Intelligent machine tool cutter wear state recognition and machining parameter adjustment method and system

Abstract

The invention provides an intelligent machine tool cutter wear state recognition and machining parameter adjustment method and system, and relates to the technical field of intelligent manufacturing, and the method comprises the steps: collecting the working condition data of a machine tool machining process, carrying out the sliding window segmentation, denoising and normalization processing, extracting a time sequence feature sequence through a bidirectional gating circulation unit, and carrying out the recognition of a machining parameter; the method comprises the following steps: acquiring an intrinsic mode component by combining Hilbert-Huang transform decomposition, extracting a frequency domain feature sequence by utilizing four-layer spatial pyramid pooling, determining a fusion weight based on mutual information entropy to obtain a tool wear state feature vector, and inputting the feature vector into a pre-training extreme learning machine to predict the wear loss and the residual life. And determining an initial adjustment scheme in combination with observation sharing information of a parallel agent structure, predicting a processing quality score through a capsule network and feeding back the score as a reward signal, and obtaining an optimization correction strategy for adjusting working parameters of a machine tool.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and particularly relates to a method and system for identifying the tool wear state of an intelligent machine tool and adjusting machining parameters. Background Art

[0002] With the rapid development of intelligent manufacturing technology, intelligent machine tools have been widely used in modern manufacturing. During the machining process of an intelligent machine tool, the tool wear state directly affects machining accuracy and product quality. Tool wear is an inevitable phenomenon in the machining process. As the tool usage time increases, the geometric shape and cutting performance of the tool will gradually deteriorate. Accurately identifying the tool wear state and timely adjusting machining parameters are of great significance for ensuring machining quality and improving production efficiency;

[0003] Tool wear state monitoring mainly adopts two methods: direct measurement and indirect measurement. The direct measurement method directly measures the change of the tool's geometric parameters through optical or tactile means, while the indirect measurement method infers the tool wear state by collecting signals such as cutting force, vibration, and acoustic emission during the machining process. With the development of sensing technology and artificial intelligence technology, the tool wear state identification method based on working condition data analysis has gradually become a research hotspot;

[0004] In the prior art, there are still problems such as the inability to fully capture the long-term and short-term dependencies during the tool wear process, resulting in low accuracy of tool wear state identification, lack of in-depth understanding of the relationship between tool wear state and machining parameters, mostly using empirical formulas or simple feedback adjustment mechanisms for parameter adjustment, unable to perform intelligent and personalized parameter optimization according to the specific wear state and remaining service life of the tool, and using a single decision-making mechanism for machining parameter adjustment, lacking a multi-agent collaborative decision-making framework, unable to comprehensively consider multiple factors such as tool wear, machining quality, and production efficiency, making it difficult to achieve optimal control in a dynamically changing machining environment, and also lacking an evaluation and optimization mechanism for decision-making results;

[0005] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for identifying the tool wear state of an intelligent machine tool and adjusting machining parameters, which can at least solve some of the problems existing in the prior art.

[0007] In the first aspect of the embodiments of the present invention, a method for identifying the tool wear state of an intelligent machine tool and adjusting machining parameters is provided, including:

[0008] Collecting the working condition data during the machining process of the intelligent machine tool and performing sliding window segmentation, and denoising and normalizing the segmented working condition data to obtain standard working condition data;

[0009] Extract the long - short - term time - series feature sequence from the standard working condition data through a bidirectional gated recurrent unit, input the long - short - term time - series feature sequence into the Hilbert - Huang transform for decomposition to obtain the intrinsic mode components, combine four - layer spatial pyramid pooling to extract the multi - scale frequency - domain feature sequence, calculate the mutual information entropy between the time - series feature sequence and the frequency - domain feature sequence, determine the fusion weight coefficient according to the mutual information entropy value and perform weighting to obtain the tool wear state feature vector;

[0010] Input the tool wear state feature vector into a pre - trained extreme learning machine to calculate the current tool wear amount value and the remaining service life value;

[0011] Add the current tool wear amount value and the remaining service life value to the parallel agent structure, determine the initial adjustment plan by combining the observed shared information collected by each agent, predict the machining quality score corresponding to the initial adjustment plan through a capsule network, and feedback the machining quality score as a reward signal to the parallel agent structure to obtain the optimized correction strategy;

[0012] Adjust the working parameters of the intelligent machine tool based on the optimized correction strategy and add them to the machine tool process parameter library.

[0013] In an alternative embodiment,

[0014] Collect the working condition data during the machining process of the intelligent machine tool and perform sliding window segmentation. Denoise and normalize the segmented working condition data to obtain the standard working condition data, including:

[0015] Collect the working condition data during the machining process of the intelligent machine tool, and the working condition data includes spindle current signal, spindle speed signal, feed speed signal, and cutting force signal;

[0016] Perform sliding window segmentation on the working condition data, with an overlap rate of 50% between adjacent windows, to obtain the segmented working condition data sequence;

[0017] For the segmented working condition data sequence, perform multi - scale decomposition and reconstruction of the signal through wavelet threshold method to remove noise, combine Butterworth low - pass filter to filter out high - frequency interference components higher than 50% of the sampling frequency, and perform maximum - minimum normalization on the filtered signal to obtain the standard working condition data.

[0018] In an alternative embodiment,

[0019] Extract the long - short - term time - series feature sequence from the standard working condition data through a bidirectional gated recurrent unit, input the long - short - term time - series feature sequence into the Hilbert - Huang transform for decomposition to obtain the intrinsic mode components, combine four - layer spatial pyramid pooling to extract the multi - scale frequency - domain feature sequence, including:

[0020] Input the standard working condition data into the bidirectional gated recurrent unit network. Establish the forward hidden layer state and the backward hidden layer state in the bidirectional gated recurrent unit network. Calculate and update the forward hidden layer state and the backward hidden layer state respectively through the reset gate and the update gate. Concatenate the updated forward hidden layer state and the backward hidden layer state to generate a long short-term time series feature sequence.

[0021] Decompose the long short-term time series feature sequence through empirical mode decomposition to obtain the intrinsic mode function components. Perform Hilbert transform on the intrinsic mode function components to obtain the Hilbert transform components. Construct an analytic signal based on the intrinsic mode function components and the Hilbert transform components. Extract the instantaneous amplitude and the instantaneous phase from the analytic signal to form the intrinsic mode components.

[0022] Set pooling scales of 1×1, 2×2, 4×4, and 8×8 respectively in the four-layer spatial pyramid pooling network. Input the intrinsic mode components into the four-layer spatial pyramid pooling network, generate a pooled output feature map and perform max pooling operation to obtain four-layer pooled features. Concatenate the four-layer pooled features to form a multi-scale frequency domain feature sequence.

[0023] In an alternative embodiment,

[0024] Calculate the mutual information entropy between the time series feature sequence and the frequency domain feature sequence. Determine the fusion weight coefficient according to the mutual information entropy value and perform weighting to obtain the tool wear state feature vector, including:

[0025] Calculate the mutual information entropy value between the two feature sequences through the operation of multiplying the logarithms of each group of components of the time series feature sequence and the frequency domain feature sequence and then summing them.

[0026] Divide the difference between the eigenvalue in the time series feature sequence and the frequency domain feature sequence and the corresponding feature mean by the bandwidth parameter to obtain the normalized eigenvalue. Substitute the normalized eigenvalue into the Gaussian kernel function. Divide the result calculated by the Gaussian kernel function by the product of the number of samples and the bandwidth parameter to obtain the probability distribution value.

[0027] Multiply the negative mutual information entropy value by the adjustment parameter and perform exponential operation to obtain the numerator of the weight coefficient. Divide the numerator of the weight coefficient by the sum of the numerator of the weight coefficient and its complement to obtain the fusion weight coefficient of the time series feature sequence. Subtract the fusion weight coefficient of the time series feature sequence from 1 to obtain the fusion weight coefficient of the frequency domain feature sequence.

[0028] Multiply the time series feature sequence by the fusion weight coefficient of the time series feature sequence to obtain the weighted time series feature. Multiply the frequency domain feature sequence by the fusion weight coefficient of the frequency domain feature sequence to obtain the weighted frequency domain feature. Add the weighted time series feature and the weighted frequency domain feature to obtain the fusion feature sequence.

[0029] Extract the eigenvector corresponding to the maximum eigenvalue in the fusion feature sequence as the main feature component to construct the tool wear state feature vector.

[0030] In an alternative embodiment,

[0031] Input the tool wear state feature vector into a pre-trained extreme learning machine, and calculate the current tool wear amount value and the remaining service life value, including:

[0032] Collect the tool wear state sample data corresponding to the tool wear state feature vector, divide the tool wear state sample data into a training data set and a test data set, train the extreme learning machine based on the training data set. The extreme learning machine includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is the dimension of the tool wear state feature vector. The hidden layer uses the Sigmoid activation function. By randomly generating the initial values of the input weight matrix and the bias vector, multiply the tool wear state feature vector in the training data set by the input weight matrix and add the bias vector, and obtain the hidden layer output through the Sigmoid activation function. Calculate the output weight based on the hidden layer output to obtain the pre-trained extreme learning machine;

[0033] Input the tool wear state feature vector into the pre-trained extreme learning machine. Multiply the input weight matrix by the tool wear state feature vector and add the bias vector, and then obtain the hidden layer output through the Sigmoid activation function. Calculate the current tool wear amount value and the remaining service life value according to the hidden layer output.

[0034] In an alternative embodiment,

[0035] Add the current tool wear amount value and the remaining service life value to the parallel agent structure, determine the initial adjustment plan by combining the observation sharing information collected by each agent, predict the processing quality score corresponding to the initial adjustment plan through the capsule network, and feedback the processing quality score as a reward signal to the parallel agent structure to obtain the optimized correction strategy, including:

[0036] Add the current tool wear amount value and the remaining service life value to the parallel agent structure. The parallel agent structure includes multiple parallel agents. Each parallel agent collects local observation sharing information, and interacts the local observation sharing information through a shared network to obtain global observation sharing information;

[0037] Calculate the attention weight coefficients between the parallel agents according to the global observation sharing information. The attention weight coefficients are obtained through the inner product operation of the query vector and the key vector corresponding to the parallel agents. Perform weighted accumulation of the attention weight coefficients and the corresponding value vectors to obtain a fused feature representation. Determine an initial adjustment scheme based on the fused feature representation and the observation sharing information collected by each parallel agent;

[0038] Input the initial adjustment scheme into the capsule network. Perform transformation on the initial adjustment scheme through the transformation matrix of the capsule network to obtain the output of the initial capsule layer. Calculate the routing weights based on the output of the initial capsule layer. Obtain the output capsule of the capsule network through the product of the routing weights and the output of the initial capsule layer after passing through a compression function. Predict the processing quality score corresponding to the initial adjustment scheme based on the output capsule;

[0039] Feed the processing quality score back to the parallel agent structure as a reward signal. The parallel agent structure encodes the processing quality score to obtain a historical information encoding. Input the historical information encoding and the currently obtained current state encoding into the policy network to generate a probability distribution of adjustment actions. Optimize according to the policy optimization criterion to obtain an optimized correction policy.

[0040] In an alternative embodiment,

[0041] Obtaining the output capsule of the capsule network through the product of the routing weights and the output of the initial capsule layer after passing through a compression function, and predicting the processing quality score corresponding to the initial adjustment scheme based on the output capsule includes:

[0042] Construct the routing weights in the capsule network and the output of the initial capsule layer as nodes in a Markov random field structure. Calculate the edge connection probability based on the spatial position relationship and feature similarity between the nodes. Construct the edge connection probability as a potential function. Construct a single-node potential function based on distribution parameters for the nodes. Construct the joint probability distribution of the node states through the product of the single-node potential function and the potential function;

[0043] Based on the joint probability distribution, initialize the state values of the nodes according to Gibbs sampling. Calculate the conditional probability of the nodes based on the single-node potential function and the potential function. Iteratively update the node states according to the conditional probability to obtain a sampling sequence. Monitor the autocorrelation of the sampling sequence and dynamically adjust the sampling step size and the number of iterations until a preset convergence condition is met;

[0044] Calculate the optimized routing weights through the marginal probability distribution of the sampling sequence. Multiply the optimized routing weights by the output of the initial capsule layer to obtain a product result. Input the product result into the compression function to obtain the output capsule of the capsule network;

[0045] Calculate the inner product of the output capsule and a preset parameter vector to obtain an initial score. Based on the initial score, obtain a normalized score through a non-linear activation function and use it as the processing quality score corresponding to the initial adjustment scheme.

[0046] In a second aspect of the embodiments of the present invention, there is provided an intelligent machine tool tool wear state recognition and processing parameter adjustment system, including:

[0047] A first unit for collecting working condition data during the machining process of an intelligent machine tool, performing sliding window segmentation on the segmented working condition data, and performing denoising and normalization processing on the segmented working condition data to obtain standard working condition data;

[0048] A second unit for extracting long-term and short-term time series feature sequences from the standard working condition data through a bidirectional gated recurrent unit, inputting the long-term and short-term time series feature sequences into a Hilbert-Huang transform decomposition to obtain intrinsic mode components, combining four-layer spatial pyramid pooling to extract multi-scale frequency domain feature sequences, calculating the mutual information entropy between the time series feature sequences and the frequency domain feature sequences, determining a fusion weight coefficient according to the mutual information entropy value, and performing weighting to obtain a tool wear state feature vector;

[0049] A third unit for inputting the tool wear state feature vector into a pre-trained extreme learning machine to calculate the current tool wear amount value and the remaining service life value;

[0050] A fourth unit for adding the current tool wear amount value and the remaining service life value to a parallel agent structure, determining an initial adjustment scheme by combining the observed shared information collected by each agent, predicting the processing quality score corresponding to the initial adjustment scheme through a capsule network, and feeding the processing quality score back to the parallel agent structure as a reward signal to obtain an optimized correction strategy;

[0051] A fifth unit for adjusting the working parameters of the intelligent machine tool based on the optimized correction strategy and adding them to the machine tool process parameter library.

[0052] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including:

[0053] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the method described above.

[0054] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0055] In the present invention, by segmenting the working condition data through a sliding window, denoising and normalizing it, and combining a bidirectional gated recurrent unit and Hilbert-Huang transform for time-frequency feature extraction, various types of characteristic information during the tool wear process can be comprehensively captured, significantly improving the accuracy and stability of tool wear state recognition, reducing the misjudgment rate. An extreme learning machine is used to accurately predict the tool wear amount and remaining service life, and an initial adjustment plan is formulated through a parallel intelligent agent structure. At the same time, a capsule network is used for plan evaluation and quality prediction to form a closed-loop feedback optimization mechanism, realizing the intelligent dynamic adjustment of machining parameters, effectively extending the tool service life. The optimized working parameters are added to the machine tool process parameter library, establishing a knowledge accumulation mechanism of self-learning and continuous optimization, not only improving the current machining quality and efficiency, but also providing a reliable reference for parameter settings under similar working conditions, reducing material waste and energy consumption, and enhancing the adaptive machining ability of intelligent machine tools. Brief Description of the Drawings

[0056] Figure 1 It is a flow schematic diagram of the method for identifying the tool wear state and adjusting machining parameters of an intelligent machine tool according to an embodiment of the present invention;

[0057] Figure 2 It is a visible view of the capsule network hierarchical structure and routing mechanism of the method for identifying the tool wear state and adjusting machining parameters of an intelligent machine tool according to an embodiment of the present invention;

[0058] Figure 3 It is a simulation effect diagram of the capsule network routing optimization of the method for identifying the tool wear state and adjusting machining parameters of an intelligent machine tool according to an embodiment of the present invention. Detailed Embodiments

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0061] Figure 1 It is a flow schematic diagram of the method for identifying the tool wear state and adjusting machining parameters of an intelligent machine tool according to an embodiment of the present invention, as Figure 1 shown, and the method includes:

[0062] Collect the working condition data during the machining process of the intelligent machine tool and perform sliding window segmentation. Denoise and normalize the segmented working condition data to obtain the standard working condition data;

[0063] Extract the long-term and short-term time series feature sequences from the standard working condition data through a bidirectional gated recurrent unit. Input the long-term and short-term time series feature sequences into the Hilbert-Huang transform decomposition to obtain the intrinsic mode components. Combine the four-layer spatial pyramid pooling to extract the multi-scale frequency domain feature sequences. Calculate the mutual information entropy between the time series feature sequences and the frequency domain feature sequences. Determine the fusion weight coefficients according to the mutual information entropy values and perform weighting to obtain the tool wear state feature vector;

[0064] Input the tool wear state feature vector into the pre-trained extreme learning machine to calculate the current tool wear amount value and the remaining service life value;

[0065] Add the current tool wear amount value and the remaining service life value to the parallel agent structure. Combine the observed shared information collected by each agent to determine the initial adjustment plan. Predict the machining quality score corresponding to the initial adjustment plan through a capsule network. Use the machining quality score as a reward signal to feedback to the parallel agent structure to obtain the optimized correction strategy;

[0066] Adjust the working parameters of the intelligent machine tool based on the optimized correction strategy and add them to the machine tool process parameter library.

[0067] In an alternative embodiment,

[0068] Collect the working condition data during the machining process of the intelligent machine tool and perform sliding window segmentation. Denoise and normalize the segmented working condition data to obtain the standard working condition data, including:

[0069] Collect the working condition data during the machining process of the intelligent machine tool. The working condition data includes spindle current signal, spindle speed signal, feed speed signal, and cutting force signal;

[0070] Perform sliding window segmentation on the working condition data, and the overlap rate between adjacent windows is 50% to obtain the segmented working condition data sequence;

[0071] For the segmented working condition data sequence, perform multi-scale decomposition and reconstruction of the signal through the wavelet threshold method to remove noise. Combine the Butterworth low-pass filter to filter out the high-frequency interference components above 50% of the sampling frequency and perform maximum-minimum normalization processing on the filtered signal to obtain the standard working condition data.

[0072] Collect the working condition data during the machining process of intelligent machine tools. Real-time collect the spindle current signal through a sensor network, monitor the instantaneous change of the motor input current during the collection process and convert it into a standard current signal; collect the spindle speed signal, and obtain the actual rotation speed of the spindle through an encoder; collect the feed speed signal and record the relative movement speed of the tool relative to the workpiece; collect the cutting force signal, arrange a force sensor in the contact area between the tool and the workpiece, and obtain the dynamic cutting force during the cutting process.

[0073] Perform sliding window segmentation on the collected working condition data. Determine the window length and the sliding step size. The window length is set according to the signal characteristics and the sampling frequency, and the sliding step size is half of the window length, ensuring that the overlapping rate of adjacent windows is 50%. Slide the window over the entire data sequence in turn, and extract a data segment each time. All data segments form the segmented working condition data sequence. The overlapping design can avoid information breakage caused by window segmentation and enhance the continuity of the data.

[0074] Perform wavelet denoising on the segmented working condition data sequence. Select an orthogonal wavelet basis function to decompose the signal into wavelet coefficients of multiple scales, including low-frequency approximation coefficients and high-frequency detail coefficients. Determine the threshold function according to the signal energy distribution characteristics, perform threshold processing on the detail coefficients of different scales, retain the effective signal components and suppress the noise components. Use the soft threshold method to shrink the coefficients exceeding the threshold, and reconstruct the denoised signal using the processed wavelet coefficients.

[0075] Design a Butterworth low-pass filter for filtering. Select the filter order according to the signal spectrum characteristics, and determine the passband and stopband by setting the cut-off frequency to half of the sampling frequency. The filter has good amplitude-frequency characteristics, with a gentle attenuation in the passband and a rapid attenuation in the stopband, and can effectively filter out high-frequency interference components. Perform filtering on the signal after wavelet denoising to further improve the signal quality.

[0076] Perform maximum-minimum normalization on the filtered signal. Calculate the maximum and minimum values of each signal sequence respectively, and linearly map the original data to the interval from zero to one. Eliminate the dimensional difference and numerical range difference between different types of signals through normalization, so that all signals have the same numerical distribution characteristics, and obtain the standard working condition data.

[0077] Exemplarily, during the machining process of a certain numerical control machine tool, the sampling frequency is set to 1000 Hz, and the working condition data is continuously collected for 5 seconds. The amplitude range of the spindle current signal is 0 - 30 A, the spindle speed signal range is 0 - 3000 rpm, the feed speed signal range is 0 - 500 mm / min, and the cutting force signal range is 0 - 2000 N. The data is segmented using a 2048-point sliding window, with 1024 points overlapping between adjacent windows, resulting in a total of 4 data segments. The db4 wavelet is selected to decompose the signal into 4 layers, and the signal is reconstructed after removing noise using the soft threshold method. A 4th-order Butterworth low-pass filter with a cut-off frequency of 500 Hz is designed to filter the reconstructed signal. Finally, the filtered signal is normalized to obtain a standard working condition data sequence with amplitudes all within the range of [0, 1].

[0078] In this embodiment, by collecting four types of signal types for the working condition data, the multi-dimensional characteristics of the machine tool processing state are fully reflected. Compared with single-signal collection, the dynamic changes in the processing process can be characterized more comprehensively. The sliding window segmentation method is adopted. By setting an appropriate window length and overlap rate, both the real-time nature of signal analysis is ensured, and the information breakage caused by signal segmentation is avoided. By combining the wavelet threshold method and the Butterworth filter, multi-scale and multi-level noise suppression is achieved, significantly improving the signal-to-noise ratio of the signal.

[0079] In an alternative embodiment,

[0080] The long-term and short-term time series feature sequences are extracted from the standard working condition data through a bidirectional gated recurrent unit. The long-term and short-term time series feature sequences are input into the Hilbert-Huang transform for decomposition to obtain the intrinsic mode components, and multi-scale frequency domain feature sequences are extracted by combining four-layer spatial pyramid pooling, including:

[0081] The standard working condition data is input into the bidirectional gated recurrent unit network. Forward hidden layer states and backward hidden layer states are established in the bidirectional gated recurrent unit network. The forward hidden layer state and the backward hidden layer state are calculated and updated respectively through the reset gate and the update gate, and the updated forward hidden layer state and the backward hidden layer state are concatenated to generate the long-term and short-term time series feature sequences;

[0082] The long-term and short-term time series feature sequences are decomposed into intrinsic mode function components through empirical mode decomposition. The Hilbert transform is performed on the intrinsic mode function components to obtain the Hilbert transform components. An analytic signal is constructed based on the intrinsic mode function components and the Hilbert transform components, and the instantaneous amplitude and the instantaneous phase are extracted from the analytic signal to form the intrinsic mode components;

[0083] In the four-layer spatial pyramid pooling network, pooling scales of 1×1, 2×2, 4×4, and 8×8 are respectively set. The intrinsic mode components are input into the four-layer spatial pyramid pooling network to generate a pooled output feature map and perform a max pooling operation to obtain four-layer pooled features. The four-layer pooled features are concatenated to form a multi-scale frequency domain feature sequence.

[0084] The standard condition data is input into the bidirectional gated recurrent unit network. Meanwhile, two processing channels, a forward channel and a backward channel, are constructed. In the forward channel, the data is input sequentially in time series order, and the forgetting degree of historical information is dynamically adjusted through the reset gate to clear the historical information that contributes little to the current state prediction. The update gate adaptively adjusts the retention ratio of the historical state according to the importance of the current input to form a forward hidden layer state sequence. In the backward channel, the data is input sequentially in reverse time series order and processed using the same gating mechanism to generate a backward hidden layer state sequence. The state sequences of the two channels respectively capture the forward and backward time series dependencies of the data. The forward and backward hidden layer states are concatenated in the feature dimension to obtain a long short-term feature sequence that fuses bidirectional time series information.

[0085] Perform empirical mode decomposition on the long short-term feature sequence. Identify all local extreme points of the sequence, and respectively construct the upper envelope and the lower envelope through cubic spline interpolation. Calculate the mean of the upper and lower envelopes to obtain the local mean curve. Subtract the local mean curve from the original sequence to obtain a new sequence. Repeat the above process until the new sequence meets the definition conditions of the intrinsic mode function to obtain the first intrinsic mode function component. Subtract this component from the original sequence, and repeat the decomposition process for the remaining sequence to sequentially obtain all intrinsic mode function components.

[0086] Perform Hilbert transform on each intrinsic mode function component. Through the transform, the real signal is converted into the imaginary part of the analytic signal, and the analytic signal is constructed together with the original real signal. Extract the instantaneous amplitude from the analytic signal to characterize the amplitude feature of the signal at each moment; extract the instantaneous phase to characterize the phase change feature of the signal. Combine the extracted instantaneous amplitude and phase information to form the intrinsic mode components that characterize the time-frequency characteristics of the signal.

[0087] Construct a multi-layer spatial pyramid pooling network for feature extraction. Set the smallest scale pooling window in the first layer to retain the local detailed features of the signal. Increase the pooling window size in the second layer to extract the statistical features of the local area. Further expand the pooling range in the third layer to capture the medium-scale feature patterns. Use the maximum pooling window in the fourth layer to extract the global-scale feature expression. For each layer, input the intrinsic mode components into the current pooling layer, scan the feature map through a sliding window, and perform a maximum pooling operation within each window area to extract the most significant feature response in that area.

[0088] Each pooling layer processes independently to obtain feature maps with different resolutions. The pooling layer with a smaller scale retains more local details, while the pooling layer with a larger scale extracts more abstract global features. The feature maps output by the four pooling layers are dimensionally reduced to generate pooling feature vectors of different scales. Finally, these feature vectors are sequentially concatenated in the feature dimension to form a feature sequence containing multi-scale frequency domain information, realizing the feature expression of the signal at different time scales and frequency scales.

[0089] Exemplarily, during the machining process of a certain numerically controlled machine tool, the collected standard working condition data includes four sensor signals, and the length of each signal sequence is one thousand data points. Through the processing of a bidirectional gated recurrent unit network, with the number of hidden layer units set to one hundred and twenty-eight, a two-hundred-and-fifty-six-dimensional long-term and short-term time series feature sequence is extracted. The feature sequence is subjected to empirical mode decomposition to obtain eight intrinsic mode function components. Through Hilbert transform and analytic signal construction, the instantaneous amplitude and phase information of each component are obtained to form the intrinsic mode components. The intrinsic mode components are input into a four-layer spatial pyramid pooling network, and four groups of features are obtained after pooling operations of different scales, with the feature dimensions being sixty-four, sixteen, four, and one respectively. These features are concatenated to obtain an eighty-five-dimensional multi-scale frequency domain feature sequence.

[0090] In this embodiment, the bidirectional gated recurrent unit network is used to extract time series features, realizing the capture of the forward and backward dual dependence relationships of the working condition data, significantly enhancing the network's learning ability for long-term dependence relationships, and improving the accuracy and robustness of time series feature extraction. The empirical mode decomposition method is used to adaptively decompose the long-term and short-term feature sequence without presetting the decomposition basis function, and can perform modal decomposition according to the characteristics of the signal itself, realizing the effective decomposition of complex non-stationary signals. Combining Hilbert transform to construct an analytic signal, extracting the instantaneous amplitude and instantaneous phase information therefrom to form the intrinsic mode components, can comprehensively characterize the instantaneous characteristics and frequency change laws of the signal, and improve the integrity of feature expression.

[0091] In an alternative embodiment,

[0092] Calculating the mutual information entropy between the time series feature sequence and the frequency domain feature sequence, determining the fusion weight coefficient according to the mutual information entropy value and performing weighting to obtain the tool wear state feature vector includes:

[0093] By performing logarithmic multiplication and summation operations on the components of the time series feature sequence and the frequency domain feature sequence, the mutual information entropy value between the two feature sequences is calculated;

[0094] The difference between the eigenvalue in the time - series feature sequence and the frequency - domain feature sequence and the corresponding feature mean is divided by the bandwidth parameter to obtain the standardized eigenvalue. The standardized eigenvalue is substituted into the Gaussian kernel function, and the result of the Gaussian kernel function calculation is divided by the product of the number of samples and the bandwidth parameter to obtain the probability distribution value;

[0095] The negative of the mutual information entropy value is multiplied by the adjustment parameter and then exponentiated to obtain the numerator of the weight coefficient. The numerator of the weight coefficient is divided by the sum of the numerator of the weight coefficient and its complement to obtain the fusion weight coefficient of the time - series feature sequence. Subtract the fusion weight coefficient of the time - series feature sequence from 1 to obtain the fusion weight coefficient of the frequency - domain feature sequence;

[0096] Multiply the time - series feature sequence by the fusion weight coefficient of the time - series feature sequence to obtain the weighted time - series feature, multiply the frequency - domain feature sequence by the fusion weight coefficient of the frequency - domain feature sequence to obtain the weighted frequency - domain feature, and add the weighted time - series feature and the weighted frequency - domain feature to obtain the fusion feature sequence;

[0097] Extract the eigenvector corresponding to the maximum eigenvalue in the fusion feature sequence as the main feature component to construct the tool wear state feature vector.

[0098] Calculate the mutual information entropy of the time - series feature sequence and the frequency - domain feature sequence. The two feature sequences are put in one - to - one correspondence according to the dimension, and each pair of feature components is extracted for processing. Take the logarithm of each pair of feature components respectively, and multiply them to get the product term. Calculate the product terms of all corresponding components in turn, and sum up these product terms to obtain the mutual information entropy value representing the statistical correlation of the two feature sequences. The mutual information entropy value reflects the degree of information sharing between the two feature sequences. The larger the entropy value, the stronger the mutual dependence between the feature sequences.

[0099] Perform feature standardization and probability distribution estimation. Calculate the feature mean of each dimension for the time - series feature sequence and the frequency - domain feature sequence respectively. Subtract the corresponding mean from each dimension eigenvalue to obtain the feature deviation value. Select an appropriate bandwidth parameter, and divide the feature deviation value by the bandwidth parameter for normalization to obtain the standardized eigenvalue. Substitute the standardized eigenvalue into the Gaussian kernel function to calculate the kernel function response value. Divide the response value by the product of the total number of samples and the bandwidth parameter to obtain the probability distribution estimate of the eigenvalue. This kernel - based probability density estimation method can effectively capture the non - linear characteristics of the feature distribution.

[0100] Calculate the adaptive weight coefficient for feature fusion. Take the negative of the mutual information entropy value obtained previously, multiply it by a preset adjustment parameter, and perform an exponential operation on the product result to obtain the numerator term of the weight coefficient. Calculate the sum of this numerator term and its complement as the denominator, and divide the numerator by the denominator to obtain the fusion weight coefficient of the time series feature sequence. The fusion weight coefficient reflects the importance of the time series feature in the fusion process. Subtract the weight coefficient of the time series feature from the unit weight to obtain the fusion weight coefficient of the frequency domain feature sequence. This weight allocation method based on mutual information entropy can adaptively adjust the fusion ratio according to the correlation between feature sequences.

[0101] Perform the weighted feature fusion operation. Multiply each feature component in the time series feature sequence by the time series feature weight coefficient to obtain the weighted time series feature sequence. Similarly, multiply each feature component in the frequency domain feature sequence by the frequency domain feature weight coefficient to obtain the weighted frequency domain feature sequence. Perform an addition operation on the two weighted feature sequences in the corresponding dimensions to generate the fused feature sequence. This weighted fusion strategy not only retains the important information components in each feature sequence but also can suppress the influence of redundant information and noise.

[0102] Extract the main feature components to construct the state feature vector. Perform eigenvalue decomposition on the fused feature sequence, and calculate the eigenvalues and the corresponding eigenvectors. Sort all the eigenvalues from largest to smallest, and select the eigenvector corresponding to the largest eigenvalue as the tool wear state feature vector. This feature vector reflects the most significant change pattern in the fused feature sequence and contains the main feature information of the tool wear state.

[0103] Exemplarily, in a certain CNC milling process, a 256-dimensional time series feature sequence and an 85-dimensional frequency domain feature sequence are obtained. The mutual information entropy value of the two sequences is calculated to be 0.8. Feature standardization is performed using a bandwidth parameter of 0.1, and the feature probability distribution is estimated through the Gaussian kernel function. The adjustment parameter is set to 2, and the fusion weight of the time series feature sequence is calculated to be 0.6, and the fusion weight of the frequency domain feature sequence is 0.4. The two feature sequences are weighted and fused to obtain a 300-dimensional fused feature sequence. The eigenvector corresponding to the largest eigenvalue is extracted as a 60-dimensional tool wear state feature vector.

[0104] In this embodiment, by calculating the mutual information entropy between the time series feature sequence and the frequency domain feature sequence, the correlation strength between the two types of features is quantitatively evaluated. The kernel density estimation method is used to model the feature distribution. By performing a non-linear transformation on the mutual information entropy, the fusion weights of the time series feature and the frequency domain feature are automatically calculated, and the main feature components are extracted through eigenvalue decomposition, realizing the dimensionality reduction and refinement of the features;

[0105] In the prior art, the method for extracting the tool wear state features usually analyzes using a single time - series feature or frequency - domain feature, which is difficult to comprehensively reflect the dynamic change process of tool wear. At the same time, using simple feature splicing or fixed - weight fusion cannot adaptively adjust according to the importance of features, resulting in insufficient accuracy and robustness of feature expression;

[0106] This embodiment realizes the complementary advantages of time - series features and frequency - domain features, improves the comprehensiveness of feature expression, introduces mutual information entropy as a measure of feature correlation, enhances the rationality of feature fusion, adopts an adaptive weight allocation mechanism, improves the flexibility of feature fusion, realizes dimensionality reduction and optimization of features through main feature extraction, and improves the efficiency of feature expression, providing reliable feature support for the accurate identification and prediction of tool wear states.

[0107] In an alternative embodiment,

[0108] Inputting the tool wear state feature vector into a pre - trained extreme learning machine and calculating the current tool wear amount value and remaining service life value includes:

[0109] Collecting the tool wear state sample data corresponding to the tool wear state feature vector, dividing the tool wear state sample data into a training data set and a test data set, training the extreme learning machine based on the training data set. The extreme learning machine includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is the dimension of the tool wear state feature vector. The hidden layer uses the Sigmoid activation function. By randomly generating the initial values of the input weight matrix and the bias vector, multiplying the tool wear state feature vector in the training data set by the input weight matrix and adding the bias vector, and passing through the Sigmoid activation function to obtain the hidden - layer output, calculating the output weight based on the hidden - layer output to obtain the pre - trained extreme learning machine;

[0110] Inputting the tool wear state feature vector into the pre - trained extreme learning machine, multiplying the input weight matrix by the tool wear state feature vector and adding the bias vector, then passing through the Sigmoid activation function to obtain the hidden - layer output, and calculating the current tool wear amount value and remaining service life value according to the hidden - layer output.

[0111] Collecting the tool wear state sample data corresponding to the tool wear state feature vector during the machining process through multiple sensors. The collection process covers the entire life cycle from the initial state of the tool to the fully worn state to ensure the integrity and representativeness of the sample data. Pre - processing the collected raw data to eliminate outliers and noise interference and improve the data quality. Randomly dividing the processed sample data set into a training data set and a test data set according to a preset ratio to ensure the consistency of the sample distribution in the two data sets.

[0112] Construct the topological structure of the extreme learning machine network. The number of nodes in the input layer is consistent with the dimension of the tool wear state feature vector to ensure that the network can completely receive all feature information. An appropriate number of neuron nodes are set in the hidden layer, and the neuron nodes realize the transformation and mapping of the feature space through non-linear activation functions. A continuous differentiable saturation activation function is selected. The saturation activation function has good non-linear characteristics and numerical stability, and can effectively extract the complex relationships between features. Two nodes are set in the output layer, which are used to predict the current wear amount and remaining service life of the tool respectively.

[0113] Initialize the network parameters. Use a random number generator to generate random numbers that satisfy a specific distribution, and construct the weight matrix from the input layer to the hidden layer. Each element of the weight matrix represents the importance of the corresponding input feature. Generate the bias vector of the hidden layer nodes to adjust the activation threshold of each hidden layer node. This random initialization strategy can enable the network to obtain a better initial state in the parameter space and avoid falling into symmetric solutions.

[0114] Execute the model training process. Input the feature vectors in the training dataset into the network in sequence. The feature vectors perform matrix multiplication operations with the input weight matrix to achieve linear transformation of the features. Add the transformed result to the bias vector to obtain the input value of each hidden layer node. Each input value undergoes non-linear mapping through the activation function to generate the actual output value of the hidden layer node. Collect the output values of the hidden layer of all training samples to construct the hidden layer output matrix.

[0115] Based on the hidden layer output matrix and the target values in the training dataset, use the least squares method to calculate the output weight matrix. The direct solution method avoids the process of repeated iterative optimization required by traditional neural networks and significantly improves the training efficiency. The calculated output weight matrix contains the optimal mapping relationship from the hidden layer features to the output targets. So far, the pre-training process of the extreme learning machine is completed, and the network has obtained all the parameters required to predict the tool wear state.

[0116] Use the trained model for online prediction. Input the feature vector of the tool wear state obtained in real time into the network, and perform feature transformation through the trained input weight matrix. Add the transformed features to the bias vector to form the input of the hidden layer nodes. Each hidden layer node generates an output value through the activation function to form the hidden layer feature expression of the current state. Multiply the hidden layer features by the output weight matrix to obtain the final prediction result. The first output node gives the numerical value of the current wear amount of the tool, and the second output node gives the numerical value of the predicted remaining service life.

[0117] Exemplarily, for a certain CNC milling process, one thousand sets of tool wear state sample data were collected. The sample data was divided into a training set and a test set according to a ratio of seven to three. An extreme learning machine network with an input layer dimension of sixty and one hundred and twenty hidden layer nodes was constructed. The initial values of the input weight matrix and bias vector were generated by a random method. The training data was input into the network, and the hidden layer output was obtained through matrix operations and activation functions, and the output weight matrix was calculated to complete the model training. The newly collected wear state feature vector was predicted, and the current tool wear amount was output as 0.3 mm, and the remaining service life was predicted to be eight hours.

[0118] In this embodiment, by reasonably dividing the tool wear state sample data, a data set containing sufficient training samples and test samples was established, ensuring that the evolution law of the tool wear state could be fully learned during the model training process. At the same time, the prediction performance of the model was verified through the test data set, ensuring the reliability of the prediction results. The extreme learning machine network structure adopted was reasonably designed. The input layer completely received the tool wear state feature information. The hidden layer realized the deep mapping and extraction of features through a non-linear activation function. The output layer directly gave the predicted values of the wear amount and remaining life, effectively capturing the complex relationship between the features and the wear state. The initial values of the input weight matrix and bias vector were determined by a random generation method, providing a good starting point for the network to learn.

[0119] In an alternative embodiment,

[0120] The current tool wear amount value and the remaining service life value are added to the parallel agent structure. The initial adjustment plan is determined by combining the observation shared information collected by each agent. The processing quality score corresponding to the initial adjustment plan is predicted through a capsule network, and the processing quality score is fed back to the parallel agent structure as a reward signal to obtain an optimized correction strategy, including:

[0121] The current tool wear amount value and the remaining service life value are added to the parallel agent structure. The parallel agent structure includes multiple parallel agents. Each parallel agent collects local observation shared information, and the local observation shared information is interacted through a shared network to obtain global observation shared information;

[0122] According to the global observation shared information, the attention weight coefficient between the parallel agents is calculated. The attention weight coefficient is obtained through the inner product operation of the query vector and the key vector corresponding to the parallel agent. The attention weight coefficient is weighted and accumulated with the corresponding value vector to obtain a fused feature representation. Based on the fused feature representation and the observation shared information collected by each parallel agent, an initial adjustment plan is determined;

[0123] Input the initial adjustment plan into the capsule network, transform the initial adjustment plan through the transformation matrix of the capsule network to obtain the output of the initial capsule layer, calculate the routing weights based on the output of the initial capsule layer, obtain the output capsules of the capsule network through the compression function based on the product of the routing weights and the output of the initial capsule layer, and predict the processing quality score corresponding to the initial adjustment plan based on the output capsules;

[0124] Use the processing quality score as a reward signal and feedback it to the parallel agent structure. The parallel agent structure encodes the processing quality score to obtain the historical information encoding, inputs the historical information encoding and the currently acquired current state encoding into the policy network to generate the probability distribution of the adjustment actions, and combines the policy optimization criterion to optimize and obtain the optimized correction policy.

[0125] Construct a collaborative optimization structure composed of multiple parallel agents, and add the predicted tool wear amount value and remaining service life value to the state space of each agent. Each parallel agent is equipped with multiple sensors to continuously collect local observation information in the area, including multi-dimensional processing data such as cutting force, vibration, temperature, and acoustic emission. Each agent establishes an information interaction channel through a shared communication network, encodes and transmits the collected local observation information to achieve two-way information flow. Each agent receives the observation information sent by other agents, integrates this information with the data collected by itself, and constructs an observation information matrix reflecting the global processing state.

[0126] On the basis of obtaining the global observation information, construct an attention calculation mechanism among agents. Extract and map the state information of each agent to generate a query vector representing the characteristics of the agent. At the same time, map the state information of other agents to the corresponding key vectors. Calculate the attention score representing the degree of association among agents through the inner product operation of the query vector and the key vector. Normalize the attention score to obtain the attention weight coefficient reflecting the importance of each agent. Perform a weighted summation operation on the attention weight coefficient and the value vector corresponding to the agent to obtain the fused feature representation. Based on the fused feature representation, combined with the real-time observation information collected by each agent, comprehensively evaluate the current processing state and formulate an initial adjustment plan including parameters such as feed speed and spindle speed.

[0127] The initial adjustment plan is input into a specially designed capsule network for in-depth optimization. The capsule network first performs feature transformation and reconstruction on the adjustment plan through a transformation matrix, extracts the key attribute information in the plan, and generates the feature vectors of the initial capsule layer. The dynamic routing algorithm is used to calculate the association strength between different feature capsules, and the routing weights representing the importance of features are obtained. The routing weights are weighted and combined with the feature vectors of the initial capsule layer, and normalized through a non-linear compression function to obtain the output capsules containing optimization suggestions. The information in the output capsules is parsed to predict the processing quality score after the implementation of the initial adjustment plan.

[0128] The predicted processing quality score is used as an evaluation signal and fed back to the parallel agent structure. After receiving the quality score, the parallel agent structure performs temporal encoding on the historical adjustment plans and their corresponding effects, and constructs a long-term memory containing historical decision-making experience. At the same time, the current processing state is encoded in real time to generate a short-term memory reflecting the immediate situation. The encoded information of the long-term memory and the short-term memory is jointly input into the policy network, and through the mapping and transformation of the multi-layer neural network, the probability distribution of each possible adjustment action is generated.

[0129] Based on the policy gradient method, the probability distribution is optimized. The expected return under the current policy is calculated, and compared with the baseline value to obtain the advantage function. According to the gradient information of the advantage function, the network parameters are updated according to the policy optimization criterion, and the action probability distribution is adjusted. Through multiple rounds of iterative optimization, the decision-making ability of the policy network is continuously improved, and an optimized correction policy with rich experience and high reliability is obtained.

[0130] Exemplarily, in a certain CNC machining workshop, five parallel agents are deployed to monitor the tool machining state. When it is detected that the tool wear amount reaches 0.3 mm and the remaining service life is expected to be 6 hours, each agent collects local machining data and shares it. The interaction weights between the agents are calculated through the attention mechanism, and the fused features are generated and an initial adjustment plan for the feed speed and spindle speed is formulated. The adjustment plan is input into the capsule network for optimization, and the predicted processing quality score is 95 points. Based on this score, reinforcement learning is performed to update the policy network parameters, and an optimized machining parameter correction policy is generated to realize the intelligent control of the machining process.

[0131] In this embodiment, through the distributed agent network, machining state information can be collected from multiple dimensions, avoiding the information loss and misjudgment problems that may be caused by single-point monitoring. By calculating the attention weights between the agents, key information can be adaptively identified and extracted, effectively filtering out redundant and noise interference. The capsule network can capture the hierarchical relationship and spatial features in the adjustment plan, and through the dynamic routing mechanism, an accurate evaluation of the feasibility of the plan is realized;

[0132] In the prior art, the monitoring of the tool wear state and the optimization of machining parameters mainly rely on a single agent for decision-making, which has problems such as incomplete information acquisition, single optimization strategy, and poor adaptability. In terms of information fusion and decision optimization, simple linear weighting or rule matching is adopted, which cannot effectively handle complex and changeable machining environments, resulting in unsatisfactory optimization effects.

[0133] In this embodiment, the information sharing mechanism among agents ensures the real-time grasp of the global state, provides a complete and reliable data basis for subsequent optimization decisions. The intelligent information fusion method significantly improves the accuracy of state perception and the reliability of decision-making. The depth evaluation mechanism effectively reduces the decision-making risk in the optimization process and improves the accuracy of parameter adjustment. The adaptive learning mechanism enables the system to effectively respond to various changes in the machining process and maintain continuous optimization effects, providing a more advanced technical solution for realizing the intelligent control of the machining process.

[0134] Figure 2 It is a viewable diagram of the capsule network hierarchical structure and routing mechanism for the tool wear state recognition and machining parameter adjustment method of the intelligent machine tool in the embodiment of the present invention, showing the complete working process of the capsule network hierarchical structure and routing mechanism adopted in this technical solution. This structure consists of five main parts: an input layer, an initial capsule layer, a transformation matrix and routing mechanism, an output capsule layer, and a prediction result layer.

[0135] In the input layer, the system collects seven types of key machining data, including the three-dimensional components of cutting force (X, Y, Z), vibration acceleration signal, temperature distribution, acoustic emission signal, current load, and the current tool wear amount (0.28 mm) and remaining service life (6.5 h). These multi-dimensional heterogeneous data serve as the initial input of the network and provide a basis for subsequent intelligent analysis.

[0136] The initial capsule layer consists of six capsule units. Each capsule unit represents features using an 8-dimensional vector, respectively capturing key information such as cutting force patterns, vibration spectra, temperature gradients, acoustic emission energy, current fluctuations, and wear trends. Different from traditional convolutional neural networks, the capsule network represents features using vectors rather than scalars, and can retain richer spatial hierarchical relationship information.

[0137] In the transformation matrix and routing mechanism section, the feature vectors of the initial capsules are linearly transformed through the feature transformation matrix W[i, j] to generate "prediction vectors". Subsequently, through the dynamic routing algorithm, three key processes are realized: feature extraction, compression activation, and feature integration. The routing algorithm iterates 3 times, and determines the importance of different features by calculating weight coefficients, realizing the effective aggregation of features.

[0138] The output capsule layer contains four dedicated capsules, corresponding to feed rate optimization (confidence 0.93), spindle speed adjustment (confidence 0.89), cutting depth control (confidence 0.87), and coolant parameter optimization (confidence 0.84) respectively. Each capsule obtains information from the previous layer through dynamic routing and generates optimization suggestions for specific parameters.

[0139] The prediction result layer shows the final output of the model, including the machining quality score (95.3), wear prediction (current 0.28mm, 0.31mm after 1 hour, 0.38mm after 4 hours), life prediction (remaining 6.5 hours, 8.2 hours after optimization), and specific parameter adjustment suggestions (feed rate 165mm / min, spindle speed 1150rpm).

[0140] The capsule network structure of this technical solution can more effectively retain the spatial hierarchical relationship in the cutting process, significantly improve the accuracy of tool life prediction (94.8%) and the effect of machining parameter optimization, and effectively extend the service life of the tool.

[0141] In an alternative embodiment,

[0142] The output capsule of the capsule network is obtained by compressing the product of the routing weight and the output of the initial capsule layer. Predicting the machining quality score corresponding to the initial adjustment plan based on the output capsule includes:

[0143] Construct the routing weight in the capsule network and the output of the initial capsule layer as nodes in the Markov random field structure, calculate the edge connection probability based on the spatial position relationship and feature similarity between the nodes, construct the edge connection probability as a potential function, construct a single-node potential function based on the distribution parameter for the nodes, and construct the joint probability distribution of the node states through the product of the single-node potential function and the potential function;

[0144] Based on the joint probability distribution, initialize the state values of the nodes according to Gibbs sampling, calculate the conditional probability of the nodes based on the single-node potential function and the potential function, iteratively update the node states according to the conditional probability to obtain a sampling sequence, monitor the autocorrelation of the sampling sequence and dynamically adjust the sampling step size and the number of iterations until the preset convergence condition is met;

[0145] Obtain the optimized routing weight by calculating the marginal probability distribution of the sampling sequence, multiply the optimized routing weight by the output of the initial capsule layer to obtain a product result, and input the product result into the compression function to obtain the output capsule of the capsule network;

[0146] Calculate the inner product of the output capsule and a preset parameter vector to obtain an initial score. Based on the initial score, obtain a normalized score through a non-linear activation function and use it as the processing quality score corresponding to the initial adjustment scheme.

[0147] Map the routing weights and the output of the initial capsule layer in the capsule network to nodes in a random field. Extract multi-dimensional feature vectors for each node, including spatial coordinate information, topological structure features, numerical distribution features, etc. By calculating the Euclidean distance and cosine similarity between nodes, evaluate the spatial correlation and feature similarity between nodes. Based on the evaluation results, construct an edge connection probability matrix, which describes the connection strength between nodes. Convert the edge connection probability into a potential function to characterize the interaction relationship between nodes.

[0148] For each independent node, construct a single-node potential energy function based on its feature distribution parameters. The single-node potential energy function takes into account multiple factors such as the prior probability distribution of the node, numerical range constraints, gradient change features, etc. Multiply the single-node potential energy function and the edge connection potential function probabilistically to obtain a complete joint probability distribution model of node states, which fully describes the state combinations of all nodes in the random field and their occurrence probabilities.

[0149] Use the Gibbs sampling method to optimize and update the node states. Randomly initialize the state values of each node according to the prior distribution of the node. In each round of sampling, fix the states of other nodes and calculate the conditional probability distribution of the target node based on the single-node potential energy function and the potential functions of adjacent nodes. Draw a new state value from the conditional probability distribution and update the state of the target node. According to the preset node traversal order, sequentially update the states of all nodes to generate a complete state sampling sequence.

[0150] During the sampling process, calculate the autocorrelation coefficient of the sampling sequence in real time to monitor the mixing degree of the sampling chain. When the autocorrelation coefficient is high, appropriately increase the sampling step size to accelerate the exploration speed of the state space. When the autocorrelation coefficient decreases, reduce the sampling step size to improve the fineness of local search. At the same time, dynamically adjust the number of iterations according to the change trend of the autocorrelation coefficient until the autocorrelation of the sampling sequence drops below a preset threshold, indicating that the sampling has reached a stable state.

[0151] Conduct statistical analysis on the obtained stable sampling sequence, calculate the marginal probability distribution of each node state. Determine the optimal state value of the node through the marginal probability distribution to obtain the optimized routing weights. Perform a tensor multiplication operation on the optimized routing weights and the output of the initial capsule layer to obtain a weighted feature representation. The weighted feature representation retains the important information of the original features and highlights the contribution of key features through the optimized weights.

[0152] Input the weighted feature representation into a specially designed compression function. The compression function adopts a non-linear transformation method to map the features to a new representation space. Through the processing of the compression function, output capsules with good feature expression ability are obtained. The output capsules contain optimized feature information and can effectively express the key features of the current processing state.

[0153] Perform machining quality scoring. Calculate the inner product of the output capsule and the pre-trained parameter vector to obtain the initial quality score. This score reflects the matching degree between the current features and the ideal state. Input the initial score into a customized non-linear activation function and obtain the final machining quality score through normalization processing.

[0154] Exemplarily, in a certain CNC milling process, the capsule network contains thirty routing weight nodes and forty initial capsule layer output nodes. Construct these nodes into a Markov random field structure and calculate the edge connection probability based on the spatial distribution characteristics of the nodes. Perform one thousand iterations of update through Gibbs sampling and stop sampling when the autocorrelation coefficient of the sampling sequence is lower than the preset threshold. Statistically analyze the marginal distribution of the sampling sequence to obtain the optimized routing weights, multiply them with the output of the initial capsule layer, and obtain the output capsule through the compression function. Finally, calculate the inner product of the output capsule and the preset parameter vector and obtain the scoring result reflecting the machining quality through normalization processing.

[0155] In this embodiment, by constructing the routing weights and capsule outputs into a Markov random field structure, a probabilistic dependence relationship model between nodes is established. The potential function is used to describe the interaction relationship between nodes, which includes both the feature potential energy of single nodes and the edge connection potential energy between nodes. The intelligent optimization of node states is realized through the Gibbs sampling method. By dynamically adjusting the sampling step size and monitoring the autocorrelation, the convergence and effectiveness of the sampling process are ensured;

[0156] In the prior art, the routing mechanism of the capsule network mainly adopts a simple dynamic routing algorithm and is optimized by directly iteratively updating the routing weights, ignoring the spatial correlation between the routing weights, resulting in the optimization process being prone to falling into local optima, lacking probabilistic modeling of the feature distribution, making the weight update lack theoretical support, the convergence of the optimization process difficult to guarantee, and the reliability of the optimization result insufficient;

[0157] This embodiment fully considers the spatial correlation and numerical similarity between features, enabling the optimization process to take into account both local features and global structures simultaneously, significantly improving the accuracy and integrity of feature extraction. The double-layer potential energy structure provides a clear theoretical basis for the optimization process, enabling the explanation and guidance of the weight update process from a probabilistic perspective, greatly enhancing the interpretability and reliability of the optimization results. The adaptive sampling strategy not only improves the optimization efficiency but also effectively avoids the risk of falling into local optima, ensuring the global optimality of the optimization results, solving the problems existing in traditional routing mechanisms, and providing new technical ideas for the performance optimization of capsule networks in complex application scenarios, with important theoretical value and practical significance.

[0158] Figure 3 This is the simulation effect diagram of the capsule network routing optimization for the intelligent machine tool tool wear state recognition and machining parameter adjustment method in the embodiment of the present invention, showing the comparison of the surface topography simulation effects of three different technical solutions in CNC milling. In the main figure, the horizontal axis represents the machining distance (0 - 50 mm), and the vertical axis represents the surface height deviation (±3.0 μm). The machining surface generated by this technical solution (solid line) is extremely close to the ideal contour. For easy observation, its trajectory has been slightly offset to Y = 198 μm, and a local enlarged view is provided in the upper right corner. The enlarged view clearly shows that the fluctuation amplitude of this solution is only ±0.08 μm, the surface roughness Ra = 0.08 μm, and the machining quality score is as high as 0.98. In contrast, the dynamic routing algorithm (short dashed line) has obvious periodic ripples, with a fluctuation amplitude of ±1.25 μm and a score of only 0.65; the EM routing algorithm (dotted dashed line) has the worst surface quality, with a fluctuation amplitude as high as ±2.35 μm and a score as low as 0.48.

[0159] This technical solution improves the surface smoothness by approximately 16 times compared to the traditional dynamic routing algorithm and by approximately 30 times compared to the EM routing algorithm, fully verifying the excellent performance of the capsule network routing optimization mechanism based on Markov random fields in the field of high-precision machining.

[0160] In the second aspect of the embodiment of the present invention, there is provided an intelligent machine tool tool wear state recognition and machining parameter adjustment system, including:

[0161] A first unit for collecting the working condition data during the machining process of the intelligent machine tool, performing sliding window segmentation, and performing denoising and normalization processing on the segmented working condition data to obtain standard working condition data;

[0162] The second unit is used to extract long - term and short - term time - series feature sequences from standard working condition data through a bidirectional gated recurrent unit, input the long - term and short - term time - series feature sequences into Hilbert - Huang transform decomposition to obtain intrinsic mode components, combine four - layer spatial pyramid pooling to extract multi - scale frequency - domain feature sequences, calculate the mutual information entropy between the time - series feature sequences and the frequency - domain feature sequences, determine the fusion weight coefficient according to the mutual information entropy value and perform weighting to obtain the tool wear state feature vector;

[0163] The third unit is used to input the tool wear state feature vector into a pre - trained extreme learning machine to calculate the current tool wear amount value and the remaining service life value;

[0164] The fourth unit is used to add the current tool wear amount value and the remaining service life value to a parallel agent structure, determine an initial adjustment plan by combining the observed shared information collected by each agent, predict the processing quality score corresponding to the initial adjustment plan through a capsule network, and feedback the processing quality score as a reward signal to the parallel agent structure to obtain an optimized correction strategy;

[0165] The fifth unit is used to adjust the working parameters of the intelligent machine tool based on the optimized correction strategy and add them to the machine tool process parameter library.

[0166] In the third aspect of the embodiments of the present invention, an electronic device is provided, including:

[0167] A processor and a memory for storing processor - executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the method described above.

[0168] In the fourth aspect of the embodiments of the present invention, a computer - readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0169] The present invention can be a method, a device, a system, and / or a computer program product. The computer program product can include a computer - readable storage medium, on which computer - readable program instructions for executing various aspects of the present invention are uploaded.

[0170] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the wear state of a tool of an intelligent machine tool and adjusting machining parameters, characterized in that Including: Collect the working condition data during the machining process of the intelligent machine tool and perform sliding window segmentation. Denoise and normalize the segmented working condition data to obtain standard working condition data; Extract the long-term and short-term time series feature sequences from the standard working condition data through a bidirectional gated recurrent unit. Input the long-term and short-term time series feature sequences into the Hilbert-Huang transform decomposition to obtain the intrinsic mode components. Combine the four-layer spatial pyramid pooling to extract the multi-scale frequency domain feature sequences. Calculate the mutual information entropy between the time series feature sequences and the frequency domain feature sequences. Determine the fusion weight coefficients according to the mutual information entropy values and perform weighting to obtain the tool wear state feature vector; Input the tool wear state feature vector into the pre-trained extreme learning machine to calculate the current tool wear amount value and the remaining service life value; Add the current tool wear amount value and the remaining service life value of the tool to the parallel agent structure. Combine the observed shared information collected by each agent to determine the initial adjustment plan. Predict the machining quality score corresponding to the initial adjustment plan through the capsule network. Use the machining quality score as a reward signal to feedback to the parallel agent structure to obtain the optimized correction strategy; Adjust the working parameters of the intelligent machine tool based on the optimized correction strategy and add them to the machine tool process parameter library.

2. The method according to claim 1, characterized in that Collect the working condition data during the machining process of the intelligent machine tool and perform sliding window segmentation. Denoise and normalize the segmented working condition data to obtain standard working condition data, including: Collect the working condition data during the machining process of the intelligent machine tool, and the working condition data includes spindle current signal, spindle speed signal, feed speed signal and cutting force signal; Perform sliding window segmentation on the working condition data, and the overlap rate between adjacent windows is 50% to obtain the segmented working condition data sequence; For the segmented working condition data sequence, perform multi-scale decomposition and reconstruction of the signal through the wavelet threshold method to remove noise. Combine the Butterworth low-pass filter to filter out the high-frequency interference components higher than 50% of the sampling frequency and perform maximum-minimum normalization processing on the filtered signal to obtain the standard working condition data.

3. The method according to claim 1, characterized in that, Extract the long-term and short-term time series feature sequences from the standard working condition data through a bidirectional gated recurrent unit. Input the long-term and short-term time series feature sequences into the Hilbert-Huang transform decomposition to obtain the intrinsic mode components. Combine the four-layer spatial pyramid pooling to extract the multi-scale frequency domain feature sequences, including: Input the standard working condition data into the bidirectional gated recurrent unit network. Establish the forward hidden layer state and the backward hidden layer state in the bidirectional gated recurrent unit network. Calculate and update the forward hidden layer state and the backward hidden layer state through the reset gate and the update gate respectively. Concatenate the updated forward hidden layer state and the backward hidden layer state to generate the long-term and short-term time series feature sequences; Obtain the intrinsic mode function components by performing empirical mode decomposition on the long-term and short-term time series feature sequences. Perform the Hilbert transform on the intrinsic mode function components to obtain the Hilbert transform components. Construct the analytical signal based on the intrinsic mode function components and the Hilbert transform components. Extract the instantaneous amplitude and the instantaneous phase from the analytical signal to form the intrinsic mode components; In the four-layer spatial pyramid pooling network, pooling scales of 1×1, 2×2, 4×4, and 8×8 are respectively set. The intrinsic mode components are input into the four-layer spatial pyramid pooling network to generate a pooled output feature map and perform a max pooling operation to obtain four-layer pooled features. The four-layer pooled features are concatenated to form a multi-scale frequency domain feature sequence.

4. The method according to claim 1, characterized in that, Calculate the mutual information entropy between the time series feature sequence and the frequency domain feature sequence, determine the fusion weight coefficient according to the mutual information entropy value, and perform weighting to obtain the tool wear state feature vector, including: Calculate the mutual information entropy value between the two feature sequences by performing the operation of multiplying the logarithms of each component of the time series feature sequence and the frequency domain feature sequence and then summing. Divide the difference between the eigenvalue in the time series feature sequence and the frequency domain feature sequence and the corresponding feature mean by the bandwidth parameter to obtain the normalized eigenvalue. Substitute the normalized eigenvalue into the Gaussian kernel function, and divide the result calculated by the Gaussian kernel function by the product of the sample number and the bandwidth parameter to obtain the probability distribution value. Multiply the negative value of the mutual information entropy value by the adjustment parameter and perform an exponential operation to obtain the numerator of the weight coefficient. Divide the numerator of the weight coefficient by the sum of the numerator of the weight coefficient and its complement to obtain the fusion weight coefficient of the time series feature sequence. Subtract the fusion weight coefficient of the time series feature sequence from 1 to obtain the fusion weight coefficient of the frequency domain feature sequence. Multiply the time series feature sequence by the fusion weight coefficient of the time series feature sequence to obtain the weighted time series feature. Multiply the frequency domain feature sequence by the fusion weight coefficient of the frequency domain feature sequence to obtain the weighted frequency domain feature. Add the weighted time series feature and the weighted frequency domain feature to obtain the fusion feature sequence. Extract the feature vector corresponding to the maximum eigenvalue in the fusion feature sequence as the main feature component, and construct the tool wear state feature vector.

5. The method according to claim 1, wherein Input the tool wear state feature vector into the pre-trained extreme learning machine, and calculate the current tool wear amount value and the remaining service life value, including: Collect the tool wear state sample data corresponding to the tool wear state feature vector, divide the tool wear state sample data into a training data set and a test data set, and train the extreme learning machine based on the training data set. The extreme learning machine includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is the dimension of the tool wear state feature vector. The hidden layer uses the Sigmoid activation function. By randomly generating the initial values of the input weight matrix and the bias vector, multiply the tool wear state feature vector in the training data set by the input weight matrix and add the bias vector, and obtain the hidden layer output through the Sigmoid activation function. Calculate the output weight based on the hidden layer output to obtain the pre-trained extreme learning machine. Input the tool wear state feature vector into the pre-trained extreme learning machine, multiply the tool wear state feature vector by the input weight matrix and add the bias vector, and then obtain the hidden layer output through the Sigmoid activation function. Calculate the current tool wear amount value and the remaining service life value according to the hidden layer output.

6. The method according to claim 1, wherein Add the current tool wear value and the remaining service life value to the parallel agent structure, determine the initial adjustment plan by combining the observation sharing information collected by each agent, predict the machining quality score corresponding to the initial adjustment plan through a capsule network, and feedback the machining quality score as a reward signal to the parallel agent structure to obtain an optimized correction strategy, including: Add the current tool wear value and the remaining service life value to the parallel agent structure, which includes multiple parallel agents. Each parallel agent collects local observation sharing information and exchanges the local observation sharing information through a shared network to obtain global observation sharing information; Calculate the attention weight coefficient between the parallel agents according to the global observation sharing information. The attention weight coefficient is obtained through the inner product operation of the query vector and the key vector corresponding to the parallel agent. Perform weighted accumulation on the attention weight coefficient and the corresponding value vector to obtain a fused feature representation, and determine the initial adjustment plan based on the fused feature representation and the observation sharing information collected by each parallel agent; Input the initial adjustment plan into the capsule network, perform transformation on the initial adjustment plan through the transformation matrix of the capsule network to obtain the output of the initial capsule layer, calculate the routing weight based on the output of the initial capsule layer, and obtain the output capsule of the capsule network through the product of the routing weight and the output of the initial capsule layer after passing through a compression function. Predict the machining quality score corresponding to the initial adjustment plan based on the output capsule; Feedback the machining quality score as a reward signal to the parallel agent structure. The parallel agent structure encodes the machining quality score to obtain a historical information encoding, and inputs the historical information encoding and the currently obtained current state encoding into the policy network to generate a probability distribution of adjustment actions, and optimize it in combination with the policy optimization criterion to obtain an optimized correction strategy.

7. The method according to claim 6, characterized in that, Obtain the output capsule of the capsule network through the product of the routing weight and the output of the initial capsule layer after passing through a compression function. Predicting the machining quality score corresponding to the initial adjustment plan based on the output capsule includes: Construct the routing weight in the capsule network and the output of the initial capsule layer as nodes in a Markov random field structure, calculate the edge connection probability based on the spatial position relationship and feature similarity between the nodes, construct the edge connection probability as a potential function, construct a single-node potential function based on the distribution parameters for the nodes, and construct the joint probability distribution of the node states through the product of the single-node potential function and the potential function; Based on the joint probability distribution, initialize the state value of the nodes according to Gibbs sampling, calculate the conditional probability of the nodes based on the single-node potential function and the potential function, iteratively update the node states according to the conditional probability to obtain a sampling sequence, monitor the autocorrelation of the sampling sequence, and dynamically adjust the sampling step size and the number of iterations until the preset convergence condition is met; The optimized routing weight is obtained by calculating the edge probability distribution of the sampling sequence, the product result is obtained by multiplying the optimized routing weight by the output of the initial capsule layer, and the output capsule of the capsule network is obtained by inputting the product result into a compression function. The inner product of the output capsule and a preset parameter vector is calculated to obtain an initial score. Based on the initial score, a normalized score is obtained through a non-linear activation function and used as the machining quality score corresponding to the initial adjustment scheme.

8. An intelligent machine tool tool wear state recognition and machining parameter adjustment system for implementing the method described in any one of the foregoing claims 1-7, characterized in that, It includes: A first unit for collecting the working condition data during the machining process of an intelligent machine tool, performing sliding window segmentation on the segmented working condition data, and performing denoising and normalization processing on the working condition data to obtain standard working condition data. A second unit for extracting long-term and short-term time series feature sequences from the standard working condition data through a bidirectional gated recurrent unit, inputting the long-term and short-term time series feature sequences into a Hilbert-Huang transform decomposition to obtain intrinsic mode components, combining four-layer spatial pyramid pooling to extract multi-scale frequency domain feature sequences, calculating the mutual information entropy between the time series feature sequences and the frequency domain feature sequences, determining the fusion weight coefficient according to the mutual information entropy value, and performing weighting to obtain a tool wear state feature vector. A third unit for inputting the tool wear state feature vector into a pre-trained extreme learning machine to calculate the current tool wear amount value and the remaining service life value. A fourth unit for adding the current tool wear amount value and the remaining service life value to a parallel agent structure, determining an initial adjustment scheme by combining the observed shared information collected by each agent, predicting the machining quality score corresponding to the initial adjustment scheme through a capsule network, and feeding the machining quality score back to the parallel agent structure as a reward signal to obtain an optimized correction strategy. A fifth unit for adjusting the working parameters of the intelligent machine tool based on the optimized correction strategy and adding them to the machine tool process parameter library.

9. An electronic device, characterized in that, It includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, The computer program instructions, when executed by the processor, implement the method according to any one of claims 1 to 7.

Images