Cutter degradation trend analysis method based on similarity measurement

By collecting and processing multiple state signals in real time, combining technologies such as sparse adaptive decomposition and adversarial similarity measurement network, the precise monitoring of tool state and degradation trend analysis are achieved, and the low accuracy and lack of real-time monitoring in the existing technology are solved. Problem, optimized maintenance strategies, extended tool life and improved production efficiency and product quality.

CN119025965APending Publication Date: 2024-11-26YOUJI TECH (SHANGHAI) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411124450.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing tool status monitoring and degradation trend analysis methods have low accuracy and reliability, lack real-time monitoring and preventive maintenance capabilities, making it difficult to accurately predict tool degradation trends and remaining life, and maintenance strategies lack intelligence and dynamic optimization.

Method used

Using real-time acquisition and processing of multiple state signals, combining sparse adaptive decomposition algorithms, variational autoencoders, adversarial similarity measurement networks and Bayesian non-parametric models, real-time monitoring and prevention are achieved through deep Q network optimization maintenance strategies. Sexual maintenance.

Benefits of technology

It significantly improves the accuracy and reliability of tool status monitoring, realizes accurate tracking and prediction of tool degradation trends, optimizes maintenance strategies, extends tool service life, reduces production costs, and improves production efficiency and product quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025965A_ABST
    Figure CN119025965A_ABST
Patent Text Reader

Abstract

The invention discloses a similarity measurement-based cutter degradation trend analysis method, which comprises the following steps of: acquiring a cutter state signal by using a vibration sensor, an acoustic emission sensor and a temperature sensor, and carrying out noise removal and standardization processing; extracting high-order statistical features, sparse coding features and deep learning features through an improved sparse adaptive decomposition algorithm; a variational automatic encoder and a self-adaptive multi-modal fusion dimension reduction technology are adopted to carry out dimension reduction processing; calculating the signal similarity through an adversarial similarity measurement network; according to a similarity result, drawing a cutter degradation trend chart by adopting a Bayesian nonparametric model, and dynamically adjusting model parameters to capture change details; dividing the degradation process into a plurality of stages by using a multi-level degradation stage recognition algorithm; and finally, proposing maintenance and replacement suggestions, and optimizing a strategy through a dynamic maintenance strategy optimization algorithm to realize real-time monitoring and preventive maintenance of the cutter.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of mechanical processing, and in particular to a tool degradation trend analysis method based on similarity measurement. Background Art

[0002] In the field of modern machining, the service life and performance of cutting tools are crucial to production efficiency and product quality. However, cutting tools will gradually wear out during the machining process, resulting in reduced machining quality and increased production costs. Therefore, how to effectively monitor the status of cutting tools and predict their degradation trends has become an important research topic. Existing methods for tool status monitoring and degradation trend analysis mainly include experience-based maintenance, replacement strategies, and monitoring methods based on traditional signal processing and pattern recognition technologies.

[0003] In the prior art, the methods of tool condition monitoring and degradation trend analysis are mainly divided into experience-based maintenance strategies and monitoring methods based on signal processing and pattern recognition technology. However, these methods have many defects and shortcomings. Traditional maintenance and replacement strategies mainly rely on the operator's experience and the preset time or number of uses. This method often leads to premature or late replacement of tools. Premature replacement will increase production costs, while late replacement may lead to a decline in processing quality and even cause safety accidents. In addition, the experience-based maintenance strategy lacks scientificity and flexibility, and it is difficult to adapt to tool wear under different processing conditions. The working conditions in each processing process may be different, and the wear speed and pattern of the tool are also different. It is difficult to accurately judge the best maintenance and replacement time based on experience alone. At the same time, this method usually performs regular inspections or maintenance based on usage time, and cannot achieve real-time monitoring and preventive maintenance of the tool status.

[0004] Monitoring methods based on signal processing and pattern recognition also have significant limitations. In the prior art, one of the vibration signals, acoustic emission signals or temperature signals is usually used to monitor the tool status. This method can only reflect one aspect of the tool status, and it is difficult to fully reflect the comprehensive operating status of the tool, resulting in low monitoring accuracy and reliability. Traditional feature extraction methods such as Fourier transform and wavelet transform are mainly used to extract features in the time domain, frequency domain or time-frequency domain, but these methods cannot effectively process complex multimodal signals and nonlinear relationships, resulting in the extracted features being insufficient to fully reflect the degradation status of the tool. Traditional pattern recognition methods have limited performance when processing multimodal and high-dimensional data, and cannot fully capture the complex relationship between signal features, resulting in low accuracy and reliability of tool status identification. In addition, the prior art can usually only monitor the current status of the tool, lacks effective dynamic trend analysis methods, and is difficult to predict the degradation trend and remaining life of the tool. This limits the comprehensive grasp of the tool status and the implementation of preventive maintenance.

[0005] The performance of existing models under different working conditions and environments is often unstable, lacking sufficient robustness and generalization capabilities. Different processing conditions and environmental changes may cause model failure or inaccurate prediction results. In the prior art, the evaluation and optimization of maintenance strategies are usually based on offline analysis of historical data, lacking real-time feedback and dynamic adjustment mechanisms. This results in the inability of maintenance strategies to respond to changes in tool status in a timely manner, reducing the effectiveness of maintenance. Existing maintenance and replacement strategies are usually based on fixed rules or simple threshold judgments, lacking intelligence and dynamic optimization capabilities. This results in the inability to adjust maintenance strategies based on real-time data and trend changes, making it difficult to achieve optimal maintenance and replacement decisions. These shortcomings significantly affect the effectiveness of tool status monitoring and degradation trend analysis, making it difficult to achieve comprehensive and accurate monitoring and prediction of tool status.

[0006] Therefore, how to provide a tool degradation trend analysis method based on similarity measurement is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] One purpose of the present invention is to propose a tool degradation trend analysis method based on similarity measurement. The present invention utilizes multiple state signals, advanced feature extraction and dimensionality reduction techniques, adversarial similarity measurement networks, Bayesian non-parametric models, and dynamic maintenance strategy optimization algorithms, and describes in detail the method of real-time monitoring of tool status and degradation trend analysis. The flexibility and accuracy of signal comparison are improved by using multiple similarity measurement methods; the introduction of variational autoencoders and adaptive multimodal fusion dimensionality reduction techniques reduces data redundancy and improves the effectiveness of feature extraction; the Bayesian non-parametric model is used to realize dynamic trend analysis, which can track the degradation process of the tool in real time and dynamically; the maintenance and replacement strategies are optimized by reinforcement learning and deep Q networks, which realizes real-time monitoring and preventive maintenance of tools and improves the scientificity and flexibility of maintenance strategies.

[0008] A tool degradation trend analysis method based on similarity measurement according to an embodiment of the present invention comprises the following steps:

[0009] S1, collecting the state signal of the tool in the machining process in real time through the vibration sensor, acoustic emission sensor and temperature sensor, and performing noise removal, filtering and standardization processing on the collected state signal;

[0010] S2, using the improved sparse adaptive decomposition algorithm to extract high-order statistical features, sparse coding features and deep learning features from the preprocessed signal;

[0011] S3, using variational autoencoder combined with adaptive multimodal fusion dimensionality reduction technology to reduce the dimensionality of the extracted features, compressing and representing the features through the variational autoencoder model, and using adaptive multimodal fusion dimensionality reduction technology to fuse different modal features and extract representative main feature components;

[0012] S4, calculating the similarity between the tool state signal and the reference signal at different times through an adversarial similarity measurement network, wherein the adversarial similarity measurement network uses a generative adversarial network to perform adversarial training on the signal to generate a virtual signal similar to the reference signal, and calculating the similarity measurement value between the signals through a discriminator;

[0013] S5. According to the similarity measurement results, an adaptive dynamic trend analysis based on the Bayesian non-parametric model is used to draw a tool degradation trend diagram, dynamically adjust the parameters of the trend analysis model to capture the change details in the tool degradation process, and show the trend of signal similarity changing over time;

[0014] S6. Identify the degradation stage of the tool through the trend chart, and use the multi-level degradation stage identification algorithm to divide the degradation process into multiple stages, including the initial running-in stage, stable use stage and degradation stage, and classify them according to the characteristics of different stages;

[0015] S7. According to the degradation trend diagram, tool maintenance and replacement suggestions are proposed, and the dynamic maintenance strategy optimization algorithm is used to optimize the maintenance and replacement strategies to achieve real-time monitoring and preventive maintenance of the tools.

[0016] Optionally, the S2 specifically includes:

[0017] S21, performing segment processing on the pre-processed state signal, dividing the signal into several segments, each of which has a length of L;

[0018] S22. Decomposing each signal segment using an improved sparse adaptive decomposition algorithm, wherein the improved sparse adaptive decomposition algorithm includes an optimization process of dynamically adjusting a sparse base dictionary and an adaptive sparse regularization parameter:

[0019]

[0020] Where X is the preprocessed signal matrix, D is the dynamically adjusted sparse basis dictionary matrix, α is the sparse coefficient vector, and λ(t) is the sparse regularization parameter that changes dynamically over time;

[0021] S23, by solving the above objective function, a sparse representation coefficient vector α of each signal segment is obtained, and sparse coding features are extracted;

[0022] S24. Calculate the high-order statistical features of each signal segment, including mean μ, variance σ2 , skewness γ and kurtosis κ:

[0023]

[0024] Among them, x i is the signal sample point, N is the number of samples;

[0025] S25. Use graph neural network to further process sparse coding features and high-order statistical features, construct a graph convolutional neural network, input sparse coding features and high-order statistical features, use the correlation between nodes to extract signal features, and output enhanced features.

[0026] Optionally, the S3 specifically includes:

[0027] S31, preliminarily integrating the sparse coding features, high-order statistical features and enhanced features extracted by the graph neural network to form an input feature vector φ;

[0028] S32, using a variational autoencoder to reduce the dimension of the input feature vector φ, the variational autoencoder includes an encoder and a decoder, the encoder compresses the input feature vector φ to a latent space representation z, and the decoder reconstructs the feature vector from the latent space representation z

[0029] L=E q(z|φ) [log p(φ|z)]-KL[q(z|φ)||p(z)];

[0030] Where q(z|φ) is the probability distribution of the encoder, p(z) is the prior distribution, and KL is the Kullback-Leibler divergence;

[0031] S33. The weighted fusion of different modal features is performed through adaptive multimodal fusion dimensionality reduction technology, and the weight of each modal feature is adaptively adjusted to extract the representative main feature components:

[0032]

[0033] Among them, φ i is the eigenvector of the ith mode, w i is the weight of the i-th mode, and M is the number of modes;

[0034] S34, combine the latent space representation z output by the variational autoencoder and the main feature components extracted by the adaptive multimodal fusion dimensionality reduction technology to generate the final reduced-dimensional feature vector φ final :

[0035]

[0036] Among them, fj (z,φ i ) is the feature function extracted by the variational autoencoder and adaptive multimodal fusion dimensionality reduction technology, β j is the dynamic weight in the fusion process, and N is the number of feature functions.

[0037] Optionally, the S4 specifically includes:

[0038] S41, constructing an adversarial similarity measurement network, the network comprising a generator G and a discriminator D, the generator G generates a virtual signal similar to a reference signal, and the discriminator D is used to distinguish between a real signal and a virtual signal;

[0039] S42, generator G inputs feature vector φ final Generate virtual feature vector φ gen :

[0040]

[0041] Among them, G(φ final )=φ gen , D(φ gen ) is the discriminant result of the discriminator on the generated feature vector;

[0042] S43, the discriminator D inputs the true feature vector φ final and the virtual eigenvector φ gen To determine:

[0043]

[0044] Among them, D(φ final ) is the discriminator’s discriminant result on the real feature vector, D(φ gen ) is the discriminant result of the virtual feature vector;

[0045] S44, through adversarial training, the virtual feature vector φ generated by the generator G gen With the true eigenvector φ final The similarity measure between them is maximized:

[0046]

[0047] Among them, · represents the vector inner product, ||·|| represents the Euclidean norm of the vector;

[0048] S45. Calculate the true feature vector φ according to the output result of the discriminator D final and the virtual eigenvector φ gen The similarity measurement value between them is taken as the final similarity measurement result.

[0049] Optionally, the S5 specifically includes:

[0050] S51. Based on the similarity measurement results calculated by the adversarial similarity measurement network, an adaptive dynamic trend analysis method based on the Bayesian non-parametric model is used to model and analyze the tool degradation process;

[0051] S52. Constructing a Bayesian non-parametric model, the model includes a Gaussian process and a dynamic Bayesian network, the Gaussian process is used to capture the continuous change characteristics in the tool degradation process, and the dynamic Bayesian network is used to capture the discrete stage change characteristics in the tool degradation process;

[0052] S53. Use Gaussian process to predict tool degradation trend, establish the relationship between similarity measurement result vector y and time series, and capture the correlation between signal features through the covariance matrix K of Gaussian process to achieve smooth prediction of tool degradation trend:

[0053]

[0054] Among them, y is the similarity measurement result vector, K is the covariance matrix, and n is the number of samples;

[0055] S54. Dynamic Bayesian network is used to identify and classify different stages in the tool degradation process. Dynamic Bayesian network models the state at different time points through probability transfer matrix. Through dynamic Bayesian network, the key stages in the tool degradation process can be accurately identified and the degradation characteristics of different stages can be classified:

[0056]

[0057] Among them, X t and X t+1 denote the states at time t and time t+1 respectively, P(X t |X t+1 ) is the transition probability between time t and time t+1;

[0058] S55. Dynamically adjust the tool degradation trend through the Bayesian non-parametric model, optimize the model parameters to adaptively capture the details of the changes in the tool degradation process, and the Bayesian update process can dynamically adjust the model parameters to adapt to the changes in the tool degradation process;

[0059] S56. Based on the dynamic adjustment results of the Bayesian non-parametric model, a tool degradation trend graph is drawn to show the trend of signal similarity over time. Combined with the prediction results of the Gaussian process and dynamic Bayesian network, real-time analysis and prediction of tool degradation are provided.

[0060] Optionally, the S6 specifically includes:

[0061] S61, preliminarily identifying the degradation stage of the tool according to the drawn tool degradation trend diagram, and further classifying and labeling each stage in the tool degradation process using a multi-level degradation stage identification algorithm;

[0062] S62, calculating the signal similarity metric value at each time point, and determining the initial degradation stage boundary according to the change trend of the similarity metric value;

[0063] S63. Evaluate the signal complexity of each initial degradation stage through multi-scale entropy analysis, and calculate the multi-scale entropy value of each stage:

[0064]

[0065] Among them, p i is the probability distribution of the signal at the i-th scale, and n is the number of scales;

[0066] S64. Based on the results of multi-scale entropy analysis, the degradation stage boundaries are further refined and the degradation stages are identified, including the initial running-in stage, the stable use stage and the degradation stage;

[0067] S65. According to the results of the multi-level degradation stage identification algorithm, the tool degradation process is divided into several specific stages, and the mixed effect model is used to classify and model the characteristics of different stages:

[0068] Y ij =β0+β1X ij +b i +∈ ij ;

[0069] Among them, Y ij is the observed value of the i-th stage at the j-th time point, β0 and β1 are fixed effect parameters, X ij is the predictor variable of the observed value, b i is a random effect, ∈ ij is the error term;

[0070] S66. Based on the classification results of the mixed effect model, the characteristics of different degradation stages are modeled and analyzed, the key characteristic parameters of each stage are extracted, and the characteristic curves of each stage are drawn;

[0071] S67. Combining the classification results of the mixed effect model and the characteristic parameters of each stage, a complete tool degradation stage identification model is formed.

[0072] Optionally, the S7 specifically includes:

[0073] S71. Based on the degradation trend diagram and tool degradation stage identification model, make suggestions for tool maintenance and replacement, and preliminarily determine the maintenance and replacement strategy, including maintenance time point, maintenance content and replacement standards;

[0074] S72. Use a dynamic maintenance strategy optimization algorithm to optimize the maintenance and replacement strategies initially formulated, and construct a dynamic maintenance strategy optimization model based on reinforcement learning, the model including a state space, an action space, and a reward function, the state space including the current state and degradation stage of the tool, the action space including different strategies for maintenance and replacement, and the reward function for evaluating the effect of each strategy;

[0075] S73. Use deep Q network to optimize dynamic maintenance strategy. Deep Q network optimizes maintenance and replacement strategy by combining neural network and Q learning algorithm:

[0076] Q(s,a)←Q(s,a)+α[r+γmax a' Q(s′,a′)-Q(s,a)];

[0077] Where Q(s,a) is the Q value of taking action a in state s, α is the learning rate, r is the immediate reward, γ is the discount factor, s' is the new state after executing the action, and a' is the best action in the new state;

[0078] S74. Through multiple rounds of training and simulation experiments, the maintenance and replacement strategies are optimized so that the maximum cumulative reward can be obtained at different degradation stages, and real-time monitoring and preventive maintenance of the tool can be achieved;

[0079] S75. According to the optimized dynamic maintenance strategy, formulate a specific maintenance and replacement plan, including specific time points and operation steps, and apply it in the actual production process.

[0080] The beneficial effects of the present invention are:

[0081] The present invention collects various state signals of the tool in real time through vibration sensors, acoustic emission sensors and temperature sensors, and uses an improved sparse adaptive decomposition algorithm to extract high-order statistical features, sparse coding features and deep learning features, which comprehensively reflect the operating status of the tool and significantly improve the monitoring accuracy and reliability. Secondly, a variational autoencoder combined with an adaptive multimodal fusion dimensionality reduction technology is used to reduce the dimensionality of the extracted multimodal features, and compression and representation learning are performed through a variational autoencoder model. At the same time, adaptive multimodal fusion dimensionality reduction technology is used to fuse different modal features to extract representative main feature components, effectively reducing data redundancy and improving the effectiveness of feature extraction.

[0082] The present invention calculates the similarity between the tool state signal and the reference signal through an adversarial similarity measurement network, uses a generative adversarial network to perform adversarial training on the signal to generate a virtual signal similar to the reference signal, and calculates the similarity measurement value between the signals through a discriminator. This method improves the flexibility and accuracy of signal comparison. In addition, based on the similarity measurement results, an adaptive dynamic trend analysis method based on a Bayesian non-parametric model is used to draw a tool degradation trend graph, dynamically adjust the parameters of the trend analysis model to capture the details of the changes in the tool degradation process, and display the trend of signal similarity over time. Through a multi-level degradation stage recognition algorithm, the degradation process is divided into multiple stages, including an initial running-in stage, a stable use stage, and a degradation stage, and is classified according to the characteristics of different stages, thereby achieving accurate tracking and prediction of the tool degradation process.

[0083] According to the degradation trend graph, tool maintenance and replacement suggestions are proposed, and the dynamic maintenance strategy optimization algorithm is used to optimize the initially formulated maintenance and replacement strategy. A dynamic maintenance strategy optimization model based on reinforcement learning is constructed, and the dynamic maintenance strategy is optimized through a deep Q network to ensure that the maintenance and replacement strategy can obtain the maximum cumulative reward at different degradation stages, thereby realizing real-time monitoring and preventive maintenance of the tool, and improving the scientificity and flexibility of the maintenance strategy. The present invention realizes real-time monitoring and preventive maintenance of the tool status through the above-mentioned technical means, and can make maintenance and replacement suggestions before the tool is severely worn, avoiding production interruptions and quality problems caused by tool failure, and significantly improving production efficiency and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0085] Figure 1 A flowchart of a tool degradation trend analysis method based on similarity measurement proposed by the present invention;

[0086] Figure 2 A flowchart of extracting high-order statistical features, sparse coding features and deep learning features using the improved sparse adaptive decomposition algorithm of the present invention;

[0087] Figure 3 The present invention provides a flowchart for calculating the similarity between tool state signals and reference signals at different times using the adversarial similarity measurement network.

[0088] Figure 4 Schematic diagram of tool degradation trend diagram drawn by adaptive dynamic trend analysis method based on Bayesian non-parametric model in the present invention DETAILED DESCRIPTION

[0089] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0090] refer to Figure 1-4 ,The tool degradation trend analysis method based on similarity measurement includes the following steps:

[0091] S1, collecting the state signal of the tool in the machining process in real time through the vibration sensor, acoustic emission sensor and temperature sensor, and performing noise removal, filtering and standardization processing on the collected state signal;

[0092] S2, using the improved sparse adaptive decomposition algorithm to extract high-order statistical features, sparse coding features and deep learning features from the preprocessed signal;

[0093] S3, using variational autoencoder combined with adaptive multimodal fusion dimensionality reduction technology to reduce the dimensionality of the extracted features, compressing and representing the features through the variational autoencoder model, and using adaptive multimodal fusion dimensionality reduction technology to fuse different modal features and extract representative main feature components;

[0094] S4, calculating the similarity between the tool state signal and the reference signal at different times through an adversarial similarity measurement network, wherein the adversarial similarity measurement network uses a generative adversarial network to perform adversarial training on the signal to generate a virtual signal similar to the reference signal, and calculating the similarity measurement value between the signals through a discriminator;

[0095] S5. According to the similarity measurement results, an adaptive dynamic trend analysis based on the Bayesian non-parametric model is used to draw a tool degradation trend diagram, dynamically adjust the parameters of the trend analysis model to capture the change details in the tool degradation process, and show the trend of signal similarity changing over time;

[0096] S6. Identify the degradation stage of the tool through the trend chart, and use the multi-level degradation stage identification algorithm to divide the degradation process into multiple stages, including the initial running-in stage, stable use stage and degradation stage, and classify them according to the characteristics of different stages;

[0097] S7. According to the degradation trend diagram, tool maintenance and replacement suggestions are proposed, and the dynamic maintenance strategy optimization algorithm is used to optimize the maintenance and replacement strategies to achieve real-time monitoring and preventive maintenance of the tools.

[0098] In this implementation, S2 specifically includes:

[0099] S21, performing segment processing on the pre-processed state signal, dividing the signal into several segments, each of which has a length of L;

[0100] S22. Decomposing each signal segment using an improved sparse adaptive decomposition algorithm, wherein the improved sparse adaptive decomposition algorithm includes an optimization process of dynamically adjusting a sparse base dictionary and an adaptive sparse regularization parameter:

[0101]

[0102] Where X is the preprocessed signal matrix, D is the dynamically adjusted sparse basis dictionary matrix, α is the sparse coefficient vector, and λ(t) is the sparse regularization parameter that changes dynamically over time;

[0103] S23, by solving the above objective function, a sparse representation coefficient vector α of each signal segment is obtained, and sparse coding features are extracted;

[0104] S24. Calculate the high-order statistical features of each signal segment, including mean μ, variance σ 2 , skewness γ and kurtosis κ:

[0105]

[0106] Among them, x i is the signal sample point, N is the number of samples;

[0107] S25. Use graph neural network to further process sparse coding features and high-order statistical features, construct a graph convolutional neural network, input sparse coding features and high-order statistical features, use the correlation between nodes to extract signal features, and output enhanced features.

[0108] In this implementation, S3 specifically includes:

[0109] S31, preliminarily integrating the sparse coding features, high-order statistical features and enhanced features extracted by the graph neural network to form an input feature vector φ;

[0110] S32, using a variational autoencoder to reduce the dimension of the input feature vector φ, the variational autoencoder includes an encoder and a decoder, the encoder compresses the input feature vector φ to a latent space representation z, and the decoder reconstructs the feature vector from the latent space representation z

[0111] L=E q(z|φ) [log p(φ|z)]-KL[q(z|φ)||p(z)];

[0112] Where q(z|φ) is the probability distribution of the encoder, p(z) is the prior distribution, and KL is the Kullback-Leibler divergence;

[0113] S33, weighted fusion of different modal features is performed through adaptive multimodal fusion dimensionality reduction technology, the weight of each modal feature is adaptively adjusted, and representative main feature components are extracted:

[0114]

[0115] Among them, φ i is the eigenvector of the ith mode, w i is the weight of the i-th mode, and M is the number of modes;

[0116] S34, combine the latent space representation z output by the variational autoencoder and the main feature components extracted by the adaptive multimodal fusion dimensionality reduction technology to generate the final reduced-dimensional feature vector φ final :

[0117]

[0118] Among them, f j (z,φ i ) is the feature function extracted by the variational autoencoder and adaptive multimodal fusion dimensionality reduction technology, β j is the dynamic weight in the fusion process, and N is the number of feature functions.

[0119] In this implementation manner, the S4 specifically includes:

[0120] S41, constructing an adversarial similarity measurement network, the network comprising a generator G and a discriminator D, the generator G generates a virtual signal similar to a reference signal, and the discriminator D is used to distinguish between a real signal and a virtual signal;

[0121] S42, generator G inputs feature vector φ final Generate virtual feature vector φ gen :

[0122]

[0123] Among them, G(φ final )=φ gen , D(φ gen ) is the discriminant result of the discriminator on the generated feature vector;

[0124] S43, the discriminator D inputs the true feature vector φ final and the virtual eigenvector φ gen To determine:

[0125]

[0126] Among them, D(φ final ) is the discriminator’s discriminant result on the real feature vector, D(φgen ) is the discriminant result of the virtual feature vector;

[0127] S44, through adversarial training, the virtual feature vector φ generated by the generator G gen With the true eigenvector φ final The similarity measure between them is maximized:

[0128]

[0129] Among them, · represents the vector inner product, ||·|| represents the Euclidean norm of the vector;

[0130] S45. Calculate the true feature vector φ according to the output result of the discriminator D final and the virtual eigenvector φ gen The similarity measurement value between them is taken as the final similarity measurement result.

[0131] In this implementation manner, S5 specifically includes:

[0132] S51. Based on the similarity measurement results calculated by the adversarial similarity measurement network, an adaptive dynamic trend analysis method based on the Bayesian non-parametric model is used to model and analyze the tool degradation process;

[0133] S52. Constructing a Bayesian non-parametric model, the model includes a Gaussian process and a dynamic Bayesian network, the Gaussian process is used to capture the continuous change characteristics in the tool degradation process, and the dynamic Bayesian network is used to capture the discrete stage change characteristics in the tool degradation process;

[0134] S53. Use Gaussian process to predict tool degradation trend, establish the relationship between similarity measurement result vector y and time series, and capture the correlation between signal features through the covariance matrix K of Gaussian process to achieve smooth prediction of tool degradation trend:

[0135]

[0136] Among them, y is the similarity measurement result vector, K is the covariance matrix, and n is the number of samples;

[0137] S54. Dynamic Bayesian network is used to identify and classify different stages in the tool degradation process. Dynamic Bayesian network models the state at different time points through probability transfer matrix. Through dynamic Bayesian network, the key stages in the tool degradation process can be accurately identified and the degradation characteristics of different stages can be classified:

[0138]

[0139] Among them, X tand X t+1 denote the states at time t and time t+1 respectively, P(X t |X t+1 ) is the transition probability between time t and time t+1;

[0140] S55. Dynamically adjust the tool degradation trend through the Bayesian non-parametric model, optimize the model parameters to adaptively capture the details of the changes in the tool degradation process, and the Bayesian update process can dynamically adjust the model parameters to adapt to the changes in the tool degradation process;

[0141] S56. Based on the dynamic adjustment results of the Bayesian non-parametric model, a tool degradation trend graph is drawn to show the trend of signal similarity over time. Combined with the prediction results of the Gaussian process and dynamic Bayesian network, real-time analysis and prediction of tool degradation are provided.

[0142] In this implementation manner, S6 specifically includes:

[0143] S61, preliminarily identifying the degradation stage of the tool according to the drawn tool degradation trend diagram, and further classifying and labeling each stage in the tool degradation process using a multi-level degradation stage identification algorithm;

[0144] S62, calculating the signal similarity metric value at each time point, and determining the initial degradation stage boundary according to the change trend of the similarity metric value;

[0145] S63. Evaluate the signal complexity of each initial degradation stage through multi-scale entropy analysis, and calculate the multi-scale entropy value of each stage:

[0146]

[0147] Among them, p i is the probability distribution of the signal at the i-th scale, and n is the number of scales;

[0148] S64. Based on the results of multi-scale entropy analysis, the degradation stage boundaries are further refined and the degradation stages are identified, including the initial running-in stage, the stable use stage and the degradation stage;

[0149] S65. According to the results of the multi-level degradation stage identification algorithm, the tool degradation process is divided into several specific stages, and the mixed effect model is used to classify and model the characteristics of different stages:

[0150] Y ij =β0+β1X ij +b i +∈ ij ;

[0151] Among them, Y ijis the observed value of the i-th stage at the j-th time point, β0 and β1 are fixed effect parameters, X ij is the predictor variable of the observed value, b i is a random effect, ∈ ij is the error term;

[0152] S66. Based on the classification results of the mixed effect model, the characteristics of different degradation stages are modeled and analyzed, the key characteristic parameters of each stage are extracted, and the characteristic curves of each stage are drawn;

[0153] S67. Combining the classification results of the mixed effect model and the characteristic parameters of each stage, a complete tool degradation stage identification model is formed.

[0154] In this implementation manner, the S7 specifically includes:

[0155] S71. Based on the degradation trend diagram and tool degradation stage identification model, make suggestions for tool maintenance and replacement, and preliminarily determine the maintenance and replacement strategy, including maintenance time point, maintenance content and replacement standards;

[0156] S72. Use a dynamic maintenance strategy optimization algorithm to optimize the maintenance and replacement strategies initially formulated, and construct a dynamic maintenance strategy optimization model based on reinforcement learning, the model including a state space, an action space, and a reward function, the state space including the current state and degradation stage of the tool, the action space including different strategies for maintenance and replacement, and the reward function for evaluating the effect of each strategy;

[0157] S73. Use deep Q network to optimize dynamic maintenance strategy. Deep Q network optimizes maintenance and replacement strategy by combining neural network and Q learning algorithm:

[0158] Q(s,a)←Q(s,a)+α[r+γmax a' Q(s′,a′)-Q(s,a)];

[0159] Where Q(s,a) is the Q value of taking action a in state s, α is the learning rate, r is the immediate reward, γ is the discount factor, s' is the new state after executing the action, and a' is the best action in the new state;

[0160] S74. Through multiple rounds of training and simulation experiments, the maintenance and replacement strategies are optimized so that the maximum cumulative reward can be obtained at different degradation stages, and real-time monitoring and preventive maintenance of the tool can be achieved;

[0161] S75. According to the optimized dynamic maintenance strategy, formulate a specific maintenance and replacement plan, including specific time points and operation steps, and apply it in the actual production process.

[0162] Embodiment 1:

[0163] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a precision machinery processing plant. In order to improve production efficiency and product quality, the plant decided to introduce the tool degradation trend analysis method of the present invention to achieve real-time monitoring and preventive maintenance of the tool status during the processing process. The plant mainly produces high-precision parts and has extremely high requirements for the status of the tool. However, since the existing tool monitoring method cannot fully and accurately reflect the comprehensive status of the tool, the problem of early or late replacement of the tool often occurs in production, which not only increases production costs, but also affects product quality.

[0164] In order to implement the method of the present invention, firstly, a vibration sensor, an acoustic emission sensor and a temperature sensor are installed on the processing equipment. These sensors can collect the status signals of the tool in the processing process in real time and transmit the signals to the central control system through the data acquisition system. In the data acquisition process, advanced noise removal and filtering technology are used to standardize the original signal to ensure the accuracy and availability of the signal.

[0165] In the data processing stage, the preprocessed signal is decomposed using an improved sparse adaptive decomposition algorithm to extract high-order statistical features, sparse coding features, and deep learning features. By dynamically adjusting the sparse basis dictionary and adaptive sparse regularization parameters, the sparse representation coefficient vector of each signal segment is obtained, and high-order statistical features such as mean, variance, skewness, and kurtosis are calculated. Next, the extracted features are reduced in dimension using a variational autoencoder combined with adaptive multimodal fusion dimensionality reduction technology. The variational autoencoder model is used to compress and represent features, and adaptive multimodal fusion dimensionality reduction technology is used to fuse different modal features to extract representative main feature components.

[0166] After feature extraction and dimensionality reduction, the adversarial similarity metric network is used to calculate the similarity between the tool state signal and the reference signal at different times. The generative adversarial network is used to generate virtual signals similar to the reference signal, and the similarity metric between the signals is calculated by the discriminator. According to the similarity metric results, the tool degradation trend diagram is drawn using an adaptive dynamic trend analysis method based on the Bayesian non-parametric model. The parameters of the trend analysis model are dynamically adjusted to capture the details of the changes in the tool degradation process and show the trend of signal similarity over time.

[0167] The degradation stage of the tool is identified through the trend chart, and the degradation process is divided into the initial running-in stage, stable use stage and degradation stage using the multi-level degradation stage identification algorithm. According to the characteristics of different stages, the tool maintenance and replacement suggestions are put forward. The dynamic maintenance strategy optimization algorithm is used to optimize the initially formulated maintenance and replacement strategies, and a dynamic maintenance strategy optimization model based on reinforcement learning is constructed. The dynamic maintenance strategy is optimized through the deep Q network to ensure that the maintenance and replacement strategies can obtain the maximum cumulative reward in different degradation stages, so as to realize the real-time monitoring and preventive maintenance of the tool.

[0168] In order to verify the effectiveness of the present invention, the comparative data of the method of the present invention and the traditional method were statistically analyzed during a six-month production process. After using the method of the present invention, the average service life of the tool was extended from the original 150 hours to 210 hours, and the production cost was reduced by 15%. The production efficiency was increased by 10%, and the product qualification rate was increased from 96% to 99%. The specific data are as follows:

[0169] Production location: a precision machinery processing plant. Implementation time: January 2023-June 2023. Data using the traditional method: average tool life: 150 hours, monthly production cost: 2 million yuan, production efficiency: 85%, product qualification rate: 96%. Data using the method of the present invention: average tool life: 210 hours, monthly production cost: 1.7 million yuan, production efficiency: 95%, product qualification rate: 99%.

[0170] Table 1 Comparative data of the traditional method and the method of the present invention in various aspects of the tool

[0171]

[0172] It can be seen from the above data that in the actual production process of a precision machinery processing plant, the tool degradation trend analysis method based on similarity measurement was applied. After six months of implementation, the data of the traditional method and the method of the present invention were compared, showing significant beneficial effects. By comparing the average service life of the tool, after using the method of the present invention, the average service life of the tool was extended from 150 hours of the traditional method to 210 hours, an increase of 40%. This shows that the method of the present invention has significant advantages in extending the tool life. In terms of production cost, after using the method of the present invention, the monthly production cost was reduced from 2 million yuan of the traditional method to 1.7 million yuan, a decrease of 15%. This reflects that the method of the present invention can effectively reduce production costs and improve economic benefits. In terms of production efficiency, after using the method of the present invention, the production efficiency was increased from 85% of the traditional method to 95%, an increase of 10 percentage points. This means that the method of the present invention has a significant effect in improving production efficiency. In terms of product qualification rate, after using the method of the present invention, the product qualification rate was increased from 96% of the traditional method to 99%, an increase of 3 percentage points. This shows that the method of the present invention has significant advantages in improving product quality.

[0173] The method of the present invention has significant advantages in tool status monitoring and degradation trend analysis. Through real-time monitoring and preventive maintenance, the service life of the tool is extended, the production cost is reduced, the production efficiency and product quality are improved, and the beneficial effects of the present invention are fully demonstrated. In summary, the tool degradation trend analysis method based on similarity measurement of the present invention has broad application prospects and significant technical advantages in actual production. It not only solves many deficiencies in the prior art, but also has achieved significant technical breakthroughs in monitoring accuracy, feature extraction, similarity measurement, trend analysis and maintenance strategy optimization, etc., and has broad application prospects and practical value.

[0174] In summary, the tool degradation trend analysis method based on similarity measurement of the present invention has broad application prospects and significant technical advantages in actual production. It not only solves many deficiencies in the prior art, but also has made significant technical breakthroughs in monitoring accuracy, feature extraction, similarity measurement, trend analysis and maintenance strategy optimization. It has broad application prospects and practical value.

[0175] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. Tool degradation trend analysis method based on similarity measurement, It is characterized in that The steps include: S1, collecting the state signal of the tool in the machining process in real time through the vibration sensor, acoustic emission sensor and temperature sensor, and performing noise removal, filtering and standardization processing on the collected state signal; S2, using the improved sparse adaptive decomposition algorithm to extract high-order statistical features, sparse coding features and deep learning features from the preprocessed signal; S3, using variational autoencoder combined with adaptive multimodal fusion dimensionality reduction technology to reduce the dimensionality of the extracted features, compressing and representing the features through the variational autoencoder model, and using adaptive multimodal fusion dimensionality reduction technology to fuse different modal features and extract representative main feature components; S4, calculating the similarity between the tool state signal and the reference signal at different times through an adversarial similarity measurement network, wherein the adversarial similarity measurement network uses a generative adversarial network to perform adversarial training on the signal to generate a virtual signal similar to the reference signal, and calculating the similarity measurement value between the signals through a discriminator; S5. According to the similarity measurement results, an adaptive dynamic trend analysis based on the Bayesian non-parametric model is used to draw a tool degradation trend diagram, dynamically adjust the parameters of the trend analysis model to capture the change details in the tool degradation process, and show the trend of signal similarity changing over time; S6. Identify the degradation stage of the tool through the trend chart, and use the multi-level degradation stage identification algorithm to divide the degradation process into multiple stages, including the initial running-in stage, stable use stage and degradation stage, and classify them according to the characteristics of different stages; S7. Propose tool maintenance and replacement suggestions based on the degradation trend graph, and use dynamic maintenance strategy optimization algorithm to optimize the maintenance and replacement strategy to achieve real-time monitoring and preventive maintenance of tools.

2. The tool degradation trend analysis method based on similarity measurement according to claim 1, It is characterized in that The S2 specifically includes: S21, performing segment processing on the pre-processed state signal, dividing the signal into several segments, each of which has a length of L; S22. Decomposing each signal segment using an improved sparse adaptive decomposition algorithm, wherein the improved sparse adaptive decomposition algorithm includes an optimization process of dynamically adjusting a sparse base dictionary and an adaptive sparse regularization parameter: min∥X-Dα∥ 2 2 +λ(t)∥α∥ 1 ; α Where X is the preprocessed signal matrix, D is the dynamically adjusted sparse basis dictionary matrix, α is the sparse coefficient vector, and λ(t) is the sparse regularization parameter that changes dynamically over time; S23, by solving the above objective function, a sparse representation coefficient vector α of each signal segment is obtained, and sparse coding features are extracted; S24. Calculate the high-order statistical features of each signal segment, including mean μ, variance σ 2 , skewness γ and kurtosis κ: Among them, x i is the signal sample point, N is the number of samples; S25. Use graph neural network to further process sparse coding features and high-order statistical features, construct a graph convolutional neural network, input sparse coding features and high-order statistical features, use the correlation between nodes to extract signal features, and output enhanced features.

3. The tool degradation trend analysis method based on similarity measurement according to claim 1, It is characterized in that The S3 specifically includes: S31, preliminarily integrating the sparse coding features, high-order statistical features and enhanced features extracted by the graph neural network to form an input feature vector φ; S32, using a variational autoencoder to reduce the dimension of the input feature vector φ, the variational autoencoder includes an encoder and a decoder, the encoder compresses the input feature vector φ to a latent space representation z, and the decoder reconstructs the feature vector from the latent space representation z L=E q(z|φ) [logp(φ|z)]-KL[q(z|φ)||p(z)]; Where q(z|φ) is the probability distribution of the encoder, p(z) is the prior distribution, and KL is the Kullback-Leibler divergence; S33, weighted fusion of different modal features is performed through adaptive multimodal fusion dimensionality reduction technology, the weight of each modal feature is adaptively adjusted, and representative main feature components are extracted: Among them, φ i is the eigenvector of the ith mode, w i is the weight of the i-th mode, and M is the number of modes; S34, combine the latent space representation z output by the variational autoencoder and the main feature components extracted by the adaptive multimodal fusion dimensionality reduction technology to generate the final reduced-dimensional feature vector φ final : Among them, f j (z,φ i ) is the feature function extracted by the variational autoencoder and adaptive multimodal fusion dimensionality reduction technology, β j is the dynamic weight in the fusion process, and N is the number of feature functions.

4. The tool degradation trend analysis method based on similarity measurement according to claim 1, It is characterized in that The S4 specifically includes: S41, constructing an adversarial similarity measurement network, the network comprising a generator G and a discriminator D, the generator G generates a virtual signal similar to a reference signal, and the discriminator D is used to distinguish between a real signal and a virtual signal; S42, generator G inputs feature vector φ final Generate virtual feature vector φ gen : Among them, G(φ final )=φ gen , D(φ gen ) is the discriminant result of the discriminator on the generated feature vector; S43, the discriminator D inputs the true feature vector φ final and the virtual eigenvector φ gen To determine: Among them, D(φ final ) is the discriminator’s discriminant result on the real feature vector, D(φ gen ) is the discriminant result of the virtual feature vector; S44, through adversarial training, the virtual feature vector φ generated by the generator G gen and the true eigenvector φ final The similarity measure between them is maximized: Among them, · represents the vector inner product, ∥·∥ represents the Euclidean norm of the vector; S45. Calculate the true feature vector φ according to the output result of the discriminator D final and the virtual eigenvector φ gen The similarity measurement value between them is taken as the final similarity measurement result.

5. The tool degradation trend analysis method based on similarity measurement according to claim 1, It is characterized in that The S5 specifically includes: S51. Based on the similarity measurement results calculated by the adversarial similarity measurement network, an adaptive dynamic trend analysis method based on the Bayesian non-parametric model is used to model and analyze the tool degradation process; S52. Constructing a Bayesian non-parametric model, the model includes a Gaussian process and a dynamic Bayesian network, the Gaussian process is used to capture the continuous change characteristics in the tool degradation process, and the dynamic Bayesian network is used to capture the discrete stage change characteristics in the tool degradation process; S53. Use Gaussian process to predict tool degradation trend, establish the relationship between similarity measurement result vector y and time series, and capture the correlation between signal features through the covariance matrix K of Gaussian process to achieve smooth prediction of tool degradation trend: Among them, y is the similarity measurement result vector, K is the covariance matrix, and n is the number of samples; S54. Dynamic Bayesian network is used to identify and classify different stages in the tool degradation process. The dynamic Bayesian network models the states at different time points through the probability transfer matrix. Through the dynamic Bayesian network, the key stages in the tool degradation process are accurately identified and the degradation characteristics of different stages are classified: Among them, X t and X t+1 denote the states at time t and time t+1 respectively, P(X t |X t+1 ) is the transition probability between time t and time t+1; S55, dynamically adjust the tool degradation trend through the Bayesian non-parametric model, optimize the model parameters to adaptively capture the details of the changes in the tool degradation process, and dynamically adjust the model parameters to adapt to the changes in the tool degradation process through the Bayesian update process; S56. Based on the dynamic adjustment results of the Bayesian non-parametric model, a tool degradation trend graph is drawn to show the trend of signal similarity over time. Combined with the prediction results of the Gaussian process and dynamic Bayesian network, real-time analysis and prediction of tool degradation are provided.

6. The tool degradation trend analysis method based on similarity measurement according to claim 1, It is characterized in that The S6 specifically includes: S61. Preliminarily identify the degradation stage of the tool according to the tool degradation trend diagram drawn, and classify and mark each stage in the tool degradation process using a multi-level degradation stage identification algorithm; S62, calculating the signal similarity metric value at each time point, and determining the initial degradation stage boundary according to the change trend of the similarity metric value; S63. Evaluate the signal complexity of each initial degradation stage through multi-scale entropy analysis, and calculate the multi-scale entropy value of each stage: Among them, p i is the probability distribution of the signal at the i-th scale, and n is the number of scales; S64. Based on the results of multi-scale entropy analysis, refine the degradation stage boundaries and identify the degradation stages, including the initial running-in stage, stable use stage and degradation stage; S65. According to the results of the multi-level degradation stage identification algorithm, the tool degradation process is divided into several specific stages, and the mixed effect model is used to classify and model the characteristics of different stages: Y ij =b 0 +b 1 X ij +b i +∈ ij ; Among them, Y ij is the observed value of the i-th stage at the j-th time point, β 0 and β 1 is the fixed effect parameter, X ij is the predictor variable of the observed value, b i is a random effect, ∈ ij is the error term; S66. Based on the classification results of the mixed effect model, the characteristics of different degradation stages are modeled and analyzed, the key characteristic parameters of each stage are extracted, and the characteristic curves of each stage are drawn; S67. Combining the classification results of the mixed effect model and the characteristic parameters of each stage, a complete tool degradation stage identification model is formed.

7. The tool degradation trend analysis method based on similarity measurement according to claim 1, It is characterized in that The S7 specifically includes: S71. Based on the degradation trend diagram and tool degradation stage identification model, make suggestions for tool maintenance and replacement, and preliminarily determine the maintenance and replacement strategy, including maintenance time point, maintenance content and replacement standards; S72. Use a dynamic maintenance strategy optimization algorithm to optimize the maintenance and replacement strategies initially formulated, and construct a dynamic maintenance strategy optimization model based on reinforcement learning, the model including a state space, an action space, and a reward function, the state space including the current state and degradation stage of the tool, the action space including different strategies for maintenance and replacement, and the reward function for evaluating the effect of each strategy; S73. Use deep Q network to optimize dynamic maintenance strategy. Deep Q network optimizes maintenance and replacement strategy by combining neural network and Q learning algorithm: Q(s,a)←Q(s,a)+α[r+γmax a' Q(s′,a′)-Q(s,a)]; Where Q(s,a) is the Q value of taking action a in state s, α is the learning rate, r is the immediate reward, γ is the discount factor, s' is the new state after executing the action, and a' is the best action in the new state; S74. Through multiple rounds of training and simulation experiments, the maintenance and replacement strategies are optimized so that the maximum cumulative reward can be obtained at different degradation stages, and real-time monitoring and preventive maintenance of the tool can be achieved; S75. According to the optimized dynamic maintenance strategy, formulate a specific maintenance and replacement plan, including specific time points and operation steps, and apply it in the actual production process.