutional layer is followed by batch normalization and ReLU activation functions to ensure the stability and nonlinear expression ability of the network. Short-circuit connection introduces residual learning mechanism, which effectively alleviates the gradient vanishing problem that may occur during the training of deep neural networks by directly adding the input signal to the output, making it easier to train deep networks. After being processed by the first residual block, the first layer of residual feature atlas is obtained. The first layer of residual feature atlas is input to the second residual block. The structure of the second residual block is the same as the first one, but the number of convolution kernels is increased to 128, allowing the model to learn more complex and more representative features. In this process, through the residual learning mechanism, the model can transfer information more effectively and obtain the second layer of residual feature atlas. The second layer of residual feature atlas is input to the third residual block. The structure of the third residual block is the same as the previous residual block, and the number of convolution kernels is increased to 256, so that the model can capture richer and more comprehensive feature information in higher dimensions. The third-layer residual feature atlas is obtained through residual learning. The feature atlas condenses a variety of frequency domain features from low-level to high-level. The third-layer residual feature atlas is subjected to global average pooling. Global average pooling compresses the spatial information of all feature maps into a 256-dimensional feature vector by taking the average of each feature map. The 256-dimensional feature vector is L2 normalized to make it unit length to obtain a feature vector with a consistent scale. The unit-length feature vector is input into a fully connected layer, which reduces the dimension of the feature vector to 128, which helps to simplify the model and reduce the computational complexity. At the same time, the ReLU activation function is applied to continue to maintain the nonlinear characteristics of the network and obtain the frequency domain feature vector.

[0019] The time domain feature vector, frequency domain feature vector and time-frequency domain feature vector are concatenated to obtain a fused feature vector, the dimension of which is the sum of the dimensions of the three feature vectors. The fused feature vector is normalized to eliminate the dimensional differences between different features, ensure that all features are compared and processed at the same scale, and obtain a standardized fused feature vector. The standardized fused feature vector is input into the first fully connected layer. The first fully connected layer contains 256 neurons and uses the ReLU activation function for nonlinear transformation. The introduction of the ReLU activation function enables the model to capture the nonlinear relationship in the data and obtain the first layer of feature representation. In order to prevent the overfitting of the model, the first layer of feature representation is Dropout, and the dropout rate is set to 0.5. During the training process, half of the neurons are randomly dropped to force the model to learn a more robust feature representation, and a sparse first layer of feature representation is obtained. Through the sparse processing, the model can better cope with new data and reduce the risk of overfitting. The sparse first layer of feature representation is input into the second fully connected layer. The second fully connected layer contains 128 neurons. The ReLU activation function is also used for nonlinear transformation to further compress and abstract the feature space, extract more representative feature information, and obtain the second-layer feature representation. In order to ensure the consistency of data distribution during the training process of the model, the Batch Normalization operation is applied to the second-layer feature representation. Through Batch Normalization, the distribution of data is standardized to reduce the instability of the model caused by changes in the training batch, and the normalized second-layer feature representation is obtained. The normalized second-layer feature representation is input into the third fully connected layer. This fully connected layer contains 4 neurons, which correspond to the four states of normal processing, tool wear, chipping, and chatter. Through the processing of this layer, the original state score vector is obtained. The score vector contains the original score of each state, which directly reflects the confidence of the model in each state. The Softmax function is applied to the original state score vector. The Softmax function converts the original score into a probability distribution to obtain the processing state simulation recognition result. Through the probability distribution, it is judged whether the processing state belongs to normal processing, tool wear, chipping, or chatter. The probability of each state reflects the confidence of the model in the occurrence of the state.

[0020] Step S5: perform deep reinforcement learning based on the machining state simulation recognition result, generate a dynamic adjustment strategy for machining parameters, and perform global parameter optimization on the initial CNC machining control parameter set according to the dynamic adjustment strategy for machining parameters to obtain a target CNC machining control parameter set.

[0021] Specifically, the probability threshold processing is performed on the machining state simulation recognition results to convert the continuous probability distribution into discrete machining states, such as normal machining, tool wear, chipping and chattering states. The discrete machining states are combined with the initial CNC machining control parameter set to construct the state space of reinforcement learning. The state space defines the environment that the model can observe under different machining states. Based on the reinforcement learning state space and the adjustable range of CNC machining control parameters, the action space is constructed. The action space defines the operations that can be performed in each state, which include the adjustment amount of machining parameters such as spindle speed, feed rate, cutting depth, tool path and coolant flow. By clarifying the adjustment amount, the model can explore the response of the CNC machining process to the changes of these parameters under different operations. The reward function is designed according to the machining quality-efficiency-energy consumption function model. The reward function is used to measure the effect of each action in a specific state to guide the model to continuously optimize the strategy during the training process. In order to ensure that the reward function has a consistent influence between different states and actions, the reward function is normalized to obtain a standardized reward function. The standardization process helps to accelerate the training process and improve the learning efficiency of the model in different scenarios. Based on the constructed reinforcement learning state space, action space and standardized reward function, a deep Q network (DQN) model is constructed. The DQN model is a common structure in reinforcement learning. It estimates the value (i.e., Q value) of each state-action pair through deep learning. The DQN model contains 4 fully connected layers, each of which uses the ReLU activation function to introduce nonlinear characteristics, thereby enhancing the model's adaptability to complex environments. The last fully connected layer outputs the Q value, which reflects the expected cumulative reward for performing a certain action in the current state. In order to train the deep Q network model, experience replay and target network technology are used. Experience replay breaks the correlation of data by storing and randomly extracting past experience data to improve training stability. The target network is used to slow down the update speed of the target value and further stabilize the training process. During the training process, the ε-greedy strategy is used to balance the relationship between exploration and utilization, that is, the model selects random actions for exploration with a certain probability, and selects the current optimal action for utilization at other times. In this way, the model can fully explore the possible strategy space and finally converge to the optimal strategy to obtain a trained deep Q network model. After the model training is completed, the discretized machining state is evaluated based on the trained deep Q network model. By comparing the Q values ​​of possible actions under different states, the action corresponding to the maximum Q value is selected as the optimal operation, and the dynamic adjustment strategy of machining parameters is obtained. This strategy clarifies how to adjust CNC machining parameters under different machining states to achieve the optimization goal. According to the generated dynamic adjustment strategy of machining parameters, the initial CNC machining control parameter set is globally optimized. By adjusting and optimizing these parameters, the target CNC machining control parameter set is obtained.

[0022] The discretized machining state and the initial CNC machining control parameter set are encoded, and different machining states and control parameters are converted into state vectors that can be input into the deep Q network model, and the complex physical state and control strategy are expressed in vector form. The state vector is input into the trained deep Q network model, and the Q value of each possible action is calculated by forward propagation. The forward propagation process refers to the step-by-step calculation of the corresponding output results, namely Q values, through each layer of the network. These Q values ​​reflect the expected cumulative rewards that can be obtained by executing each possible action in the current state. These Q values ​​are combined into a Q value vector to provide an evaluation basis for each possible action. The ε-greedy strategy is applied to the Q value vector, and the action with the largest Q value is selected with a probability of 1-ε, and the action is randomly selected with a probability of ε to obtain the candidate action. The core of the ε-greedy strategy is to balance exploration and utilization: most of the time (i.e., with a probability of 1-ε), the algorithm will select the currently known best action (i.e., the action with the largest Q value), but it will also leave a certain amount of exploration space (i.e., with a probability of ε) to randomly select other actions to ensure that the model has the opportunity to discover potential better strategies. Based on the candidate actions, the initial CNC machining control parameter set is fine-tuned. According to the candidate actions, the initial parameters are slightly adjusted. On the basis of maintaining the existing processing performance, certain specific processing indicators are optimized to obtain the adjusted processing control parameters. The adjusted processing control parameters are input into the processing quality-efficiency-energy consumption function model to calculate the expected performance indicators. The model comprehensively considers the performance of multiple aspects such as processing quality, efficiency and energy consumption. By simulating the adjusted parameters, the processing effect and resource consumption that these adjustments may bring are predicted, and a quantitative expected performance indicator is provided. According to the comparison results between the expected performance indicator and the current performance indicator, it is determined whether to accept the candidate action. If the expected performance indicator is significantly improved compared with the current indicator, it indicates that the adjustment is effective. At this time, the candidate action is accepted and output as the final dynamic adjustment strategy for processing parameters. Otherwise, if the expected performance fails to meet expectations, other actions are reselected and the process is repeated until the optimal dynamic adjustment strategy for processing parameters is found.

[0023] In the embodiment of the present invention, by comprehensively modeling the CNC processing quality, efficiency and energy consumption, a processing quality-efficiency-energy consumption function model is obtained, and the quantitative expression of the multi-objective optimization problem is realized. A multi-sensor data acquisition scheme is adopted to obtain multi-dimensional time series state data in the CNC processing process, and time-frequency domain feature extraction and noise reduction processing are performed. The perception ability of the processing state is improved, and a real-time recognition model of the processing state is constructed by training a multi-branch hybrid deep neural network. This network structure can make full use of the time domain, frequency domain and time-frequency domain features to improve the accuracy and robustness of processing state recognition. Five-axis linkage CNC processing simulation is performed based on the initial processing parameters, and the simulation data is input into the state recognition model to realize the prediction and evaluation of the processing process. The risk of actual processing is reduced and the efficiency of parameter optimization is improved. A deep reinforcement learning method is used to generate a dynamic adjustment strategy for processing parameters, and adaptive optimization of processing parameters is realized. This method can adjust parameters according to the real-time processing state, which improves the intelligence level and adaptability of the processing process. The initial CNC processing control parameter set is globally optimized based on the dynamic adjustment strategy to obtain the target parameter set. This optimization method takes the entire machining process into consideration and can achieve the overall optimization of machining quality, efficiency and energy consumption. A dynamic weight adjustment mechanism is introduced, an adaptive QEE function is constructed, and sensitivity analysis and optimization are performed. This method improves the model's adaptability to different machining requirements. By constructing a multi-scale feature fusion matrix and performing feature dimensionality reduction, a more representative feature data set is extracted. The efficiency and accuracy of state recognition are improved.

[0024] In a specific embodiment, the process of executing step S1 may specifically include the following steps: The processing quality index is quantified to obtain the processing quality sub-model, which includes surface roughness, dimensional accuracy and shape accuracy. The processing quality score is obtained by weighted summation. The processing efficiency index is quantified to obtain a processing efficiency sub-model, which includes material removal rate and processing time. The processing efficiency score is obtained by establishing a nonlinear relationship function. The energy consumption index is quantified to obtain an energy consumption sub-model, which includes spindle power, feed power and auxiliary equipment power consumption. The total energy consumption is calculated by integration to obtain an energy consumption score. Based on the processing quality sub-model, processing efficiency sub-model and energy consumption sub-model, a multi-objective optimization problem is constructed to obtain the initial QEE function, and the initial QEE function is normalized to obtain the normalized QEE function. Based on the normalized QEE function, a dynamic weight adjustment mechanism is introduced to obtain an adaptive QEE function. A sensitivity analysis is performed on the adaptive QEE function to obtain the influence of each parameter on the QEE function. The adaptive QEE function is optimized based on the influence to obtain the machining quality-efficiency-energy consumption function model.

[0025] Specifically, the processing quality indexes are quantified to obtain the processing quality sub-model, which contains key quality evaluation indicators, namely surface roughness, dimensional accuracy and shape accuracy. ) reflects the microscopic geometric shape changes on the machined surface and is usually expressed as the average profile deviation; dimensional accuracy ( ) measures the difference between the actual size of the workpiece after processing and the designed size; shape accuracy ( ) evaluates the degree of conformity between the actual shape of the workpiece and the ideal shape. In order to quantify these indicators, a weight is assigned to each indicator to reflect its importance in the overall quality evaluation. Assume that the weights of surface roughness, dimensional accuracy and shape accuracy are , and , then the processing quality score ( ) can be expressed as: ; The selection of weights should be based on the priority setting of specific processing tasks. For example, in high-precision processing, dimensional accuracy may be given a higher weight. For the quantitative processing of processing efficiency indicators, a processing efficiency sub-model is established. The model includes material removal rate ( ) and processing time ) are two core parameters. Material removal rate is the volume of material removed per unit time, which is usually an important indicator for improving machining efficiency. Machining time is the total time required to complete a machining task. Since there may be a nonlinear relationship between these two parameters, a nonlinear function is used to represent the machining efficiency score ( ),For example: ; in, , and are empirical coefficients used to adjust the contribution of different parameters to the efficiency score. By adjusting these coefficients, the efficiency priority in different processing scenarios can be reflected. For example, when processing time is critical, you can increase The value of is used to highlight the influence of time. For the quantitative processing of energy consumption indicators, an energy consumption sub-model is constructed, which includes the spindle power ( )、Feed power( ) and auxiliary equipment power consumption ( ). These power consumption indicators reflect the energy consumption in different links of the processing process. In order to obtain the total energy consumption score ( ), the integral calculation of the entire processing process is performed, which is specifically expressed as: ; in, is the total processing time, , and Respectively represent the spindle power, feed power and auxiliary equipment power consumption that change with time. The total energy consumption of the entire machining process is calculated through the integral formula, and this is used as the basis for energy consumption scoring. Based on the machining quality sub-model, machining efficiency sub-model and energy consumption sub-model, a multi-objective optimization problem is constructed to obtain the initial quality-efficiency-energy consumption (QEE) function. The form of the initial QEE function can be expressed as: ; in, , and is a weight coefficient used to balance the importance of quality, efficiency and energy consumption in optimization. The weight setting should be adjusted according to actual needs. For example, in some high-efficiency scenarios, The value of may be relatively large, and when energy saving is given priority, The value of should be increased. In order to ensure that the indicators of the initial QEE function are compared at the same scale, the initial QEE function is normalized to obtain a normalized QEE function. Normalization can be achieved by scaling the value of each sub-item to the interval [0, 1]. Based on the normalized QEE function, a dynamic weight adjustment mechanism is introduced to obtain an adaptive QEE function. The dynamic weight adjustment mechanism can adjust the weight according to the real-time feedback information during the processing. , and , so that the optimization target can adapt to the actual processing needs. For example, when the processing quality fluctuates, The weight of can be increased dynamically, prompting the system to pay more attention to quality improvement. Similarly, if energy consumption exceeds expectations, The weight of can also be increased accordingly. A sensitivity analysis is performed on the adaptive QEE function to obtain the influence of each parameter on the QEE function. The results of the sensitivity analysis can reveal which parameters contribute most to the optimization target during the machining process. Based on the influence, the adaptive QEE function is optimized to obtain an adjusted and optimized machining quality-efficiency-energy consumption function model.

[0026] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Based on the machining quality-efficiency-energy consumption function model, the target parameters in the CNC machining process are analyzed to obtain the sensor layout scheme, which includes the installation positions and sampling frequencies of the spindle current sensor, feed axis current sensor, vibration sensor, temperature sensor and acoustic emission sensor. According to the sensor arrangement scheme, real-time data collection is performed on the CNC machining process to obtain a multi-dimensional time series state data set, which includes spindle current, feed axis current, vibration, temperature and acoustic emission signals; Perform time domain analysis on the multidimensional time series state data set, extract time domain features, and obtain a time domain feature set. The time domain features include mean, variance, peak factor, skewness, and kurtosis. Based on the fast Fourier transform algorithm, the frequency domain analysis is performed on the multi-dimensional time series state data set to extract the frequency domain features and obtain the frequency domain feature set. The frequency domain features include spectrum energy, main frequency and frequency band energy ratio. The wavelet packet transform algorithm is used to perform time-frequency domain analysis on the original multi-dimensional time series state data set, extract time-frequency domain features, and obtain a time-frequency domain feature set. The time-frequency domain features include wavelet packet energy entropy and wavelet packet energy ratio. Based on the time domain feature set, the frequency domain feature set and the time-frequency domain feature set, a multi-scale feature fusion matrix is ​​constructed, and feature dimension reduction is performed on the multi-scale feature fusion matrix to obtain a feature data set.

[0027] Specifically, based on the machining quality-efficiency-energy consumption function model, the target parameters of the CNC machining process are analyzed to determine the monitoring requirements for the machining process. The model comprehensively considers the three aspects of machining quality, efficiency and energy consumption, and evaluates the performance of the entire machining process from the perspectives of machining effect, production efficiency and energy consumption. Through the analysis of the target parameters, the key factors that need to be monitored during the machining process are determined, such as spindle current, feed shaft current, vibration, temperature and acoustic emission signals. After the analysis is completed, the sensor layout plan is designed, including the installation position and sampling frequency of the spindle current sensor, feed shaft current sensor, vibration sensor, temperature sensor and acoustic emission sensor. For example, the spindle current sensor is usually installed near the spindle motor to monitor the current consumed by the motor during the cutting process. This current is closely related to the cutting force and can therefore reflect the cutting state of the tool. The feed axis current sensor is installed on the feed axis motor to monitor the feed axis current to evaluate the load of the feed process. The vibration sensor can be installed on the spindle end or the workpiece fixture to capture the vibration signal during the machining process. The amplitude and frequency of the vibration can reflect the mechanical stability during the machining process. The temperature sensor is installed in the contact area of ​​the tool or workpiece to monitor the heat generated during the machining process. The temperature change can affect the surface quality of the workpiece and the tool life. The acoustic emission sensor is used to capture the tiny cracks and friction sounds of the material during the machining process. These sounds can be used as signals for early detection of tool wear or breakage. The installation position and sampling frequency of the sensor should be set according to the specific requirements of the machining process to ensure that sufficiently detailed and accurate signal data can be captured. Based on the sensor layout scheme, real-time data acquisition is performed on the CNC machining process to obtain a multi-dimensional time series state data set. The data set includes a variety of real-time collected signals such as spindle current, feed axis current, vibration, temperature and acoustic emission signals, reflecting the physical state and changes during the machining process. Perform time domain analysis on the multidimensional time series state data set, and extract signal features from the perspective of time. Common time domain features include mean, variance, peak factor, skewness, and kurtosis. The mean reflects the average level of the signal, the variance indicates the fluctuation amplitude of the signal, the peak factor (i.e., the ratio of the peak value to the mean) is used to measure the sharpness of the signal, the skewness reflects the symmetry of the signal, and the kurtosis is a measure of the sharpness of the signal. Through the time domain features, the fluctuation characteristics and stability of the signal are preliminarily judged. Based on the fast Fourier transform algorithm, perform frequency domain analysis on the multidimensional time series state data set. Frequency features are extracted by converting the time domain signal into the frequency domain signal. These features include spectral energy, main frequency, and band energy ratio. Spectral energy represents the energy distribution of the signal on each frequency component. The main frequency is the frequency with the most concentrated energy in the spectrum, and the band energy ratio is the ratio of energy in different frequency bands. For example, the fast Fourier transform can convert the signal Converting from time domain to frequency domain is represented as ,in is the frequency, and the spectrum energy can be expressed as: ; By analyzing the frequency domain features, the potential periodicity and vibration patterns in the signal are revealed, which helps to detect abnormal conditions (such as vibration and noise) during the processing process. The wavelet packet transform algorithm is used to perform time-frequency domain analysis on the original multidimensional time series state data set. Wavelet packet transform is a method that can perform signal analysis in time and frequency at the same time. It captures the frequency characteristics in different time periods by decomposing the signal into multiple frequency bands. Time-frequency domain features include wavelet packet energy entropy and wavelet packet energy ratio. These features can effectively describe the energy distribution and complexity of the signal in different frequency bands. Wavelet packet energy entropy measures the distribution complexity of signal energy, while the energy ratio reflects the energy proportion in each frequency band. A multi-scale feature fusion matrix is ​​constructed based on time domain, frequency domain and time-frequency domain features. Feature dimensionality reduction is performed on the multi-scale feature fusion matrix to obtain an optimized feature data set. The dimension of the feature is reduced while ensuring the amount of information to improve computational efficiency and reduce model complexity.

[0028] In a specific embodiment, the process of executing step S3 may specifically include the following steps: The feature data set is segmented to obtain a time domain feature subset, a frequency domain feature subset, and a time-frequency domain feature subset; A time-domain branch network is constructed based on the time-domain feature subset. The time-domain branch network includes three one-dimensional convolutional layers and two fully connected layers. Each one-dimensional convolutional layer is followed by a batch normalization layer and a ReLU activation function. The convolution kernel size of the first one-dimensional convolutional layer is 5 and the stride is 2. The convolution kernel size of the second and third one-dimensional convolutional layers is 3 and the stride is 1. The number of neurons in the two fully connected layers is 128 and 64, respectively. A frequency domain branch network is constructed based on the frequency domain feature subset. The frequency domain branch network adopts the ResNet structure, including 1 input convolution layer, 3 residual blocks and 1 global average pooling layer. Each residual block contains 2 two-dimensional convolution layers and 1 short-circuit connection. The convolution kernel size of the convolution layer is 3x3, the stride is 1, and a batch normalization layer and ReLU activation function are added after each convolution layer. A time-frequency domain branch network is constructed based on the time-frequency domain feature subset. The time-frequency domain branch network adopts an attention-enhanced LSTM structure, including 1 bidirectional LSTM layer and 2 self-attention layers. The number of hidden units in the LSTM layer is 128, and the number of heads in the self-attention layer is 4. The output features of the time domain branch network, the frequency domain branch network and the time-frequency domain branch network are concatenated to obtain a fused feature vector; Apply feature selection algorithm to the fused feature vector, select important features and obtain optimized feature vector. The feature selection algorithm adopts feature importance ranking method based on XGBoost. A fully connected classification network is constructed based on the optimized feature vector. The fully connected classification network includes three fully connected layers with 256, 128 and 64 neurons respectively. Each fully connected layer is followed by a Dropout layer and a ReLU activation function. The last layer uses a Softmax function to output the classification probability of the processing state. The time domain branch network, frequency domain branch network, time-frequency domain branch network and fully connected classification network are combined in series to obtain a multi-branch hybrid deep neural network. The multi-branch hybrid deep neural network is end-to-end trained using the cross entropy loss function and Adam optimizer to obtain a real-time recognition model for processing status.

[0029] Specifically, the feature data set is segmented to obtain a time domain feature subset, a frequency domain feature subset, and a time-frequency domain feature subset. The time domain feature subset includes features directly extracted from the time series signal, such as the mean ,variance , peak, etc. The frequency domain feature subset is obtained by fast Fourier transforming the signal to obtain the frequency components, such as spectrum energy ( )、Main frequency( ) etc. The time-frequency domain feature subset captures the joint characteristics of the signal in time and frequency through methods such as wavelet packet transform, such as wavelet packet energy entropy ( ) and energy ratio ( ). For the time domain feature subset, a time domain branch network is constructed. The network consists of three one-dimensional convolutional layers and two fully connected layers. The convolution operation can extract the local pattern of the signal. Suppose the input signal is , the convolution kernel is , the convolution operation of the first one-dimensional convolutional layer can be expressed as: ; Among them, * represents convolution operation, BN represents batch normalization, is the ReLU activation function, is the bias term. The size of the convolution kernel of the first layer is 5 and the stride is 2. The second and third convolution layers operate similarly, with the size of the convolution kernel being 3 and the stride being 1. The output of the convolution layer is processed by two fully connected layers, outputting 128-dimensional and 64-dimensional feature vectors: ; in, and are the weight matrices of the fully connected layers, and is the bias term. For the frequency domain feature subset, a frequency domain branch network is constructed, using the ResNet structure. ResNet solves the gradient vanishing problem in deep networks through residual learning. The frequency domain features are processed by the input convolution layer, and the convolution operation of the convolution layer can be expressed as: ; in, is the frequency domain feature input. Then it passes through three residual blocks, each of which contains two two-dimensional convolutional layers and a short-circuit connection. The output of the first residual block is: ; in, is the first layer of convolution kernels. The structure of subsequent residual blocks is similar. A fixed-dimensional feature vector is obtained through a global average pooling layer. For the time-frequency domain feature subset, an attention-enhanced LSTM network is used to construct a time-frequency domain branch network. The bidirectional LSTM layer is used to capture the front-end dependency relationship in the time series. The output of the LSTM layer can be expressed as: ; in, For time The hidden state of is the cell state at the previous moment. The self-attention mechanism is used to strengthen the representation of important features. The output of the self-attention layer is: ; in, are query, key, and value matrices respectively, is a scaling factor, and the self-attention mechanism is used to capture global dependencies. After the construction is completed, the output features of the time domain branch network, the frequency domain branch network, and the time-frequency domain branch network are concatenated to obtain a fused feature vector. The feature selection algorithm is applied to the fused feature vector to select important features and obtain an optimized feature vector. The feature importance ranking method based on XGBoost is adopted. The XGBoost model is trained to evaluate the importance of each feature to the classification result, and the most critical features are selected according to the importance score. The objective function of the XGBoost model can be expressed as: ; in, represents the loss function, is the actual label, is the predicted value, is the regularization term of model complexity. By optimizing this objective function, the importance of features is determined and the most useful features are selected. After obtaining the optimized feature vector, a fully connected classification network is constructed. The network consists of three fully connected layers with 256, 128 and 64 neurons respectively. Each fully connected layer is followed by a Dropout layer and a ReLU activation function. The Dropout layer is used to randomly discard some neurons during training to prevent the model from overfitting, while the ReLU activation function provides nonlinear capabilities for the model. The last fully connected layer uses the Softmax function to output the classification probability of the processing state. The calculation formula of the Softmax function is: ; in, For Category The corresponding score, is the total number of categories, Represents input Belongs to category The probability of. The time domain branch network, frequency domain branch network, time-frequency domain branch network and fully connected classification network are connected in series to obtain a multi-branch hybrid deep neural network. This network realizes efficient recognition of CNC machining status by comprehensively utilizing the feature information of time domain, frequency domain and time-frequency domain. In order to train this model, the cross entropy loss function is used as the objective function. The calculation formula of the cross entropy loss is: ; in, is the one-hot encoding of the actual label, is the probability value predicted by the model. Combined with the Adam optimizer for end-to-end training, the update formula of the Adam optimizer is: ; in, Update the value for the parameter, is the learning rate, and are the first-order and second-order momentum correction terms, is a smoothing term. In this way, the Adam optimizer can adaptively adjust the learning rate to ensure the efficiency and convergence of the training. After training, the final processing state real-time recognition model can accurately identify different states in the processing process, providing important support for improving the intelligence and accuracy of CNC processing. For example, when the model detects a specific processing state (such as tool wear), it can issue a warning signal in time, thereby avoiding potential quality problems and improving production efficiency.

[0030] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Parameter encoding is performed on a preset initial CNC machining control parameter set to obtain a parameter encoding vector, wherein the parameter encoding vector includes spindle speed, feed speed, cutting depth, tool path and coolant flow parameters; A five-axis CNC machining simulation model is constructed based on parameter coding vectors. The simulation model uses finite element analysis method to simulate and calculate the cutting force, heat distribution and workpiece deformation during the machining process. The five-axis linkage CNC machining simulation model is processed by time discretization to obtain a discrete time series, and simulation calculations are performed at each discrete time point to obtain a machining simulation state data set, which includes simulated spindle current, feed axis current, vibration, temperature and acoustic emission signals; Feature extraction is performed on the processing simulation state data set to obtain a simulation feature data set, and the simulation feature data set is input into the time domain branch network of the processing state real-time recognition model. After being processed by three one-dimensional convolutional layers and two fully connected layers, a time domain feature vector is obtained. The simulation feature data set is input into the frequency domain branch network of the real-time recognition model of the processing state, and processed by one input convolution layer, three residual blocks and one global average pooling layer to obtain the frequency domain feature vector; The simulation feature data set is input into the time-frequency domain branch network of the real-time recognition model of the processing state, and processed by a bidirectional LSTM layer and two self-attention layers to obtain the time-frequency domain feature vector; The time domain feature vector, frequency domain feature vector and time-frequency domain feature vector are concatenated and input into the fully connected classification network. After being processed by three fully connected layers and the Softmax function, the machining state simulation recognition result is obtained. The machining state simulation recognition result includes the probability distribution of normal machining, tool wear, chipping and chattering states.

[0031] Specifically, the preset initial CNC machining control parameter set is parameter-encoded, and each parameter is mapped to a numerical vector to form a parameter encoding vector. For example, assuming the spindle speed is , the feed speed is , cutting depth is , the tool path is , the coolant flow rate is , then the parameter encoding vector can be expressed as: ; Based on the parameter encoding vector, a five-axis CNC machining simulation model is constructed. The simulation model uses the finite element analysis (FEA) method to simulate the physical phenomena in the machining process, such as cutting force, heat distribution, and workpiece deformation. Finite element analysis simulates the complex physical processes that may occur during the machining process by dividing the workpiece and tool into a finite number of elements and calculating the physical behavior of these elements. The cutting force can be expressed as: ; in, is the cutting force, is the cutting force coefficient, is the cutting area. The heat distribution can be described by the heat conduction equation: ; in, is the material density, is the specific heat capacity, is the temperature, is the thermal conductivity, is the heat source term. Through finite element simulation, the temperature distribution and stress-strain distribution inside the workpiece at each time point are obtained, and then the deformation of the workpiece during the processing is predicted. The simulation model is time discretized, and the continuous processing process is divided into multiple discrete time points so that simulation calculations can be performed at each time point. The discretized time series can be expressed as: ; At each discrete time point, the processing simulation state data set is obtained through simulation calculation. These data sets usually include spindle current, feed axis current, vibration, temperature and acoustic emission signals, etc., which reflect the physical state during the processing. Each signal can be regarded as time series data and recorded as a multidimensional state vector. Feature extraction is performed on the simulation state data set to obtain a simulation feature data set. The feature extraction process includes analyzing each time series signal in the time domain, frequency domain and time-frequency domain to extract features that are representative of the processing state. The simulation feature data set is input into the real-time recognition model of the processing state. The simulation feature data set is input into the time domain branch network, which includes 3 one-dimensional convolutional layers and 2 fully connected layers. Through the processing of these convolutional layers and fully connected layers, the time domain feature vector is extracted. The calculation process of the convolutional layer can be expressed as: ; in, For the output features, is the input signal, is the convolution kernel weight, is the bias, BN represents the batch normalization operation, and ReLU represents the activation function. The simulation feature data set is input into the frequency domain branch network. The network adopts the ResNet structure, including 1 input convolution layer, 3 residual blocks and 1 global average pooling layer. The residual block retains the information of the previous layer through short-circuit connection, which effectively alleviates the gradient vanishing problem in the deep network. The operation of the convolution layer can be described as: ; in, and is the convolution kernel weight matrix, and is the bias term, is the input feature, is the output feature. At the same time, the simulation feature dataset is input into the time-frequency domain branch network. The time-frequency domain branch network adopts an attention-enhanced LSTM structure, including 1 bidirectional LSTM layer and 2 self-attention layers. The function of the LSTM layer is to capture the long-term dependencies in the time series, while the self-attention layer focuses on the most important time-frequency domain features by calculating the attention weights. The formula of the attention mechanism is: ; in, , and are query matrix, key matrix and value matrix respectively, is the scaling factor. After obtaining the time domain, frequency domain, and time-frequency domain feature vectors, they are concatenated together to form a fused feature vector: ; The fused feature vector is input into the fully connected classification network. The network consists of three fully connected layers with 256, 128 and 64 neurons respectively. After being processed by the fully connected layer, the classification probability of the processing state is output through the Softmax function. The Softmax function converts the output of the network into a probability distribution: ; in, For the The score of the class, is the total number of categories, For input Belongs to category Through the above steps, the processing state simulation recognition results are finally obtained, including the probability distribution of normal processing, tool wear, chipping and chattering states.

[0032] In a specific embodiment, the execution step inputs the simulation feature data set into the frequency domain branch network of the real-time recognition model of the processing state, and the process of obtaining the frequency domain feature vector after processing by 1 input convolution layer, 3 residual blocks and 1 global average pooling layer can specifically include the following steps: Performing data preprocessing on the frequency domain features in the simulation feature data set, including normalization and standardization operations, to obtain a preprocessed frequency domain feature matrix; The preprocessed frequency domain feature matrix is ​​input into the input convolution layer, and 64 3x3 convolution kernels are used for convolution operation to obtain 64 feature maps. Batch normalization and ReLU activation function are applied to the feature maps to obtain the initial feature map set. The initial feature atlas is input into the first residual block. The first residual block contains two 3x3 convolutional layers and a short-circuit connection. Each convolutional layer is followed by batch normalization and ReLU activation function. The first layer of residual feature atlas is obtained through residual learning. The first-layer residual feature atlas is input into the second residual block. The structure of the second residual block is the same as the first residual block, but the number of convolution kernels is increased to 128. The second-layer residual feature atlas is obtained through residual learning. The second-layer residual feature atlas is input into the third residual block. The structure of the third residual block is the same as that of the second residual block, but the number of convolution kernels is increased to 256. The third-layer residual feature atlas is obtained through residual learning. Perform a global average pooling operation on the third-layer residual feature atlas to obtain a 256-dimensional feature vector, and perform L2 normalization on the 256-dimensional feature vector to obtain a feature vector of unit length; The unit-length feature vector is passed through a fully connected layer to reduce the dimension to 128, and the ReLU activation function is applied to obtain the frequency domain feature vector.

[0033] Specifically, the frequency domain features are preprocessed. This includes normalization and standardization operations. The purpose of normalization is to scale the value ranges of different features to the same range (for example, [0, 1]) to prevent certain features from dominating the learning process of the model due to excessive dimensions. Standardization is to adjust the mean of the feature to 0 and the standard deviation to 1 so that the feature conforms to the standard normal distribution. After normalization and standardization, the preprocessed frequency domain feature matrix is ​​obtained. The preprocessed frequency domain feature matrix is ​​input into the input convolution layer of the convolutional neural network. In this layer, 64 convolution kernels of size 3x3 are used for convolution operations. The convolution operation slides the convolution kernel on the feature matrix and calculates the inner product to extract local features. The mathematical representation of convolution is: ; in, is the convolutional layer Output feature map, For the The weight of the convolution kernel, is the input feature matrix, is the bias term. Through the convolution operation, 64 feature maps are obtained, capturing the local patterns and change trends in the input frequency domain feature matrix. In order to improve the training efficiency and stability of the model, batch normalization and ReLU activation function are applied to the feature maps. Batch normalization adjusts the mean and variance of the output to make the data distribution of each batch more consistent. The formula is: ; in, is the batch mean, is the batch variance, is a small positive number used to prevent division by zero errors. The formula for the ReLU activation function is: ; Through batch normalization and smooth nonlinear activation function ReLU, the overfitting of the model is suppressed, the nonlinear ability of feature expression is improved, and the initial feature atlas is obtained. The initial feature atlas is input into the first residual block. The residual block implements residual learning by adding a short-circuit connection between every two convolutional layers to pass the input signal directly to the output layer. This mechanism can effectively solve the gradient vanishing problem in deep neural networks and accelerate model convergence. The first residual block contains two convolutional layers of size 3x3, each of which is followed by a batch normalization layer and a ReLU activation function. The output of the residual block can be expressed as: ; in, For the The input features of the layer, and is the convolution kernel weight, is the output feature of the residual block. The first-layer residual feature atlas is input to the second residual block. The structure of the second residual block is the same as that of the first residual block, except that the number of convolution kernels is increased to 128. This enables the model to extract more and more complex features. After obtaining the second-layer residual feature atlas through residual learning, the model further extracts high-order features in the signal. The second-layer residual feature atlas is input to the third residual block. The structure of the third residual block is the same as the previous one, and the number of convolution kernels is further increased to 256. As the number of convolution kernels increases, the expressive power of the model is also enhanced, enabling it to capture subtle feature changes in the input signal. The third-layer residual feature atlas is obtained through residual learning. A global average pooling operation is performed on the third-layer residual feature atlas, and all elements of each feature map are averaged to obtain a 256-dimensional feature vector. The formula for global average pooling is: ; in, It is eigenvector components, For the The feature map is at position The value at and is the height and width of the feature map. Global average pooling effectively compresses the spatial information of the feature map into a vector of fixed size while retaining the most important global features. In order to improve the numerical stability and representation ability of the features, the 256-dimensional feature vector is L2 normalized, and the modulus of the feature vector is scaled to 1 so that each component of the feature vector falls on the same scale. The formula for L2 normalization is: ; in, is the eigenvector before normalization, is its L2 norm, is the normalized feature vector. The normalized feature vector is input into a fully connected layer, which reduces the feature dimension from 256 to 128, and applies the ReLU activation function to obtain the frequency domain feature vector.

[0034] In a specific embodiment, the execution step concatenates the time domain feature vector, the frequency domain feature vector and the time-frequency domain feature vector and inputs them into a fully connected classification network, and after being processed by three fully connected layers and a Softmax function, a processing state simulation recognition result is obtained. The processing state simulation recognition result includes a probability distribution of normal processing, tool wear, chipping and chattering states, and the process can specifically include the following steps: The time domain feature vector, the frequency domain feature vector and the time-frequency domain feature vector are concatenated to obtain a fused feature vector, wherein the dimension of the fused feature vector is the sum of the dimension of the time domain feature vector, the dimension of the frequency domain feature vector and the dimension of the time-frequency domain feature vector; The fused feature vector is normalized to obtain a standardized fused feature vector, and the standardized fused feature vector is input into the first fully connected layer. The first fully connected layer contains 256 neurons and uses the ReLU activation function for nonlinear transformation to obtain the first layer of feature representation. Perform Dropout operation on the first layer feature representation, and set the dropout rate to 0.5 to obtain a sparse first layer feature representation; The sparse first-layer feature representation is input into the second fully connected layer, which contains 128 neurons and uses the ReLU activation function for nonlinear transformation to obtain the second-layer feature representation; Apply Batch Normalization operation to the second-layer feature representation to standardize the data distribution and obtain the normalized second-layer feature representation; The normalized second-layer feature representation is input into the third fully connected layer. The third fully connected layer contains 4 neurons, corresponding to the four states of normal machining, tool wear, chipping and chatter, to obtain the original state score vector. The Softmax function is applied to the original state score vector to convert the score into a probability distribution, and the machining state simulation recognition results are obtained, including the probability distribution of normal machining, tool wear, chipping and chattering states.

[0035] Specifically, the time domain feature vector, the frequency domain feature vector, and the time-frequency domain feature vector are concatenated to obtain a fused feature vector. The fused feature vector is subjected to feature normalization to adjust the numerical ranges of different features to the same scale. The commonly used normalization method is to scale the feature value to the interval [0, 1] or [-1, 1], or to standardize the feature so that the mean value of the feature is 0 and the standard deviation is 1. The normalization formula is: ; in, is the original eigenvector, is its mean, is its standard deviation, is the normalized feature vector. Through the normalization operation, a standardized fused feature vector is obtained. The standardized fused feature vector is input into the first fully connected layer. The fully connected layer contains 256 neurons, which are used to map high-dimensional features into a new feature space. In order to enhance the nonlinear expression ability of the network, the ReLU activation function is used for nonlinear transformation. The mathematical expression of the ReLU activation function is: ; The ReLU function suppresses negative values ​​in the input to zero and retains the positive part, thereby introducing nonlinearity and improving the network's ability to model complex data. After processing the fully connected layer, the first layer of feature representation is obtained. In order to prevent the network from overfitting, the first layer of feature representation is subjected to the Dropout operation. Dropout is a regularization technique that forces the network to learn more robust feature representations by randomly discarding a portion of neurons during training. Setting the drop rate to 0.5 means that during training, 50% of the neurons will be randomly discarded each time, making the network unable to rely on certain specific neurons, thereby improving its generalization ability. After the Dropout operation, the sparse first layer of feature representation is obtained. The sparse first layer of feature representation is input to the second fully connected layer. The second fully connected layer contains 128 neurons, and the ReLU activation function is also used for nonlinear transformation to obtain the second layer of feature representation. In order to ensure the consistency of data during training, the Batch Normalization operation is applied to the second layer of feature representation. Batch Normalization adjusts the distribution of each small batch of data to make the distribution of each layer of input more stable, speed up the training of the model, and help prevent the gradient from disappearing or exploding. The calculation formula of Batch Normalization is: ; in, is the standardized feature, E is the batch mean, Var is the batch variance, is a small constant to prevent division by zero errors. After Batch Normalization operation, a normalized second-layer feature representation is obtained. The normalized second-layer feature representation is input into the third fully connected layer. This fully connected layer contains 4 neurons, corresponding to the four states of normal processing, tool wear, chipping and chatter. After processing by the fully connected layer, an original state score vector is obtained. Each component in this vector represents the score of the model for each processing state, and these scores will be passed as input to the Softmax function. Apply the Softmax function to the original state score vector. The Softmax function converts the score into a probability distribution to ensure that the sum of the probabilities of all states is 1. The expression of the Softmax function is: ; in, Represents the input feature vector Corresponding status The probability of Yes Status The original score, is the total number of all possible states. After Softmax transformation, the model outputs a probability vector containing 4 components, corresponding to normal machining, tool wear, chipping and chattering states.

[0036] In a specific embodiment, the process of executing step S5 may specifically include the following steps: The processing state simulation recognition results are processed by probability threshold to obtain the discretized processing state, which is then combined with the initial CNC processing control parameter set to construct the reinforcement learning state space. Based on the reinforcement learning state space and the adjustable range of CNC machining control parameters, the action space is constructed, which includes the adjustment amount of spindle speed, feed rate, cutting depth, tool path and coolant flow rate; Design a reward function based on the processing quality-efficiency-energy consumption function model, normalize the reward function, and obtain a standardized reward function; Based on the reinforcement learning state space, action space and standardized reward function, a deep Q network model is constructed. The deep Q network model includes 4 fully connected layers, each layer uses the ReLU activation function, and the last layer outputs the Q value; The deep Q network model is trained by using experience replay and target network technology, and the ε-greedy strategy is used to balance exploration and utilization to obtain the trained deep Q network model; Based on the trained deep Q network model, the discretized processing state is evaluated, the action corresponding to the maximum Q value is selected, and the dynamic adjustment strategy of the processing parameters is obtained; According to the dynamic adjustment strategy of machining parameters, the initial CNC machining control parameter set is globally optimized to obtain the target CNC machining control parameter set.

[0037] Specifically, a probability threshold is set. If the probability of a state exceeds the preset threshold (for example, 0.7), the current processing state is considered to be this state. For example, if the probability of the chipping state is 0.75, and the probabilities of other states are all lower than 0.7, the current state is discretized as "chipping". The processing state at each moment is clearly divided into a discrete category. The discretized processing state is combined with the initial CNC processing control parameter set (including spindle speed, feed rate, cutting depth, tool path, and coolant flow, etc.) to construct the state space of reinforcement learning. The state space is the core concept of reinforcement learning, which represents the set of all possible states of the system at a certain moment. The state space includes not only processing states (such as normal, worn, chipping, etc.), but also the current processing control parameters. Assume that the processing state has possibilities, and each control parameter has several discrete value combinations, then the dimension of the state space is ,in is the number of all combinations of machining control parameters. Based on the constructed state space, an action space is constructed. The action space defines all operations or actions that can be taken in each state. For CNC machining, these actions usually involve adjusting machining control parameters. Specifically, the action space can include adjusting spindle speed, feed rate, cutting depth, tool path, and coolant flow, etc. Assuming that each parameter can have several different adjustment amounts, the dimension of the action space is the combination of all these adjustment amounts. For example, if the spindle speed can be increased or decreased by 100 rpm, the feed rate can be adjusted by 0.1 mm / s, etc., then all possible action combinations constitute the action space. In order to guide the reinforcement learning model to select the optimal action, a reward function is designed. The reward function is used to evaluate the effect of each action in a specific state, which is usually related to machining quality, efficiency, and energy consumption. The reward function is designed based on a machining quality-efficiency-energy consumption function model, and higher quality, higher efficiency, and lower energy consumption are set to correspond to higher reward values. Assume that the reward function can be expressed as: ; in, Indicates in status Take action The rewards received, is the quality score, For efficiency score, For energy consumption, , and is the corresponding weight coefficient. In order to ensure that the reward function has consistent comparability under different states and actions, the reward function is normalized and its value is scaled to the interval [0, 1]. Based on the constructed state space, action space and standardized reward function, a deep Q network model (DQN) is constructed. DQN is a reinforcement learning-based model that uses deep neural networks to approximate the Q value of a state-action pair. The Q value represents the expected cumulative reward after taking an action in a given state. The DQN model usually consists of multiple fully connected layers, each of which uses the ReLU activation function to introduce nonlinear transformations. Assume that the DQN model contains 4 fully connected layers, the first The output of the layer can be expressed as: ; in, For the The output features of the layer, is the weight matrix, is the bias vector. After the last fully connected layer, the DQN model outputs the Q value corresponding to each action. In order to train the DQN model, experience replay and target network technology are used. Experience replay means that during the training process, the model stores past experience in a buffer and randomly extracts these experiences for learning during training, thereby breaking the correlation between data and improving training stability. The target network means that a fixed network is used to calculate the target Q value during the training process, thereby preventing the training from being unstable due to too fast network updates. When balancing exploration and utilization, the ε-greedy strategy is used, that is, the current optimal action (that is, the action with the highest Q value) is selected in most cases, but other actions are also randomly selected with a certain probability (ε) to prevent the model from falling into the local optimum. After the model training is completed, the discretized processing state is evaluated based on the trained DQN model. For each state, the model calculates the Q value of all possible actions and selects the action with the largest Q value as the optimal action. The parameter adjustment strategy corresponding to this optimal action is the dynamic adjustment strategy of processing parameters. According to the dynamic adjustment strategy of processing parameters, the initial CNC processing control parameter set is globally optimized. By continuously adjusting and optimizing these parameters, an optimal set of CNC machining control parameters is obtained, thereby maximizing efficiency and minimizing energy consumption while ensuring machining quality.

[0038] In a specific embodiment, the execution step evaluates the discretized processing state based on the trained deep Q network model, selects the action corresponding to the maximum Q value, and obtains the dynamic adjustment strategy of the processing parameters, including: S1: Encode the discretized machining state and the initial CNC machining control parameter set to obtain the state vector; S2: Input the state vector into the trained deep Q network model, calculate the Q value of each possible action through forward propagation, and form a Q value vector; S3: Apply the ε-greedy strategy to the Q value vector, select the action with the largest Q value with a probability of 1-ε, and randomly select an action with a probability of ε to obtain candidate actions; S4: Based on the candidate actions, fine-tune the initial CNC machining control parameter set to obtain adjusted machining control parameters; S5: inputting the adjusted processing control parameters into the processing quality-efficiency-energy consumption function model to calculate the expected performance indicators; S6: Determine whether to accept the candidate action based on the comparison result between the expected performance index and the current performance index. If accepted, output it as a dynamic adjustment strategy for processing parameters. Otherwise, return to step S3 to reselect the action.

[0039] Specifically, the discretized processing state and the initial CNC processing control parameter set are encoded. Different types of data (such as processing state, spindle speed, feed speed, etc.) are converted into a unified vector representation. Assuming there is a discretized processing state (such as "normal processing", "tool wear", "edge breaking", etc.), and the corresponding initial CNC processing control parameters (such as spindle speed of 2000RPM, feed speed of 0.5mm / s, etc.), assign a value or label to each state and parameter and convert it into a vector representation. For example, suppose the discretized processing state is , the initial CNC machining control parameter set is , then the state vector can be expressed as:

[0040] in, are numerical representations of each processing control parameter. The state vector is used to represent the current system state. Input into the trained deep Q network model. The DQN model calculates the Q value of each possible action through forward propagation to form a Q value vector. The forward propagation process involves the operation of multiple fully connected layers. Assuming that the DQN model contains 4 fully connected layers, the calculation of each layer can be expressed as:

[0041] in, Indicates The output feature vector of the layer, and are the weight matrix and bias vector of the layer respectively, and ReLU is the activation function. In the last layer, the output of the fully connected layer directly generates the Q value vector , each component Indicates selecting an action in the current state The expected cumulative reward. After obtaining the Q value vector After that, the ε-greedy strategy is applied to select the optimal action. The ε-greedy strategy is a strategy that balances exploration and exploitation. The probability of choosing the action with the largest Q value is called "greedy choice", but it can also be used as The probability of randomly selecting an action to avoid falling into the local optimal solution is called "exploration" selection. Assuming the value of ε is set to 0.1, there is a 90% probability of selecting the current optimal action and a 10% probability of randomly selecting other actions. Through this strategy, a candidate action is obtained Based on the selected candidate action , for the initial CNC machining control parameter set Fine-tune and obtain the adjusted processing control parameters The specific fine-tuning method depends on the selected action. For example, if the candidate action is to increase the spindle speed, then the spindle speed parameter is appropriately increased, and other parameters remain unchanged; if the candidate action is to reduce the feed speed, then the feed speed is adjusted accordingly. Input into the processing quality-efficiency-energy consumption function model to calculate the expected performance indicators. This function model is usually designed according to the requirements of the specific processing task and can comprehensively evaluate the quality, efficiency and energy consumption. For example, a possible function model form is:

[0042] in, Represents quality indicators, represents the efficiency index, Indicates the energy consumption index. , and is the corresponding weight coefficient. Through this function, the impact of different parameter combinations on the overall processing effect is quantified, and the expected performance score is obtained. The calculated expected performance index is compared with the current performance index. If the expected performance is better than the current performance (that is, the performance score is higher), it means that the selected candidate action is effective, and the action is accepted and output as the final dynamic adjustment strategy for processing parameters; if the expected performance fails to improve, the candidate action is considered unsatisfactory, the action is rejected, and the process returns to step S3 to reselect the action. In this iterative way, the system can gradually optimize the processing parameters until the optimal adjustment strategy is found.

[0043] The intelligent CNC machining method in the embodiment of the present invention is described above. The intelligent CNC machining system in the embodiment of the present invention is described below. Figure 2 , an embodiment of the intelligent CNC machining system in the embodiment of the present invention includes: Modeling module, used to comprehensively model the processing quality, efficiency and energy consumption of CNC, and obtain the processing quality-efficiency-energy consumption function model; An extraction module is used to create a multi-sensor data acquisition scheme based on a processing quality-efficiency-energy consumption function model and obtain a multi-dimensional time series state data set in the CNC processing process, perform time-frequency domain feature extraction and noise reduction on the multi-dimensional time series state data set, and obtain a feature data set; The training module is used to train a multi-branch hybrid deep neural network based on the feature data set and build a real-time recognition model for processing status; The recognition module is used to perform five-axis linkage CNC machining simulation based on a preset initial CNC machining control parameter set to obtain a machining simulation state data set, and input the machining simulation state data set into a machining state real-time recognition model to perform machining state classification recognition to obtain a machining state simulation recognition result; The optimization module is used to perform deep reinforcement learning based on the machining state simulation recognition results, generate a dynamic adjustment strategy for machining parameters, and perform global parameter optimization on the initial CNC machining control parameter set according to the dynamic adjustment strategy for machining parameters to obtain the target CNC machining control parameter set.

[0044] Through the cooperation of the above components, the quality, efficiency and energy consumption of CNC machining are comprehensively modeled to obtain the machining quality-efficiency-energy consumption function model, which realizes the quantitative expression of multi-objective optimization problems. A multi-sensor data acquisition scheme is adopted to obtain the multi-dimensional time series state data in the CNC machining process, and the time-frequency domain feature extraction and noise reduction processing are performed. The perception ability of the machining state is improved, and a real-time recognition model of the machining state is constructed by training a multi-branch hybrid deep neural network. This network structure can make full use of the time domain, frequency domain and time-frequency domain features to improve the accuracy and robustness of machining state recognition. Five-axis linkage CNC machining simulation is carried out based on the initial machining parameters, and the simulation data is input into the state recognition model to realize the prediction and evaluation of the machining process. The risk of actual machining is reduced and the efficiency of parameter optimization is improved. The deep reinforcement learning method is used to generate a dynamic adjustment strategy for machining parameters, and the adaptive optimization of machining parameters is realized. This method can adjust parameters according to the real-time machining state, which improves the intelligence level and adaptability of the machining process. The initial CNC machining control parameter set is globally optimized based on the dynamic adjustment strategy to obtain the target parameter set. This optimization method takes the entire machining process into consideration and can achieve the overall optimization of machining quality, efficiency and energy consumption. A dynamic weight adjustment mechanism is introduced, an adaptive QEE function is constructed, and sensitivity analysis and optimization are performed. This method improves the model's adaptability to different machining requirements. By constructing a multi-scale feature fusion matrix and performing feature dimensionality reduction, a more representative feature data set is extracted. The efficiency and accuracy of state recognition are improved.

[0045] The present invention also provides a computer device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the intelligent CNC machining method in the above-mentioned embodiments.

[0046] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the intelligent CNC machining method.

[0047] 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.

[0048] 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.

[0049] 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. An intelligent CNC processing method, characterized in that: The intelligent CNC processing method comprises: Comprehensively model the machining quality, efficiency and energy consumption of CNC to obtain the machining quality-efficiency-energy consumption function model; Based on the processing quality-efficiency-energy consumption function model, a multi-sensor data acquisition scheme is created to obtain a multi-dimensional time series state data set in the CNC processing process, and time-frequency domain feature extraction and noise reduction processing are performed on the multi-dimensional time series state data set to obtain a feature data set; Training a multi-branch hybrid deep neural network based on the feature data set to build a real-time recognition model for processing status; Based on a preset initial CNC machining control parameter set, a five-axis linkage CNC machining simulation is performed to obtain a machining simulation state data set, and the machining simulation state data set is input into the machining state real-time recognition model to perform machining state classification recognition to obtain a machining state simulation recognition result; Based on the machining state simulation recognition result, deep reinforcement learning is performed to generate a machining parameter dynamic adjustment strategy, and the initial CNC machining control parameter set is globally optimized according to the machining parameter dynamic adjustment strategy to obtain a target CNC machining control parameter set.

2. The intelligent CNC processing method according to claim 1, characterized in that: The CNC machining quality, efficiency and energy consumption are comprehensively modeled to obtain a machining quality-efficiency-energy consumption function model, including: Quantifying the processing quality index to obtain a processing quality sub-model, wherein the processing quality sub-model includes surface roughness, dimensional accuracy and shape accuracy, and obtaining a processing quality score through weighted summation; Quantifying the processing efficiency index to obtain a processing efficiency sub-model, wherein the processing efficiency sub-model includes material removal rate and processing time, and obtaining a processing efficiency score by establishing a nonlinear relationship function; The energy consumption index is quantified to obtain an energy consumption sub-model, wherein the energy consumption sub-model includes spindle power, feed power and auxiliary equipment power consumption, and the total energy consumption is calculated by integration to obtain an energy consumption score; Based on the processing quality sub-model, the processing efficiency sub-model and the energy consumption sub-model, a multi-objective optimization problem is constructed to obtain an initial QEE function, and the initial QEE function is normalized to obtain a normalized QEE function; Based on the normalized QEE function, a dynamic weight adjustment mechanism is introduced to obtain an adaptive QEE function, and a sensitivity analysis is performed on the adaptive QEE function to obtain the influence of each parameter on the QEE function. The adaptive QEE function is optimized based on the influence to obtain a processing quality-efficiency-energy consumption function model.

3. The intelligent CNC processing method according to claim 1, characterized in that: The method of creating a multi-sensor data acquisition scheme based on the processing quality-efficiency-energy consumption function model and obtaining a multi-dimensional time series state data set in the CNC processing process, performing time-frequency domain feature extraction and noise reduction processing on the multi-dimensional time series state data set to obtain a feature data set includes: Based on the machining quality-efficiency-energy consumption function model, the target parameters in the CNC machining process are analyzed to obtain a sensor layout scheme, which includes the installation positions and sampling frequencies of the spindle current sensor, the feed axis current sensor, the vibration sensor, the temperature sensor, and the acoustic emission sensor; According to the sensor arrangement scheme, real-time data collection is performed on the CNC machining process to obtain a multi-dimensional time series state data set, wherein the multi-dimensional time series state data set includes spindle current, feed axis current, vibration, temperature and acoustic emission signals; Performing time domain analysis on the multidimensional time series state data set, extracting time domain features, and obtaining a time domain feature set, wherein the time domain features include mean, variance, peak factor, skewness, and kurtosis; Based on the fast Fourier transform algorithm, frequency domain analysis is performed on the multidimensional time series state data set to extract frequency domain features to obtain a frequency domain feature set, wherein the frequency domain features include spectrum energy, main frequency and frequency band energy ratio; Using a wavelet packet transform algorithm, the original multi-dimensional time series state data set is analyzed in the time-frequency domain, and time-frequency domain features are extracted to obtain a time-frequency domain feature set, wherein the time-frequency domain features include wavelet packet energy entropy and wavelet packet energy ratio; Based on the time domain feature set, the frequency domain feature set and the time-frequency domain feature set, a multi-scale feature fusion matrix is ​​constructed, and feature dimension reduction is performed on the multi-scale feature fusion matrix to obtain a feature data set.

4. The intelligent CNC processing method according to claim 1, characterized in that: The method of training a multi-branch hybrid deep neural network according to the feature data set to construct a real-time processing state recognition model includes: Segmenting the feature data set to obtain a time domain feature subset, a frequency domain feature subset, and a time-frequency domain feature subset; A time-domain branch network is constructed based on the time-domain feature subset, wherein the time-domain branch network includes three one-dimensional convolutional layers and two fully connected layers, each one-dimensional convolutional layer is followed by a batch normalization layer and a ReLU activation function, the convolution kernel size of the first one-dimensional convolutional layer is 5 and the stride is 2, the convolution kernel size of the second and third one-dimensional convolutional layers is 3 and the stride is 1, and the number of neurons in the two fully connected layers is 128 and 64, respectively; A frequency domain branch network is constructed based on the frequency domain feature subset, wherein the frequency domain branch network adopts a ResNet structure, including 1 input convolution layer, 3 residual blocks and 1 global average pooling layer, each residual block contains 2 two-dimensional convolution layers and 1 short-circuit connection, the convolution kernel size of the convolution layer is 3x3, the stride is 1, and a batch normalization layer and a ReLU activation function are added after each convolution layer; Based on the time-frequency domain feature subset, a time-frequency domain branch network is constructed, wherein the time-frequency domain branch network adopts an attention-enhanced LSTM structure, including one bidirectional LSTM layer and two self-attention layers, the number of hidden units in the LSTM layer is 128, and the number of heads in the self-attention layer is 4; The output features of the time domain branch network, the frequency domain branch network and the time-frequency domain branch network are concatenated to obtain a fused feature vector; Applying a feature selection algorithm to the fused feature vector to select important features to obtain an optimized feature vector, wherein the feature selection algorithm adopts a feature importance ranking method based on XGBoost; Based on the optimized feature vector, a fully connected classification network is constructed, wherein the fully connected classification network includes three fully connected layers, the number of neurons is 256, 128 and 64 respectively, each fully connected layer is followed by a Dropout layer and a ReLU activation function, and the last layer uses a Softmax function to output the classification probability of the processing state; The time domain branch network, frequency domain branch network, time-frequency domain branch network and fully connected classification network are combined in series to obtain a multi-branch hybrid deep neural network, and the multi-branch hybrid deep neural network is end-to-end trained using a cross entropy loss function and an Adam optimizer to obtain a real-time recognition model for processing status.

5. The intelligent CNC processing method according to claim 1, characterized in that: The five-axis linkage CNC machining simulation is performed based on the preset initial CNC machining control parameter set to obtain a machining simulation state data set, and the machining simulation state data set is input into the machining state real-time recognition model to perform machining state classification recognition to obtain a machining state simulation recognition result, including: Parameter encoding is performed on a preset initial CNC machining control parameter set to obtain a parameter encoding vector, wherein the parameter encoding vector includes spindle speed, feed speed, cutting depth, tool path and coolant flow parameters; A five-axis linkage CNC machining simulation model is constructed based on the parameter coding vector, wherein the simulation model uses a finite element analysis method to simulate and calculate cutting force, heat distribution and workpiece deformation during machining; Performing time discretization processing on the five-axis linkage CNC machining simulation model to obtain a discrete time series, and performing simulation calculations at each discrete time point to obtain a machining simulation state data set, wherein the machining simulation state data set includes simulated spindle current, feed axis current, vibration, temperature, and acoustic emission signals; Performing feature extraction on the processing simulation state data set to obtain a simulation feature data set, and inputting the simulation feature data set into the time domain branch network of the processing state real-time recognition model, and processing through three one-dimensional convolutional layers and two fully connected layers to obtain a time domain feature vector; Inputting the simulation feature data set into the frequency domain branch network of the real-time recognition model of the processing state, and processing it through 1 input convolution layer, 3 residual blocks and 1 global average pooling layer to obtain a frequency domain feature vector; Input the simulation feature data set into the time-frequency domain branch network of the real-time recognition model of the processing state, and process it through one bidirectional LSTM layer and two self-attention layers to obtain a time-frequency domain feature vector; The time domain feature vector, frequency domain feature vector and time-frequency domain feature vector are spliced ​​and input into a fully connected classification network. After being processed by three fully connected layers and a Softmax function, a machining state simulation recognition result is obtained. The machining state simulation recognition result includes the probability distribution of normal machining, tool wear, chipping and chattering states.

6. The intelligent CNC processing method according to claim 5, characterized in that: The simulation feature data set is input into the frequency domain branch network of the real-time recognition model of the processing state, and processed by 1 input convolution layer, 3 residual blocks and 1 global average pooling layer to obtain a frequency domain feature vector, including: Performing data preprocessing on the frequency domain features in the simulation feature data set, including normalization and standardization operations, to obtain a preprocessed frequency domain feature matrix; Inputting the preprocessed frequency domain feature matrix into the input convolution layer, performing convolution operation using 64 3x3 convolution kernels to obtain 64 feature maps, and applying batch normalization and ReLU activation function to the feature maps to obtain an initial feature map set; Inputting the initial feature atlas into the first residual block, the first residual block comprises two 3x3 convolutional layers and a short circuit connection, each convolutional layer is followed by batch normalization and a ReLU activation function, and the first layer of residual feature atlas is obtained through residual learning; The first layer of residual feature atlas is input into the second residual block. The structure of the second residual block is the same as that of the first residual block, but the number of convolution kernels is increased to 128. The second layer of residual feature atlas is obtained through residual learning. The second layer residual feature atlas is input into the third residual block. The structure of the third residual block is the same as that of the second residual block, but the number of convolution kernels is increased to 256. The third layer residual feature atlas is obtained through residual learning. Performing a global average pooling operation on the third-layer residual feature atlas to obtain a 256-dimensional feature vector, and performing an L2 normalization process on the 256-dimensional feature vector to obtain a feature vector of unit length; The unit-length feature vector is passed through a fully connected layer to reduce the dimension to 128, and a ReLU activation function is applied to obtain a frequency-domain feature vector.

7. The intelligent CNC processing method according to claim 5, characterized in that: The time domain feature vector, the frequency domain feature vector and the time-frequency domain feature vector are spliced ​​and input into a fully connected classification network, and processed by three fully connected layers and a Softmax function to obtain a processing state simulation recognition result, which includes the probability distribution of normal processing, tool wear, chipping and chattering states, including: Performing a splicing operation on the time domain feature vector, the frequency domain feature vector and the time-frequency domain feature vector to obtain a fused feature vector, wherein the dimension of the fused feature vector is the sum of the dimension of the time domain feature vector, the dimension of the frequency domain feature vector and the dimension of the time-frequency domain feature vector; Performing feature normalization processing on the fused feature vector to obtain a standardized fused feature vector, and inputting the standardized fused feature vector into a first fully connected layer, the first fully connected layer includes 256 neurons, and performing nonlinear transformation using a ReLU activation function to obtain a first layer of feature representation; Perform a Dropout operation on the first-layer feature representation, with the dropout rate set to 0.5, to obtain a sparse first-layer feature representation; Input the sparse first-layer feature representation into a second fully connected layer, the second fully connected layer includes 128 neurons, and uses a ReLU activation function to perform nonlinear transformation to obtain a second-layer feature representation; Applying Batch Normalization operation to the second-layer feature representation to standardize data distribution and obtain a normalized second-layer feature representation; Input the normalized second-layer feature representation into a third fully connected layer, the third fully connected layer includes 4 neurons, corresponding to four states of normal machining, tool wear, chipping and chatter, to obtain an original state score vector; The Softmax function is applied to the original state score vector to convert the score into a probability distribution, thereby obtaining a machining state simulation recognition result, including the probability distribution of normal machining, tool wear, chipping, and chattering states.

8. The intelligent CNC processing method according to claim 1, characterized in that: The deep reinforcement learning is performed based on the machining state simulation recognition result to generate a machining parameter dynamic adjustment strategy, and the initial CNC machining control parameter set is globally optimized according to the machining parameter dynamic adjustment strategy to obtain a target CNC machining control parameter set, including: Performing probability threshold processing on the machining state simulation recognition result to obtain a discretized machining state, combining the discretized machining state with an initial CNC machining control parameter set to construct a reinforcement learning state space; Based on the reinforcement learning state space and the adjustable range of the CNC machining control parameters, an action space is constructed, wherein the action space includes the adjustment amounts of the spindle speed, feed speed, cutting depth, tool path and coolant flow rate; Designing a reward function according to the processing quality-efficiency-energy consumption function model, and normalizing the reward function to obtain a standardized reward function; Based on the reinforcement learning state space, the action space and the standardized reward function, a deep Q network model is constructed, wherein the deep Q network model includes 4 fully connected layers, each layer uses a ReLU activation function, and the last layer outputs a Q value; The deep Q network model is trained by using experience replay and target network technology, and an ε-greedy strategy is used to balance exploration and utilization to obtain a trained deep Q network model; Based on the trained deep Q network model, the discretized processing state is evaluated, the action corresponding to the maximum Q value is selected, and a dynamic adjustment strategy for processing parameters is obtained; According to the machining parameter dynamic adjustment strategy, global parameter optimization is performed on the initial CNC machining control parameter set to obtain a target CNC machining control parameter set.

9. The intelligent CNC processing method according to claim 8, characterized in that: The method of evaluating the discretized processing state based on the trained deep Q network model, selecting the action corresponding to the maximum Q value, and obtaining a dynamic adjustment strategy for processing parameters includes: S1: Encode the discretized machining state and the initial CNC machining control parameter set to obtain a state vector; S2: Input the state vector into the trained deep Q network model, and calculate the Q value of each possible action through forward propagation to form a Q value vector; S3: Apply the ε-greedy strategy to the Q value vector, select the action with the largest Q value with a probability of 1-ε, and randomly select an action with a probability of ε to obtain a candidate action; S4: fine-tuning the initial CNC machining control parameter set based on the candidate action to obtain adjusted machining control parameters; S5: inputting the adjusted processing control parameters into a processing quality-efficiency-energy consumption function model to calculate the expected performance index; S6: Determine whether to accept the candidate action based on the comparison result between the expected performance index and the current performance index. If accepted, output it as a dynamic adjustment strategy for processing parameters. Otherwise, return to step S3 to reselect an action.

10. An intelligent CNC processing system, characterized in that: Used to perform the intelligent CNC machining method according to any one of claims 1 to 9, the intelligent CNC machining system comprises: Modeling module, used to comprehensively model the processing quality, efficiency and energy consumption of CNC, and obtain the processing quality-efficiency-energy consumption function model; An extraction module is used to create a multi-sensor data acquisition scheme based on the processing quality-efficiency-energy consumption function model and obtain a multi-dimensional time series state data set in the CNC processing process, perform time-frequency domain feature extraction and noise reduction processing on the multi-dimensional time series state data set to obtain a feature data set; A training module, used for training a multi-branch hybrid deep neural network according to the feature data set to construct a real-time recognition model for processing status; An identification module is used to perform five-axis linkage CNC machining simulation based on a preset initial CNC machining control parameter set to obtain a machining simulation state data set, and input the machining simulation state data set into the machining state real-time identification model to perform machining state classification identification to obtain a machining state simulation identification result; The optimization module is used to perform deep reinforcement learning based on the machining state simulation recognition result, generate a dynamic adjustment strategy for machining parameters, and perform global parameter optimization on the initial CNC machining control parameter set according to the dynamic adjustment strategy for machining parameters to obtain a target CNC machining control parameter set.

Images