Intelligent machine tool tool wear state identification and processing parameter adjustment method and system

Abstract

The present invention provides a method and system for identifying tool wear status and adjusting machining parameters of intelligent machine tools, which relate to the field of intelligent manufacturing technology. The method comprises the following steps: collecting working condition data of a machine tool machining process and performing sliding window segmentation, denoising and normalization processing; extracting a time series feature sequence through a bidirectional gated cyclic unit; obtaining intrinsic modal components by combining Hilbert-Huang transform decomposition; extracting a frequency domain feature sequence by using four-layer spatial pyramid pooling; determining fusion weights based on mutual information entropy to obtain a tool wear state feature vector; inputting the feature vector into a pre-trained extreme learning machine to predict wear amount and remaining life; determining an initial adjustment plan by combining observation sharing information of a parallel intelligent agent structure; predicting a machining quality score through a capsule network and feeding it back as a reward signal; and obtaining an optimized correction strategy for adjusting machine tool operating parameters.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a method and system for identifying tool wear status and adjusting machining parameters of intelligent machine tools. 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 intelligent machine tools, the tool wear status directly affects the machining accuracy and product quality. Tool wear is an inevitable phenomenon in the machining process. As the tool is used for a longer time, the tool geometry and cutting performance will gradually deteriorate. Accurately identifying the tool wear status and adjusting the machining parameters in a timely manner are of great significance to ensuring machining quality and improving production efficiency.

[0003] Tool wear status monitoring mainly uses two methods: direct measurement and indirect measurement. Direct measurement directly measures the changes in the tool's geometric parameters through optical or tactile means, while indirect measurement infers the tool wear status by collecting cutting force, vibration, acoustic emission and other signals during the machining process. With the development of sensing technology and artificial intelligence technology, tool wear status identification methods based on working condition data analysis have gradually become a research hotspot.

[0004] Existing technologies still cannot fully capture the long-term and short-term dependencies in the tool wear process, resulting in low tool wear state identification accuracy and a lack of in-depth understanding of the relationship between tool wear state and processing parameters. Parameter adjustment often relies on empirical formulas or simple feedback adjustment mechanisms, and is unable to perform intelligent and personalized parameter optimization based on the tool's specific wear state and remaining service life, nor is it possible to use a single decision-making mechanism to adjust processing parameters. There is a lack of a multi-agent collaborative decision-making framework, and it is impossible to comprehensively consider multiple factors such as tool wear, processing quality, and production efficiency. This makes it difficult to achieve optimal control in a dynamically changing processing environment. There is also a lack of an evaluation and optimization mechanism for decision-making results.

[0005] Therefore, a solution is urgently needed 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 tool wear status and adjusting machining parameters of intelligent machine tools, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for identifying tool wear status and adjusting machining parameters of an intelligent machine tool, comprising:

[0008] Collect the working condition data of the intelligent machine tool during machining and perform sliding window segmentation. Then, perform denoising and normalization on the segmented working condition data to obtain standard working condition data.

[0009] A bidirectional gated recurrent unit is used to extract long- and short-term time series feature sequences from standard working condition data. The long- and short-term time series feature sequences are input into the Hilbert-Huang transform decomposition to obtain intrinsic modal components. A multi-scale frequency domain feature sequence is extracted by combining the four-layer spatial pyramid pooling. The mutual information entropy of the time series feature sequence and the frequency domain feature sequence is calculated. The fusion weight coefficient is determined according to the mutual information entropy value and weighted to obtain the tool wear state feature vector.

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

[0011] The current wear value and remaining service life value of the tool 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 the capsule network. The processing quality score is fed back to the parallel agent structure as a reward signal to obtain the optimized correction strategy.

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

[0013] In an optional embodiment,

[0014] The working condition data of the intelligent machine tool during machining is collected and segmented by sliding windows. The segmented working condition data is denoised and normalized to obtain standard working condition data including:

[0015] Collecting working condition data during the intelligent machine tool processing, the working condition data including spindle current signal, spindle speed signal, feed speed signal and cutting force signal;

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

[0017] For the segmented operating condition data sequence, the signal is decomposed and reconstructed at multiple scales using the wavelet threshold method to remove noise. A Butterworth low-pass filter is used to filter out high-frequency interference components that are 50% higher than the sampling frequency, and the filtered signal is normalized to the maximum and minimum values ​​to obtain standard operating condition data.

[0018] In an optional embodiment,

[0019] 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 decomposition to obtain the intrinsic modal components. The multi-scale frequency domain feature sequences are extracted by combining the four-layer spatial pyramid pooling, including:

[0020] The standard working condition data is input into the bidirectional gated recurrent unit network, and the forward hidden layer state and the backward hidden layer state are established in the bidirectional gated recurrent unit network. The forward hidden layer state and the backward hidden layer state are updated through the reset gate and the update gate respectively, and the updated forward hidden layer state and the backward hidden layer state are spliced ​​to generate a long-term and short-term time series feature sequence;

[0021] The long-term and short-term time series feature sequences are subjected to empirical mode decomposition to obtain intrinsic mode function components, the intrinsic mode function components are subjected to Hilbert transform to obtain Hilbert transform components, an analytical signal is constructed based on the intrinsic mode function components and the Hilbert transform components, and the instantaneous amplitude and instantaneous phase are extracted from the analytical signal to form the intrinsic mode components;

[0022] In a four-layer spatial pyramid pooling network, pooling scales of one by one, two by two, four by four, and eight by eight are set respectively, the intrinsic modal component is input into the four-layer spatial pyramid pooling network, a pooling output feature map is generated, and a maximum pooling operation is performed to obtain four layers of pooled features, and the four layers of pooled features are spliced ​​to form a multi-scale frequency domain feature sequence.

[0023] In an optional embodiment,

[0024] Calculating the mutual information entropy of the time series feature sequence and the frequency domain feature sequence, determining the fusion weight coefficient according to the mutual information entropy value, and weighting to obtain the tool wear state feature vector includes:

[0025] The mutual information entropy between the two feature sequences is calculated by multiplying the logarithms of the components of the time series feature sequence and the frequency domain feature sequence and summing them up.

[0026] The difference between the eigenvalue and the corresponding eigenmean in the time series feature sequence and the frequency domain feature sequence is divided by the bandwidth parameter to obtain the standardized eigenvalue. The standardized eigenvalue is substituted into the Gaussian kernel function. The result of the Gaussian kernel function is divided by the product of the number of samples and the bandwidth parameter to obtain the probability distribution value.

[0027] The mutual information entropy value is negative and multiplied by the adjustment parameter and the exponential operation is performed to obtain the weight coefficient numerator. The weight coefficient numerator is divided by the sum of the weight coefficient numerator and its complement to obtain the fusion weight coefficient of the time series feature sequence. The fusion weight coefficient of the time series feature sequence is subtracted from 1 to obtain the fusion weight coefficient of the frequency domain feature sequence.

[0028] Multiply the time series feature sequence with the fusion weight coefficient of the time series feature sequence to obtain the weighted time series feature, multiply the frequency domain feature sequence with 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;

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

[0030] In an optional embodiment,

[0031] The tool wear state feature vector is input into a pre-trained extreme learning machine to calculate the current wear value and the remaining service life value of the tool, including:

[0032] Collecting tool wear state sample data corresponding to a tool wear state feature vector, dividing the tool wear state sample data into a training data set and a test data set, training an extreme learning machine based on the training data set, wherein 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 a Sigmoid activation function, randomly generates initial values ​​of an input weight matrix and a bias vector, multiplies the tool wear state feature vector in the training data set by the input weight matrix and adds the result to the bias vector, obtains a hidden layer output through the Sigmoid activation function, calculates an output weight based on the hidden layer output, and obtains a pre-trained extreme learning machine;

[0033] The tool wear state feature vector is input into the pre-trained extreme learning machine, and the tool wear state feature vector is multiplied by the input weight matrix and added to the bias vector, and then the result is activated by the Sigmoid function to obtain the hidden layer output. The current wear value and the remaining service life value of the tool are calculated based on the hidden layer output.

[0034] In an optional embodiment,

[0035] The current wear value and the remaining service life value of the tool 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 by the capsule network. The processing quality score is fed back to the parallel agent structure as a reward signal. The optimization correction strategy includes:

[0036] Adding the current wear value and the remaining service life value of the tool to a parallel agent structure, wherein the parallel agent structure includes multiple parallel agents, each of which collects local observation shared information, and obtains global observation shared information by exchanging the local observation shared information through a shared network;

[0037] Calculating the attention weight coefficients between the parallel agents based on the global observation sharing information, wherein the attention weight coefficients are obtained by performing an inner product operation on the query vector corresponding to the parallel agent and the key vector, performing weighted accumulation on the attention weight coefficients and the corresponding value vectors to obtain a fused feature representation, and determining an initial adjustment scheme based on the fused feature representation and the observation sharing information collected by each of the parallel agents;

[0038] Inputting the initial adjustment plan into a capsule network, transforming the initial adjustment plan using a transformation matrix of the capsule network to obtain an initial capsule layer output, calculating a routing weight based on the initial capsule layer output, obtaining an output capsule of the capsule network based on a compression function obtained by multiplying the routing weight by the initial capsule layer output, and predicting a processing quality score corresponding to the initial adjustment plan based on the output capsule;

[0039] The processing quality score is fed back to the parallel intelligent agent structure as a reward signal. The parallel intelligent agent structure encodes the processing quality score to obtain the historical information code. The historical information code and the current state code obtained in real time are input into the strategy network to generate the probability distribution of the adjustment action. The optimization is carried out in combination with the strategy optimization criterion to obtain the optimized correction strategy.

[0040] In an optional embodiment,

[0041] Obtaining an output capsule of the capsule network according to the product of the routing weight and the output of the initial capsule layer through a compression function, and predicting a processing quality score corresponding to the initial adjustment solution based on the output capsule includes:

[0042] The routing weights in the capsule network and the output of the initial capsule layer are constructed as nodes in a Markov random field structure, the edge connection probability is calculated based on the spatial position relationship and feature similarity between the nodes, the edge connection probability is constructed as a potential function, a single-node potential energy function based on a distribution parameter is constructed for the node, and a joint probability distribution of the node state is constructed by multiplying the single-node potential energy function and the potential function;

[0043] Based on the joint probability distribution, the state value of the node is initialized according to Gibbs sampling, the conditional probability of the node is calculated based on the single-node potential energy function and the potential function, the node state is iteratively updated according to the conditional probability to obtain a sampling sequence, the autocorrelation of the sampling sequence is monitored, and the sampling step size and the number of iterations are dynamically adjusted until a preset convergence condition is met;

[0044] Calculating the sampling sequence through marginal probability distribution to obtain an optimized routing weight, multiplying the optimized routing weight by the output of the initial capsule layer to obtain a product result, and inputting the product result into a compression function to obtain an output capsule of the capsule network;

[0045] An initial score is obtained by calculating the inner product of the output capsule and a preset parameter vector. Based on the initial score, a normalized score is obtained through a nonlinear activation function and used as the processing quality score corresponding to the initial adjustment scheme.

[0046] A second aspect of an embodiment of the present invention provides a system for identifying tool wear status and adjusting machining parameters of an intelligent machine tool, comprising:

[0047] The first unit is used to collect the working condition data of the intelligent machine tool during the processing and perform sliding window segmentation, and perform denoising and normalization processing on the segmented working condition data to obtain standard working condition data;

[0048] The second unit is used to extract long-term and short-term time series feature sequences from the standard working condition data through a bidirectional gated cyclic unit, input the long-term and short-term time series feature sequences into the Hilbert-Huang transform decomposition to obtain intrinsic modal components, combine the four-layer spatial pyramid pooling to extract the multi-scale frequency domain feature sequences, calculate the mutual information entropy of 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 weighted analysis to obtain the tool wear state feature vector;

[0049] The third unit is used to input the tool wear state feature vector into a pre-trained extreme learning machine to calculate the current wear value and the remaining service life value of the tool;

[0050] The fourth unit is used to add the current wear value and the remaining service life value of the tool to the parallel agent structure, determine an initial adjustment plan based on the observation shared information collected by each agent, predict the processing quality score corresponding to the initial adjustment plan through the capsule network, and feed the processing quality score back to the parallel agent structure as a reward signal to obtain an optimized correction strategy;

[0051] The fifth unit is used to adjust the working parameters of the intelligent machine tool based on the optimization correction strategy and add them to the machine tool process parameter library.

[0052] According to a third aspect of an embodiment of the present invention, an electronic device is provided, 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 aforementioned method.

[0054] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0055] In the present invention, by performing sliding window segmentation, denoising and normalization processing on the working condition data, and combining the bidirectional gated cyclic unit and Hilbert-Huang transform to perform time-frequency feature extraction, it is possible to comprehensively capture various characteristic information in the tool wear process, significantly improve the accuracy and stability of tool wear state identification, and reduce 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 body structure. At the same time, a capsule network is used to evaluate the plan and predict the quality, forming a closed-loop feedback optimization mechanism, realizing intelligent dynamic adjustment of processing parameters, effectively extending the tool life, adding the optimized working parameters to the machine tool process parameter library, and establishing a knowledge accumulation mechanism of self-learning and continuous optimization, which not only improves the current processing quality and efficiency, but also provides a reliable reference for parameter setting under similar working conditions, reduces material waste and energy consumption, and enhances the adaptive processing capability of intelligent machine tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of the process of the method for identifying tool wear status and adjusting machining parameters of an intelligent machine tool according to an embodiment of the present invention;

[0057] Figure 2 A visualization diagram of the capsule network hierarchy and routing mechanism of the method for identifying tool wear status and adjusting machining parameters of intelligent machine tools according to an embodiment of the present invention;

[0058] Figure 3 This is a simulation effect diagram of capsule network routing optimization of the intelligent machine tool tool wear status identification and processing parameter adjustment method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0060] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG. 1 is a flow chart of a method for identifying tool wear status and adjusting machining parameters of an intelligent machine tool according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] Collect the working condition data of the intelligent machine tool during machining and perform sliding window segmentation. Then, perform denoising and normalization on the segmented working condition data to obtain standard working condition data.

[0063] A bidirectional gated recurrent unit is used to extract long- and short-term time series feature sequences from standard working condition data. The long- and short-term time series feature sequences are input into the Hilbert-Huang transform decomposition to obtain intrinsic modal components. A multi-scale frequency domain feature sequence is extracted by combining the four-layer spatial pyramid pooling. The mutual information entropy of the time series feature sequence and the frequency domain feature sequence is calculated. The fusion weight coefficient is determined according to the mutual information entropy value and weighted to obtain the tool wear state feature vector.

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

[0065] The current wear value and remaining service life value of the tool 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 the capsule network. The processing quality score is fed back to the parallel agent structure as a reward signal to obtain the optimized correction strategy.

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

[0067] In an optional embodiment,

[0068] The working condition data of the intelligent machine tool during machining is collected and segmented by sliding windows. The segmented working condition data is denoised and normalized to obtain standard working condition data including:

[0069] Collecting working condition data during the intelligent machine tool processing, the working condition data including spindle current signal, spindle speed signal, feed speed signal and cutting force signal;

[0070] Perform sliding window segmentation on the operating condition data, with an overlap rate of 50% between adjacent windows, to obtain a segmented operating condition data sequence;

[0071] For the segmented operating condition data sequence, the signal is decomposed and reconstructed at multiple scales using the wavelet threshold method to remove noise. A Butterworth low-pass filter is used to filter out high-frequency interference components that are 50% higher than the sampling frequency, and the filtered signal is normalized to the maximum and minimum values ​​to obtain standard operating condition data.

[0072] Collect working condition data during the intelligent machine tool processing process. The sensor network collects the spindle current signal in real time. During the collection process, the instantaneous changes in the motor input current are monitored and converted into a standard current signal. The spindle speed signal is collected and the actual spindle rotation speed is obtained through the encoder. The feed rate signal is collected to record the relative movement speed of the tool with respect to the workpiece. The cutting force signal is collected and a force sensor is placed in the contact area between the tool and the workpiece to obtain the dynamic cutting force during the cutting process.

[0073] Sliding window segmentation is performed on the collected operating condition data. The window length and sliding step size are determined. The window length is set based on the signal characteristics and sampling frequency, and the sliding step size is half the window length. The overlap between adjacent windows is guaranteed to be 50%. The window is sequentially slid across the entire data sequence, extracting a data segment with each slide. All data segments constitute the segmented operating condition data sequence. The overlapping design avoids information fragmentation caused by window segmentation and enhances data continuity.

[0074] Wavelet denoising is performed on the segmented operating condition data sequence. Orthogonal wavelet basis functions are selected to decompose the signal into wavelet coefficients at multiple scales, including low-frequency approximate coefficients and high-frequency detail coefficients. A threshold function is determined based on the signal energy distribution characteristics, and threshold processing is performed on the detail coefficients at different scales to retain the valid signal components and suppress the noise components. A soft thresholding method is used to shrink the coefficients exceeding the threshold. The processed wavelet coefficients are then used to reconstruct the denoised signal.

[0075] A Butterworth low-pass filter was designed for filtering. The filter order was selected based on the signal's spectral characteristics, and the passband and stopband were determined by setting the cutoff frequency to half the sampling frequency. The filter exhibited excellent amplitude-frequency characteristics, with gentle attenuation in the passband and rapid attenuation in the stopband, effectively filtering out high-frequency interference. The signal after wavelet denoising was filtered to further improve signal quality.

[0076] Perform maximum and minimum normalization on the filtered signals. Calculate the maximum and minimum values ​​for each signal sequence and linearly map the raw data to a range between zero and one. Normalization eliminates dimensional differences and numerical ranges between different signal types, ensuring that all signals have the same numerical distribution characteristics, resulting in the standard operating condition data.

[0077] For example, during the machining process of a certain CNC machine tool, the sampling frequency is set to 1000Hz, and the working condition data is continuously collected for 5 seconds. The spindle current signal amplitude range is 0-30A, the spindle speed signal range is 0-3000rpm, the feed speed signal range is 0-500mm / min, and the cutting force signal range is 0-2000N. A 2048-point sliding window is used to segment the data, and adjacent windows overlap by 1024 points, resulting in a total of 4 data segments. The db4 wavelet is used to perform a 4-layer decomposition of the signal, and the soft threshold method is used to remove noise and reconstruct the signal. A 4th-order Butterworth low-pass filter with a cutoff frequency of 500Hz is designed to filter the reconstructed signal. Finally, the filtered signal is normalized to obtain a standard working condition data sequence with amplitudes in the range of [0, 1].

[0078] In this embodiment, by collecting four types of signals from the working condition data, the multi-dimensional characteristics of the machine tool processing status are fully reflected. Compared with single signal collection, it can more comprehensively characterize the dynamic changes of the processing process. The sliding window segmentation method is adopted. By setting appropriate window length and overlap rate, the real-time signal analysis is guaranteed and the information fragmentation caused by signal segmentation is avoided. The combination of wavelet threshold method and Butterworth filter is adopted to achieve multi-scale and multi-level noise suppression, which significantly improves the signal-to-noise ratio of the signal.

[0079] In an optional 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 decomposition to obtain the intrinsic modal components. The multi-scale frequency domain feature sequences are extracted by combining the four-layer spatial pyramid pooling, including:

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

[0082] The long-term and short-term time series feature sequences are subjected to empirical mode decomposition to obtain intrinsic mode function components, the intrinsic mode function components are subjected to Hilbert transform to obtain Hilbert transform components, an analytical signal is constructed based on the intrinsic mode function components and the Hilbert transform components, and the instantaneous amplitude and instantaneous phase are extracted from the analytical signal to form the intrinsic mode components;

[0083] In a four-layer spatial pyramid pooling network, pooling scales of one by one, two by two, four by four, and eight by eight are set respectively, the intrinsic modal component is input into the four-layer spatial pyramid pooling network, a pooling output feature map is generated, and a maximum pooling operation is performed to obtain four layers of pooled features, and the four layers of pooled features are spliced ​​to form a multi-scale frequency domain feature sequence.

[0084] Standard operating condition data is input into a bidirectional gated recurrent unit network. Two processing channels, forward and backward, are constructed simultaneously. In the forward channel, data is input sequentially in chronological order. The reset gate dynamically adjusts the degree of forgetting of historical information, removing historical information that has little contribution to the current state prediction. The update gate adaptively adjusts the retention ratio of historical states based on the importance of the current input, forming a forward hidden layer state sequence. In the backward channel, data is input sequentially in reverse chronological order and processed using the same gating mechanism to generate a backward hidden layer state sequence. The state sequences of the two channels capture the forward and backward temporal dependencies of the data, respectively. The forward and backward hidden layer states are concatenated in the feature dimension to obtain a long-term and short-term feature sequence that integrates bidirectional temporal information.

[0085] Perform empirical mode decomposition on the long- and short-term characteristic sequences. Identify all local extreme points in the sequence and construct the upper and lower envelopes using 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 this process until the new sequence meets the definition of an eigenmode function (IMF), obtaining the first IMF component. Subtract this component from the original sequence and repeat the decomposition process for the remaining sequences to obtain all IMF components.

[0086] A Hilbert transform is performed on each intrinsic mode function component. This transform converts the real signal into the imaginary part of the analytical signal, which is then combined with the original real signal to construct the analytical signal. The instantaneous amplitude is extracted from the analytical signal to characterize the amplitude characteristics of the signal at each moment, and the instantaneous phase is extracted to characterize the phase variation characteristics of the signal. The extracted instantaneous amplitude and phase information are combined to form the intrinsic mode component that characterizes the time-frequency characteristics of the signal.

[0087] A multi-layer spatial pyramid pooling network is constructed for feature extraction. The first layer sets a minimum pooling window to preserve local signal details. The second layer increases the pooling window size to extract statistical features of local regions. The third layer further expands the pooling range to capture medium-scale feature patterns. The fourth layer uses a maximum pooling window to extract global-scale feature expressions. For each layer, the intrinsic modal components are input into the current pooling layer. A sliding window is used to scan the feature map. Maximum pooling is performed within each window region to extract the most significant feature response in that region.

[0088] Each pooling layer independently processes feature maps of varying resolutions. Smaller-scale pooling layers retain more local details, while larger-scale pooling layers extract more abstract global features. The feature maps output by the four pooling layers undergo dimensionality reduction to generate pooled feature vectors of varying scales. Finally, these feature vectors are concatenated sequentially along the feature dimension to form a feature sequence containing frequency-domain information at multiple scales, enabling the representation of signal features at different time and frequency scales.

[0089] For example, during the machining process of a certain CNC machine tool, the standard working condition data collected contains four sensor signals, and each signal sequence is one thousand data points long. Through bidirectional gated recurrent unit network processing, the number of hidden layer units is set to 128, and a 256-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 analytical signal construction, the instantaneous amplitude and phase information of each component are obtained to form the intrinsic mode component. The intrinsic mode component is input into a four-layer spatial pyramid pooling network. After pooling operations at different scales, four sets of features are obtained, with feature dimensions of 64, 16, 4, and 1, respectively. These features are spliced ​​to obtain an 85-dimensional multi-scale frequency domain feature sequence.

[0090] In this embodiment, the timing features are extracted through a bidirectional gated recurrent unit network, which realizes the capture of the forward and backward dual dependencies of the working condition data, significantly enhances the network's learning ability of long-term dependencies, and improves the accuracy and robustness of timing feature extraction. The empirical mode decomposition method is used to adaptively decompose the long-term and short-term feature sequences without the need to preset the decomposition basis function. It can perform modal analysis according to the characteristics of the signal itself, and realizes the effective decomposition of complex non-stationary signals. The analytical signal is constructed in combination with the Hilbert transform, and the instantaneous amplitude and instantaneous phase information are extracted from it to form the inherent modal component, which can comprehensively characterize the instantaneous characteristics and frequency change law of the signal, and improve the integrity of the feature expression.

[0091] In an optional embodiment,

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

[0093] The mutual information entropy between the two feature sequences is calculated by multiplying the logarithms of the components of the time series feature sequence and the frequency domain feature sequence and summing them up.

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

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

[0096] Multiply the time series feature sequence with the fusion weight coefficient of the time series feature sequence to obtain the weighted time series feature, multiply the frequency domain feature sequence with 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] The eigenvector corresponding to the maximum eigenvalue in the fusion feature sequence is extracted as the main feature component to construct the tool wear state feature vector.

[0098] Mutual information entropy is calculated for the time series and frequency domain feature sequences. The two feature sequences are mapped one-to-one according to their dimensions, and each pair of feature components is extracted and processed. The logarithms of each pair of feature components are taken and multiplied together to obtain product terms. The product terms of all corresponding components are calculated sequentially and summed to obtain the mutual information entropy value, which represents the statistical correlation between the two feature sequences. The mutual information entropy value reflects the degree of information sharing between the two feature sequences; a higher entropy value indicates a stronger interdependence between the feature sequences.

[0099] Perform feature normalization and probability distribution estimation. Calculate the feature mean for each dimension of the time series and frequency domain feature sequences. Subtract the corresponding mean from the feature value in each dimension to obtain the feature deviation. Select an appropriate bandwidth parameter and normalize the feature deviation by dividing it by the bandwidth parameter to obtain the standardized feature value. Substitute the standardized feature value into the Gaussian kernel function and calculate the kernel function response. Divide the response by the product of the total number of samples and the bandwidth parameter to obtain the probability distribution estimate of the feature value. This kernel function-based probability density estimation method effectively captures the nonlinear 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 the pre-set adjustment parameter, and perform an exponential operation on the product to obtain the numerator of the weight coefficient. Calculate the sum of this numerator and its complement as the denominator, and divide the numerator by the denominator to obtain the fusion weight coefficient for 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 for the frequency domain feature sequence. This mutual information entropy-based weight distribution method can adaptively adjust the fusion ratio based on the correlation between feature sequences.

[0101] Perform a weighted feature fusion operation. Multiply each feature component in the time series feature sequence by the time series feature weight coefficient to obtain a 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 a weighted frequency domain feature sequence. Add the two weighted feature sequences in the corresponding dimensions to generate a fused feature sequence. This weighted fusion strategy preserves the important information components in each feature sequence while suppressing the influence of redundant information and noise.

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

[0103] For example, 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. The bandwidth parameter is 0.1 for feature normalization, and the feature probability distribution is estimated by 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 calculated to be 0.4. The two feature sequences are weightedly fused to obtain a 300-dimensional fused feature sequence. The eigenvector corresponding to the maximum eigenvalue is extracted as the 60-dimensional tool wear state feature vector.

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

[0105] Existing methods for extracting tool wear state features typically use single time series features or frequency domain features for analysis, which makes it difficult to fully reflect the dynamic changes in tool wear. Furthermore, simple feature splicing or fixed weight fusion cannot be used to adaptively adjust features based on their importance, resulting in insufficient accuracy and robustness in 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 feature correlation measure, enhances the rationality of feature fusion, adopts an adaptive weight allocation mechanism, improves the flexibility of feature fusion, realizes feature dimensionality reduction optimization through main feature extraction, improves the efficiency of feature expression, and provides reliable feature support for the accurate identification and prediction of tool wear status.

[0107] In an optional embodiment,

[0108] The tool wear state feature vector is input into a pre-trained extreme learning machine to calculate the current wear value and the remaining service life value of the tool, including:

[0109] Collecting tool wear state sample data corresponding to a tool wear state feature vector, dividing the tool wear state sample data into a training data set and a test data set, training an extreme learning machine based on the training data set, wherein 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 a Sigmoid activation function, randomly generates initial values ​​of an input weight matrix and a bias vector, multiplies the tool wear state feature vector in the training data set by the input weight matrix and adds the result to the bias vector, obtains a hidden layer output through the Sigmoid activation function, calculates an output weight based on the hidden layer output, and obtains a pre-trained extreme learning machine;

[0110] The tool wear state feature vector is input into the pre-trained extreme learning machine, and the tool wear state feature vector is multiplied by the input weight matrix and added to the bias vector, and then the result is activated by the Sigmoid function to obtain the hidden layer output. The current wear value and the remaining service life value of the tool are calculated based on the hidden layer output.

[0111] Multi-sensor technology collects sample data on tool wear, corresponding to the tool wear state feature vectors during machining. The acquisition process covers the entire tool lifecycle, from initial state to fully worn state, ensuring the integrity and representativeness of the sample data. The collected raw data is preprocessed to eliminate outliers and noise, improving data quality. The processed sample dataset is randomly divided into training and test datasets according to a preset ratio to ensure consistent sample distribution across the two datasets.

[0112] Construct the topology of the extreme learning machine network. The number of input layer nodes matches the dimension of the tool wear state feature vector to ensure the network fully receives all feature information. An appropriate number of neuron nodes are set in the hidden layer. These neurons transform and map the feature space using nonlinear activation functions. A continuously differentiable saturated activation function is selected. Saturated activation functions exhibit excellent nonlinear characteristics and numerical stability, effectively extracting complex relationships between features. Two nodes are set in the output layer, one for predicting the current tool wear and the other for predicting the remaining tool life.

[0113] Initialize the network parameters. A random number generator is used to generate random numbers from a specific distribution to construct a weight matrix from the input layer to the hidden layer. Each element of the weight matrix represents the importance of the corresponding input feature. A bias vector is generated for the hidden layer nodes to adjust the activation threshold of each hidden layer node. This random initialization strategy ensures that the network obtains a good initial state in parameter space, avoiding falling into a symmetric solution.

[0114] Execute the model training process. The feature vectors from the training dataset are sequentially input into the network. Matrix multiplication is performed on the feature vectors and the input weight matrix to achieve a linear transformation of the features. The transformed result is added to the bias vector to obtain the input value for each hidden layer node. Each input value is nonlinearly mapped through an activation function to generate the actual output value of the hidden layer node. The hidden layer output values ​​of all training samples are collected to construct the hidden layer output matrix.

[0115] The output weight matrix is ​​calculated using the least squares method based on the hidden layer output matrix and the target values ​​in the training dataset. This direct solution avoids the iterative optimization process required by traditional neural networks and significantly improves training efficiency. The calculated output weight matrix contains the optimal mapping from hidden layer features to output targets. This completes the pre-training process of the extreme learning machine, and the network obtains all the parameters required to predict tool wear.

[0116] Online prediction is performed using the trained model. The real-time tool wear state feature vector is input into the network and transformed using the trained input weight matrix. The transformed features are added to the bias vector to form the input of the hidden layer nodes. Each hidden layer node generates an output value through an activation function, forming a hidden layer feature representation of the current state. The hidden layer features are multiplied by the output weight matrix to obtain the final prediction result. The first output node provides the current tool wear value, and the second output node provides the estimated remaining service life value.

[0117] For example, for a CNC milling process, one thousand sets of tool wear status sample data were collected. The sample data was divided into training and test sets in a ratio of seven to three. An extreme learning machine network was constructed with an input layer dimension of sixty and a hidden layer node count of one hundred and twenty. Initial values ​​for the input weight matrix and bias vector were generated using a random method. The training data was input into the network, and the hidden layer output was obtained through matrix operations and activation functions. The output weight matrix was calculated to complete model training. A prediction was performed on the newly collected wear status feature vector, outputting a current tool wear of 0.3 mm and an estimated remaining service life of 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 is established, ensuring that the evolution law of the tool wear state can be fully learned during the model training process. At the same time, the prediction performance of the model is verified by the test data set, ensuring the reliability of the prediction results. The extreme learning machine network structure used is reasonably designed. The input layer completely receives the tool wear state feature information, the hidden layer realizes deep mapping and extraction of features through nonlinear activation functions, and the output layer directly gives the predicted values ​​of wear amount and remaining life, effectively capturing the complex relationship between features and wear state, and determining the initial values ​​of the input weight matrix and bias vector by random generation, providing a good learning starting point for the network.

[0119] In an optional embodiment,

[0120] The current wear value and the remaining service life value of the tool 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 by the capsule network. The processing quality score is fed back to the parallel agent structure as a reward signal. The optimization correction strategy includes:

[0121] Adding the current wear value and the remaining service life value of the tool to a parallel agent structure, wherein the parallel agent structure includes multiple parallel agents, each of which collects local observation shared information, and obtains global observation shared information by exchanging the local observation shared information through a shared network;

[0122] Calculating the attention weight coefficients between the parallel agents based on the global observation sharing information, wherein the attention weight coefficients are obtained by performing an inner product operation on the query vector corresponding to the parallel agent and the key vector, performing weighted accumulation on the attention weight coefficients and the corresponding value vectors to obtain a fused feature representation, and determining an initial adjustment scheme based on the fused feature representation and the observation sharing information collected by each of the parallel agents;

[0123] Inputting the initial adjustment plan into a capsule network, transforming the initial adjustment plan using a transformation matrix of the capsule network to obtain an initial capsule layer output, calculating a routing weight based on the initial capsule layer output, obtaining an output capsule of the capsule network based on a compression function obtained by multiplying the routing weight by the initial capsule layer output, and predicting a processing quality score corresponding to the initial adjustment plan based on the output capsule;

[0124] The processing quality score is fed back to the parallel intelligent agent structure as a reward signal. The parallel intelligent agent structure encodes the processing quality score to obtain the historical information code. The historical information code and the current state code obtained in real time are input into the strategy network to generate the probability distribution of the adjustment action. The optimization is carried out in combination with the strategy optimization criterion to obtain the optimized correction strategy.

[0125] A collaborative optimization structure consisting of multiple parallel agents is constructed, and the predicted values ​​for tool wear and remaining tool life are added to the state space of each agent. Each parallel agent is equipped with multiple sensors to continuously collect local observation information in its area, including multi-dimensional machining data such as cutting forces, vibration, temperature, and acoustic emissions. Each agent establishes an information exchange channel through a shared communication network, encoding and transmitting the collected local observation information, achieving a two-way flow of information. Each agent receives observation information sent by other agents and integrates this information with its own collected data to construct an observation information matrix reflecting the global machining state.

[0126] Based on global observation information, an attention calculation mechanism is constructed between agents. Feature extraction and mapping are performed on the state information of each agent to generate a query vector that represents the characteristics of the agent. At the same time, the state information of other agents is mapped to corresponding key vectors. By performing an inner product operation on the query vector and the key vector, an attention score that represents the degree of association between the agents is calculated. The attention score is normalized to obtain an attention weight coefficient that reflects the importance of each agent. The attention weight coefficient is weighted and summed with the value vector corresponding to the agent to obtain a fused feature representation. Based on the fused feature representation and combined with the real-time observation information collected by each agent, a comprehensive assessment of the current machining state is made, and an initial adjustment plan for parameters such as feed rate and spindle speed is formulated.

[0127] The initial adjustment plan is fed into a specially designed capsule network for deep optimization. The capsule network first transforms and reconstructs the adjustment plan's features using a transformation matrix, extracting key attribute information from the plan and generating feature vectors for the initial capsule layer. A dynamic routing algorithm is then used to calculate the correlation strength between different feature capsules, generating routing weights that represent the importance of these features. The routing weights are then weighted and combined with the feature vectors of the initial capsule layer, normalized using a nonlinear compression function, to generate output capsules containing the optimization recommendations. The information in the output capsules is then analyzed to predict the processing quality score after the initial adjustment plan is implemented.

[0128] The predicted processing quality score is fed back to the parallel agent structure as an evaluation signal. After receiving the quality score, the parallel agent structure performs temporal encoding of historical adjustment plans and their corresponding effects, building a long-term memory that incorporates historical decision-making experience. Simultaneously, the current processing state is encoded in real time to generate a short-term memory that reflects the immediate situation. The encoded information from both long-term and short-term memories is fed into the policy network. Through mapping and transformation within a multi-layer neural network, a probability distribution for each possible adjustment action is generated.

[0129] The probability distribution is optimized using a policy gradient method. The expected reward under the current policy is calculated and compared with the baseline value to obtain the advantage function. Based on the gradient information of the advantage function, the network parameters are updated according to the policy optimization criteria, and the action probability distribution is adjusted. Through multiple rounds of iterative optimization, the decision-making ability of the policy network is continuously improved, resulting in an experienced and highly reliable optimized and corrected policy.

[0130] For example, in a CNC machining workshop, five parallel agents were deployed to monitor tool processing status. When tool wear reached 0.3 mm and the estimated remaining service life was six hours, each agent collected and shared local processing data. An attention mechanism was used to calculate interaction weights between agents, generate fusion features, and formulate an initial adjustment plan for feed rate and spindle speed. This adjustment plan was input into a capsule network for optimization, resulting in a predicted processing quality score of 95. Reinforcement learning was then performed based on this score, updating the policy network parameters and generating an optimized processing parameter correction strategy, enabling intelligent control of the machining process.

[0131] In this embodiment, a distributed network of intelligent agents can collect processing status information from multiple dimensions, avoiding the information loss and misjudgment problems that may arise from single-point monitoring. By calculating the attention weights between intelligent agents, key information can be adaptively identified and extracted, effectively filtering out redundancy and noise interference. The capsule network can capture the hierarchical relationships and spatial features in the adjustment plan, and a dynamic routing mechanism can achieve accurate assessment of the feasibility of the plan.

[0132] In existing technologies, tool wear status monitoring and machining parameter optimization mainly rely on a single intelligent agent for decision-making. This suffers from problems such as incomplete information acquisition, a single optimization strategy, and poor adaptability. Simple linear weighting or rule matching is used for information fusion and decision optimization, which cannot effectively handle complex and changing machining environments, resulting in unsatisfactory optimization results.

[0133] The information sharing mechanism between intelligent agents in this embodiment ensures real-time grasp of the global state, providing 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 deep 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 processing process and maintain continuous optimization effects, providing a more advanced technical solution for realizing intelligent regulation of the processing process.

[0134] Figure 2 This is a visualization of the capsule network hierarchy and routing mechanism used in the method for identifying tool wear and adjusting machining parameters for intelligent machine tools according to an embodiment of the present invention. It illustrates the complete workflow of the capsule network hierarchy and routing mechanism employed in this technical solution. The structure consists of five main components: an input layer, an initial capsule layer, a transformation matrix and routing mechanism, an output capsule layer, and a prediction result layer.

[0135] At the input layer, the system collects seven key types of machining data, including the three-dimensional components of cutting force (X, Y, and Z), vibration acceleration signals, temperature distribution, acoustic emission signals, current load, and the current wear of the tool (0.28mm) and remaining tool life (6.5h). This multi-dimensional, heterogeneous data serves as the initial input to the network and provides the foundation for subsequent intelligent analysis.

[0136] The initial capsule layer consists of six capsule units, each of which uses an 8-dimensional vector to represent features, capturing key information such as cutting force patterns, vibration spectra, temperature gradients, acoustic emission energy, current fluctuations, and wear trends. Unlike traditional convolutional neural networks, capsule networks use vectors rather than scalars to represent features, enabling them to retain richer spatial hierarchical relationship information.

[0137] In the transformation matrix and routing mechanism, the feature transformation matrix W[i, j] is used to linearly transform the initial capsule's feature vector to generate a "prediction vector." A dynamic routing algorithm then performs three key steps: feature extraction, compression activation, and feature integration. The routing algorithm iterates three times, calculating weight coefficients to determine the importance of different features and achieve effective feature aggregation.

[0138] The output capsule layer contains four specialized capsules, one for feed rate optimization (confidence 0.93), one for spindle speed adjustment (confidence 0.89), one for depth of cut control (confidence 0.87), and one for cooling parameter optimization (confidence 0.84). Each capsule uses dynamic routing to extract information from the previous layer and generate optimization recommendations for specific parameters.

[0139] The prediction result layer shows the final output of the model, including the processing quality score (95.3), wear prediction (currently 0.28mm, 0.31mm after 1 hour, and 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 preserve the spatial hierarchical relationship in the cutting process, significantly improving the accuracy of tool life prediction (94.8%) and the effect of machining parameter optimization, thereby effectively extending the tool life.

[0141] In an optional embodiment,

[0142] Obtaining an output capsule of the capsule network according to the product of the routing weight and the output of the initial capsule layer through a compression function, and predicting a processing quality score corresponding to the initial adjustment solution based on the output capsule includes:

[0143] The routing weights in the capsule network and the output of the initial capsule layer are constructed as nodes in a Markov random field structure, the edge connection probability is calculated based on the spatial position relationship and feature similarity between the nodes, the edge connection probability is constructed as a potential function, a single-node potential energy function based on a distribution parameter is constructed for the node, and a joint probability distribution of the node state is constructed by multiplying the single-node potential energy function and the potential function;

[0144] Based on the joint probability distribution, the state value of the node is initialized according to Gibbs sampling, the conditional probability of the node is calculated based on the single-node potential energy function and the potential function, the node state is iteratively updated according to the conditional probability to obtain a sampling sequence, the autocorrelation of the sampling sequence is monitored, and the sampling step size and the number of iterations are dynamically adjusted until a preset convergence condition is met;

[0145] Calculating the sampling sequence through marginal probability distribution to obtain an optimized routing weight, multiplying the optimized routing weight by the output of the initial capsule layer to obtain a product result, and inputting the product result into a compression function to obtain an output capsule of the capsule network;

[0146] An initial score is obtained by calculating the inner product of the output capsule and a preset parameter vector. Based on the initial score, a normalized score is obtained through a nonlinear activation function and used as the processing quality score corresponding to the initial adjustment scheme.

[0147] The routing weights and initial capsule layer outputs in the capsule network are mapped to nodes in a random field. A multidimensional feature vector is extracted for each node, containing information such as the node's spatial coordinates, topological structure, and numerical distribution. The spatial correlation and feature similarity between nodes are evaluated by calculating the Euclidean distance and cosine similarity between them. Based on the evaluation results, an edge connection probability matrix is ​​constructed, which describes the connection strength between nodes. The edge connection probabilities are converted into potential functions to characterize the interactions between nodes.

[0148] For each independent node, a single-node potential energy function is constructed based on its characteristic distribution parameters. This single-node potential energy function takes into account multiple factors, including the node's prior probability distribution, numerical range constraints, and gradient variation characteristics. By probabilistically multiplying the single-node potential energy function with the edge connection potential function, a complete joint probability distribution model of the node state is obtained, which fully describes the state combinations and occurrence probabilities of all nodes in the random field.

[0149] The Gibbs sampling method is used to optimize and update node states. Each node's state value is randomly initialized based on its prior distribution. In each sampling round, the states of other nodes are fixed, and the conditional probability distribution of the target node is calculated based on the potential energy function of the single node and the potential functions of its neighboring nodes. A new state value is extracted from the conditional probability distribution to update the target node's state. Following the pre-set node traversal order, the states of all nodes are updated sequentially, generating a complete state sampling sequence.

[0150] During the sampling process, the autocorrelation coefficient of the sampling sequence is calculated in real time to monitor the degree of mixing in the sampling chain. When the autocorrelation coefficient is high, the sampling step size is appropriately increased to accelerate the exploration of the state space. When the autocorrelation coefficient decreases, the sampling step size is reduced to improve the precision of the local search. The number of iterations is also dynamically adjusted based on the changing 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] Statistical analysis is performed on the obtained stable sampling sequence to calculate the marginal probability distribution of each node state. The optimal node state value is determined from the marginal probability distribution, resulting in an optimized routing weight. The optimized routing weight is then multiplied by the initial capsule layer output to obtain a weighted feature representation. This weighted feature representation retains the important information of the original features while highlighting the contribution of key features through the optimized weights.

[0152] The weighted feature representation is fed into a specially designed compression function, which uses a nonlinear transformation to map the features into a new representation space. This compression function produces output capsules with excellent feature representation capabilities. These capsules contain optimized feature information and effectively represent the key characteristics of the current processing state.

[0153] Perform a processing quality score. Perform an inner product operation on the output capsule and the pre-trained parameter vector to obtain an initial quality score. This score reflects the degree of match between the current features and the ideal state. This initial score is input into a customized nonlinear activation function and normalized to obtain the final processing quality score.

[0154] For example, in a CNC milling process, a capsule network consists of thirty routing weight nodes and forty initial capsule layer output nodes. These nodes are constructed as a Markov random field structure, and edge connection probabilities are calculated based on the spatial distribution characteristics of the nodes. Gibbs sampling is used for one thousand iterations, and sampling is stopped when the autocorrelation coefficient of the sampling sequence falls below a preset threshold. The marginal distribution of the sampling sequence is statistically analyzed to obtain optimized routing weights. These weights are then multiplied by the initial capsule layer output and applied to a compression function to obtain the output capsule. Finally, the inner product of the output capsule and a preset parameter vector is calculated and normalized to obtain a score reflecting the machining quality.

[0155] In this embodiment, a probabilistic dependency model between nodes is established by constructing routing weights and capsule outputs into a Markov random field structure. A potential function is used to describe the interaction between nodes, including both the characteristic potential of a single node and the edge connection potential between nodes. Intelligent optimization of node states is achieved through the Gibbs sampling method. The convergence and effectiveness of the sampling process are ensured by dynamically adjusting the sampling step size and monitoring autocorrelation.

[0156] In existing technologies, the routing mechanism of capsule networks mainly uses a simple dynamic routing algorithm, which optimizes by directly iteratively updating routing weights. This ignores the spatial correlation between routing weights, causing the optimization process to easily fall into local optimality. The lack of probabilistic modeling of feature distribution makes weight updates lack theoretical support, making it difficult to ensure the convergence of the optimization process and the reliability of the optimization results insufficient.

[0157] This embodiment fully considers the spatial correlation and numerical similarity between features, so that the optimization process can take into account both local features and global structure, significantly improving the accuracy and completeness of feature extraction. The double-layer potential energy structure provides a clear theoretical basis for the optimization process, and can explain and guide the weight update process from a probabilistic perspective, greatly improving 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 optimality, ensuring the global optimality of the optimization results, and solving the problems existing in the traditional routing mechanism. At the same time, it also provides a new technical idea for the performance optimization of capsule networks in complex application scenarios, which has important theoretical value and practical significance.

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

[0159] Compared with the traditional dynamic routing algorithm, this technical solution improves the surface smoothness by about 16 times and by about 30 times compared with 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] A second aspect of an embodiment of the present invention provides a system for identifying tool wear status and adjusting machining parameters of an intelligent machine tool, comprising:

[0161] The first unit is used to collect the working condition data of the intelligent machine tool during the processing and perform sliding window segmentation, and perform 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 the standard working condition data through a bidirectional gated cyclic unit, input the long-term and short-term time series feature sequences into the Hilbert-Huang transform decomposition to obtain intrinsic modal components, combine the four-layer spatial pyramid pooling to extract the multi-scale frequency domain feature sequences, calculate the mutual information entropy of 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 weighted analysis 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 wear value and the remaining service life value of the tool;

[0164] The fourth unit is used to add the current wear value and the remaining service life value of the tool to the parallel agent structure, determine an initial adjustment plan based on the observation shared information collected by each agent, predict the processing quality score corresponding to the initial adjustment plan through the capsule network, and feed the processing quality score back to the parallel agent structure as a reward signal 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 optimization correction strategy and add them to the machine tool process parameter library.

[0166] According to a third aspect of an embodiment 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 aforementioned method.

[0168] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0169] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying tool wear status and adjusting machining parameters of intelligent machine tools, characterized in that: include: Collect the working condition data of the intelligent machine tool during machining and perform sliding window segmentation. Then, perform denoising and normalization on the segmented working condition data to obtain standard working condition data. A bidirectional gated recurrent unit is used to extract long- and short-term time series feature sequences from standard working condition data. The long- and short-term time series feature sequences are input into the Hilbert-Huang transform decomposition to obtain intrinsic modal components. A multi-scale frequency domain feature sequence is extracted by combining the four-layer spatial pyramid pooling. The mutual information entropy of the time series feature sequence and the frequency domain feature sequence is calculated. The fusion weight coefficient is determined according to the mutual information entropy value and weighted to obtain the tool wear state feature vector. Inputting the tool wear state feature vector into a pre-trained extreme learning machine to calculate the current wear value and the remaining service life value of the tool; The current wear value and remaining service life value of the tool 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 the capsule network. The processing quality score is fed back to the parallel agent structure as a reward signal to obtain the optimized correction strategy. Adjust the working parameters of the intelligent machine tool based on the optimization correction strategy and add them to the machine tool process parameter library.

2. The method according to claim 1, characterized in that The working condition data of the intelligent machine tool during machining is collected and segmented by sliding windows. The segmented working condition data is denoised and normalized to obtain standard working condition data including: Collecting working condition data during the intelligent machine tool processing, the working condition data including spindle current signal, spindle speed signal, feed speed signal and cutting force signal; Perform sliding window segmentation on the operating condition data, with an overlap rate of 50% between adjacent windows, to obtain a segmented operating condition data sequence; For the segmented operating condition data sequence, the signal is decomposed and reconstructed at multiple scales using the wavelet threshold method to remove noise. A Butterworth low-pass filter is used to filter out high-frequency interference components that are 50% higher than the sampling frequency, and the filtered signal is normalized to the maximum and minimum values ​​to obtain standard operating condition data.

3. The method according to claim 1, characterized in that 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 decomposition to obtain the intrinsic modal components. The multi-scale frequency domain feature sequences are extracted by combining the four-layer spatial pyramid pooling, including: The standard working condition data is input into the bidirectional gated recurrent unit network, and the forward hidden layer state and the backward hidden layer state are established in the bidirectional gated recurrent unit network. The forward hidden layer state and the backward hidden layer state are updated through the reset gate and the update gate respectively, and the updated forward hidden layer state and the backward hidden layer state are spliced ​​to generate a long-term and short-term time series feature sequence; The long-term and short-term time series feature sequences are subjected to empirical mode decomposition to obtain intrinsic mode function components, the intrinsic mode function components are subjected to Hilbert transform to obtain Hilbert transform components, an analytical signal is constructed based on the intrinsic mode function components and the Hilbert transform components, and the instantaneous amplitude and instantaneous phase are extracted from the analytical signal to form the intrinsic mode components; In a four-layer spatial pyramid pooling network, pooling scales of one by one, two by two, four by four, and eight by eight are set respectively, the intrinsic modal component is input into the four-layer spatial pyramid pooling network, a pooling output feature map is generated, and a maximum pooling operation is performed to obtain four layers of pooled features, and the four layers of pooled features are spliced ​​to form a multi-scale frequency domain feature sequence.

4. The method according to claim 1, wherein Calculating the mutual information entropy of the time series feature sequence and the frequency domain feature sequence, determining the fusion weight coefficient according to the mutual information entropy value, and weighting to obtain the tool wear state feature vector includes: The difference between the eigenvalue and the corresponding eigenmean in the time series feature sequence and the frequency domain feature sequence is divided by the bandwidth parameter to obtain the standardized eigenvalue. The standardized eigenvalue is substituted into the Gaussian kernel function. The result of the Gaussian kernel function is divided by the product of the number of samples and the bandwidth parameter to obtain the probability distribution value. The mutual information entropy between the two feature sequences is calculated by multiplying the logarithms of the components of the time series feature sequence and the frequency domain feature sequence and summing them up. The mutual information entropy value is negative and multiplied by the adjustment parameter and the exponential operation is performed to obtain the weight coefficient numerator. The weight coefficient numerator is divided by the sum of the weight coefficient numerator and its complement to obtain the fusion weight coefficient of the time series feature sequence. The fusion weight coefficient of the time series feature sequence is subtracted from 1 to obtain the fusion weight coefficient of the frequency domain feature sequence. Multiply the time series feature sequence with the fusion weight coefficient of the time series feature sequence to obtain the weighted time series feature, multiply the frequency domain feature sequence with 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; The eigenvector corresponding to the maximum eigenvalue in the fusion feature sequence is extracted as the main feature component to construct the tool wear state feature vector.

5. The method according to claim 1, characterized in that The tool wear state feature vector is input into a pre-trained extreme learning machine to calculate the current wear value and the remaining service life value of the tool, including: Collecting tool wear state sample data corresponding to a tool wear state feature vector, dividing the tool wear state sample data into a training data set and a test data set, training an extreme learning machine based on the training data set, wherein 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 a Sigmoid activation function, randomly generates initial values ​​of an input weight matrix and a bias vector, multiplies the tool wear state feature vector in the training data set by the input weight matrix and adds the result to the bias vector, obtains a hidden layer output through the Sigmoid activation function, calculates an output weight based on the hidden layer output, and obtains a pre-trained extreme learning machine; The tool wear state feature vector is input into the pre-trained extreme learning machine, and the tool wear state feature vector is multiplied by the input weight matrix and added to the bias vector, and then the result is activated by the Sigmoid function to obtain the hidden layer output. The current wear value and the remaining service life value of the tool are calculated based on the hidden layer output.

6. The method according to claim 1, characterized in that The current wear value and the remaining service life value of the tool 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 by the capsule network. The processing quality score is fed back to the parallel agent structure as a reward signal. The optimization correction strategy includes: Adding the current wear value and the remaining service life value of the tool to a parallel agent structure, wherein the parallel agent structure includes multiple parallel agents, each of which collects local observation shared information, and obtains global observation shared information by exchanging the local observation shared information through a shared network; Calculating the attention weight coefficients between the parallel agents based on the global observation sharing information, wherein the attention weight coefficients are obtained by performing an inner product operation on the query vector corresponding to the parallel agent and the key vector, performing weighted accumulation on the attention weight coefficients and the corresponding value vectors to obtain a fused feature representation, and determining an initial adjustment scheme based on the fused feature representation and the observation sharing information collected by each of the parallel agents; Inputting the initial adjustment plan into a capsule network, transforming the initial adjustment plan using a transformation matrix of the capsule network to obtain an initial capsule layer output, calculating a routing weight based on the initial capsule layer output, obtaining an output capsule of the capsule network based on a compression function obtained by multiplying the routing weight by the initial capsule layer output, and predicting a processing quality score corresponding to the initial adjustment plan based on the output capsule; The processing quality score is fed back to the parallel intelligent agent structure as a reward signal. The parallel intelligent agent structure encodes the processing quality score to obtain the historical information code. The historical information code and the current state code obtained in real time are input into the strategy network to generate the probability distribution of the adjustment action. The optimization is carried out in combination with the strategy optimization criterion to obtain the optimized correction strategy.

7. The method according to claim 6, characterized in that Obtaining an output capsule of the capsule network according to the product of the routing weight and the output of the initial capsule layer through a compression function, and predicting a processing quality score corresponding to the initial adjustment solution based on the output capsule includes: The routing weights in the capsule network and the output of the initial capsule layer are constructed as nodes in a Markov random field structure, the edge connection probability is calculated based on the spatial position relationship and feature similarity between the nodes, the edge connection probability is constructed as a potential function, a single-node potential energy function based on a distribution parameter is constructed for the node, and a joint probability distribution of the node state is constructed by multiplying the single-node potential energy function and the potential function; Based on the joint probability distribution, the state value of the node is initialized according to Gibbs sampling, the conditional probability of the node is calculated based on the single-node potential energy function and the potential function, the node state is iteratively updated according to the conditional probability to obtain a sampling sequence, the autocorrelation of the sampling sequence is monitored, and the sampling step size and the number of iterations are dynamically adjusted until a preset convergence condition is met; Calculating the sampling sequence through marginal probability distribution to obtain an optimized routing weight, multiplying the optimized routing weight by the output of the initial capsule layer to obtain a product result, and inputting the product result into a compression function to obtain an output capsule of the capsule network; An initial score is obtained by calculating the inner product of the output capsule and a preset parameter vector. Based on the initial score, a normalized score is obtained through a nonlinear activation function and used as the processing quality score corresponding to the initial adjustment scheme.

8. Intelligent machine tool tool wear state identification and processing parameter adjustment system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect the working condition data of the intelligent machine tool during the processing and perform sliding window segmentation, and perform denoising and normalization processing on the segmented working condition data to obtain standard working condition data; The second unit is used to extract long-term and short-term time series feature sequences from the standard working condition data through a bidirectional gated cyclic unit, input the long-term and short-term time series feature sequences into the Hilbert-Huang transform decomposition to obtain intrinsic modal components, combine the four-layer spatial pyramid pooling to extract the multi-scale frequency domain feature sequences, calculate the mutual information entropy of 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 weighted analysis to obtain the tool wear state feature vector; The third unit is used to input the tool wear state feature vector into a pre-trained extreme learning machine to calculate the current wear value and the remaining service life value of the tool; The fourth unit is used to add the current wear value and the remaining service life value of the tool to the parallel agent structure, determine an initial adjustment plan based on the observation shared information collected by each agent, predict the processing quality score corresponding to the initial adjustment plan through the capsule network, and feed the processing quality score back to the parallel agent structure as a reward signal to obtain an optimized correction strategy; The fifth unit is used to adjust the working parameters of the intelligent machine tool based on the optimization correction strategy and add them to the machine tool process parameter library.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; 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: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images