Cutter clamping state evaluation method based on self-encoding reconstruction characteristics

Through multimodal sensor and autoencoder technology, combined with multi-scale decomposition and fuzzy logic optimization, the limitations of single signal monitoring in the existing technology are solved, and high-precision and efficient monitoring of tool clamping status are achieved, adapting to complex machining environments , improves production efficiency and processing quality.

CN119046875AActive Publication Date: 2024-11-29YOUJI TECH (SHANGHAI) CO LTD

Patent Information

Application Number
CN202411134078.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-29
Estimated Expiration
2044-08-19

AI Technical Summary

Technical Problem

The existing tool clamping status monitoring system relies on a single signal type and cannot fully reflect the clamping status changes in complex processing environments. It lacks dynamic adaptability and comprehensive analysis capabilities, resulting in inaccurate monitoring results and low efficiency.

Method used

Multimodal sensors are used to collect vibration, temperature, acoustic emission and strain signals, and deep feature extraction and noise filtering are performed through the autoencoder. Combined with multi-scale decomposition and fuzzy logic optimization strategies, real-time clamping status evaluation and automatic adjustment are achieved.

Benefits of technology

It significantly improves the clamping status monitoring accuracy and system robustness, reduces noise interference, adapts to complex processing environments, reduces manual intervention and economic losses, and improves production efficiency and processing quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046875A_ABST
    Figure CN119046875A_ABST
Patent Text Reader

Abstract

The invention discloses a self-encoding reconstruction feature-based tool clamping state evaluation method, which comprises the following steps of: firstly, acquiring vibration, temperature, acoustic emission and strain signals in real time by using a multi-mode sensor, and performing preprocessing and feature fusion on the signals to generate a unified feature vector; and then, inputting the feature vector into a dynamic sparsity adaptive auto-encoder, performing depth feature extraction, and dynamically adjusting a sparsity parameter. And then, comparing the reconstructed feature vector with the original feature vector to construct a residual signal set, and identifying an abnormal signal through a multi-scale decomposition and dynamic reconstruction technology. And finally, in combination with a biologically inspired adaptive signal processing technology and a mixed attention mechanism, analyzing a signal time sequence, generating a clamping state evaluation result, and optimizing clamping operation through a multi-objective optimization strategy assisted by fuzzy logic. According to the method, the precision loss caused by improper clamping in the machining process is effectively reduced, and the accuracy of tool clamping state evaluation and the production efficiency are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of tool clamping fault diagnosis, and in particular to a tool clamping state evaluation method based on self-encoding reconstruction features. Background Art

[0002] In modern manufacturing, the clamping state of the tool directly affects the stability of the machining process and the final quality of the workpiece. Reasonable clamping can not only ensure the machining accuracy of the workpiece, but also avoid machining failures and increased scrap rates. However, in the actual production process, there are still many technical challenges in monitoring and evaluating the tool clamping state, which makes it difficult to meet the increasingly complex machining environment and high-precision requirements.

[0003] Currently, in most processing scenarios, the monitoring of the clamping status still mainly relies on the operator's experience, and manual inspection and adjustment are performed to ensure the correctness of the clamping. Although this traditional method can guarantee the clamping quality of the tool to a certain extent, it also has significant defects. First of all, manual monitoring is highly dependent on the worker's experience and judgment ability, is highly subjective, and is easily affected by the operator's level and fatigue. In addition, manual operation is not only time-consuming and labor-intensive, but also prone to improper clamping due to negligence in high-paced production lines. With the complexity of processing tasks, manual detection methods are difficult to cope with the changing signal characteristics under different working conditions, especially when real-time and high-precision monitoring is required, the limitations of manual methods are more obvious.

[0004] In order to improve production efficiency, some manufacturing companies have introduced automated monitoring systems, which usually rely on a single type of sensor to monitor the clamping status of the tool. However, the existing automated monitoring systems still have the following prominent problems in practical applications:

[0005] 1. Limitations of a single signal source: Current monitoring systems rely on a single signal type to determine the clamping status of the tool. This method tends to ignore other important influencing factors, such as temperature, acoustic emission, and strain signals. Since the monitoring method with a single signal source cannot fully reflect the changes in the clamping status, it may lead to misjudgment or missed judgment in a complex processing environment, thus affecting the processing quality.

[0006] 2. Insufficient signal processing capabilities: The signals generated by the tool during the machining process often contain a lot of noise and irrelevant information. Existing technologies usually use simple filtering or threshold judgment methods to process these complex signals. However, these methods show obvious deficiencies in the face of signal volatility and uncertainty, and are easily interfered by noise, resulting in inaccurate monitoring results.

[0007] 3. Lack of dynamic adaptability: Existing automated monitoring systems usually rely on preset fixed rules or models, which lack dynamic adjustment capabilities and cannot adapt to real-time changes in the processing environment. When processing conditions change, existing systems are often unable to adjust monitoring strategies in a timely manner, resulting in delayed monitoring results and affecting processing efficiency.

[0008] 4. Lack of comprehensive analysis capabilities: Most existing systems only analyze a single type of signal and lack the ability to comprehensively process multimodal signals. This limitation prevents the system from utilizing the complementary information of multiple signal sources and from comprehensively and accurately evaluating the clamping status. Although multimodal sensor technology has been developed, existing monitoring systems have not yet fully utilized the comprehensive analysis capabilities of multi-source signals to improve monitoring accuracy.

[0009] Therefore, how to provide a tool clamping state assessment method based on self-encoding reconstruction features is an urgent problem that technicians in this field need to solve. Summary of the invention

[0010] One purpose of the present invention is to propose a tool clamping state evaluation method based on autoencoder reconstruction features. The present invention uses an autoencoder algorithm to train a model of a single type of normal data, learn effective feature representations, discover potential patterns and structures in the data, improve the generalization ability of the model, and effectively filter noise and redundant information in the input data to reduce interference with the evaluation results. The present invention also solves the problem of tool clamping accuracy in machine tool processing by indirectly monitoring the sensor information during the processing, realizes the transition from manual confirmation to intelligent adaptation, greatly reduces labor costs, and significantly reduces processing time.

[0011] A tool clamping state assessment method based on self-encoding reconstruction features according to an embodiment of the present invention comprises the following steps:

[0012] S1. Use multi-modal sensors installed near the tool to collect multiple modal signals in real time, including vibration, temperature, acoustic emission, and strain signals;

[0013] S2, preprocessing the collected multimodal signals, and fusing the multimodal signal features to generate a unified feature vector;

[0014] S3, input the generated feature vector into the dynamic sparsity adaptive autoencoder, perform deep feature extraction on the fusion signal through the autoencoder, and dynamically adjust the sparsity parameters to adapt to signal changes under different working conditions;

[0015] S4, compare the reconstructed feature vector output by the autoencoder with the original feature vector, construct a residual signal set, perform multi-scale decomposition and dynamic reconstruction on the residual signal based on the adaptive multi-scale decomposition and dynamic reconstruction residual detection mechanism, and identify instantaneous anomalies and long-term accumulated abnormal signals;

[0016] S5. Adaptive signal processing technology based on bioinspiration, which adaptively adjusts the signal processing method according to the changes of real-time signals to cope with different signal characteristics and noise levels;

[0017] S6. Perform real-time analysis on the preprocessed signal, combine it with the time series deep learning model based on the hybrid attention mechanism, analyze the time series pattern and characteristics of the signal, and generate the clamping state evaluation result;

[0018] S7. Based on historical data and current evaluation results, a multi-objective optimization strategy assisted by fuzzy logic is used to perform fuzzy reasoning and trade-offs between multiple optimization objectives, generate an optimization strategy for the clamping state, and automatically adjust relevant parameters to optimize the clamping operation;

[0019] S8. Through the interaction between the operator and the system, the model parameters are adjusted in real time, and the parameter adjustment behavior is recorded and analyzed. The system automatically optimizes the parameter settings through the self-learning mechanism to adapt to different processing tasks and environments.

[0020] Optionally, the S3 specifically includes:

[0021] S31, input the generated feature vector into the dynamic sparsity adaptive autoencoder, and the feature vector is represented as X = [x1, x2, ..., x n ], where n is the feature dimension;

[0022] S32, the adaptive self-encoder includes an encoder and a decoder, and the encoder compresses the input feature vector into a latent space representation:

[0023]

[0024] Among them, Z is the latent space representation, f enc is the encoder function, θ enc is the encoder parameter, W enc is the encoder weight matrix, b enc is the bias vector, α,∈ and δ are small constants, ∥W enc X∥ 2 is the square of the two-norm of the vector after the encoder linear transformation, is the natural logarithm of the square of the eigenvalue;

[0025] S33. Reconstruct the input feature vector from the latent space representation through the decoder:

[0026]

[0027] in, is the reconstructed eigenvector, f dec is the decoder function, θ dec is the decoder parameter, W dec is the decoder weight matrix, b dec is the bias vector, σ is the activation function, λ, μ and ν are adjustment parameters, z i is the i-th component in the latent space representation, sin(W dec Z) is the sine term after linear transformation of the latent space representation, is the cosine term represented in the latent space;

[0028] S34, calculating the reconstruction error between the original feature vector and the reconstructed feature vector:

[0029]

[0030] in, is the reconstruction error, x i is the i-th component in the original eigenvector, is the i-th component in the reconstructed feature vector, β is the adjustment parameter, is the natural logarithm of the square of the error, is the fourth power of the error;

[0031] S35. Use a dynamic sparsity adjustment mechanism to adjust the sparsity parameters of the autoencoder based on the reconstruction error:

[0032]

[0033] Among them, θ spar is the sparsity parameter, η is the learning rate, γ and δ are regularization coefficients, z i is the i-th eigenvalue in the latent space, is the partial derivative of the reconstruction error with respect to the sparsity parameter, k is the specific number of sparsity parameters, log e (1+exp(-||z i ||)) is the logarithmic term of the latent space representation, is the square term of the latent space representation;

[0034] S36. Through multiple iterative training, the sparsity parameters are adaptively adjusted to optimize the reconstruction capability of the autoencoder and adapt to signal changes under different working conditions.

[0035] Optionally, the S4 specifically includes:

[0036] S41, comparing the reconstructed feature vector output by the autoencoder with the original feature vector, and constructing a residual signal set, each residual signal representing the difference between the original feature value and the reconstructed feature value;

[0037] S42, performing adaptive multi-scale decomposition on the constructed residual signal set to generate components of different scales, each component representing a change of the residual signal at a different time scale;

[0038] S43, dynamically reconstructing each component after multi-scale decomposition, and adjusting the reconstructed signal according to the components of multi-scale decomposition to adapt to changes in the signal;

[0039] S44, calculating the instantaneous abnormal value and the long-term cumulative abnormal value of each reconstructed signal, wherein the instantaneous abnormal value indicates the abnormal degree of the signal at the current moment, and the long-term cumulative abnormal value indicates the abnormal degree accumulated over a period of time;

[0040] S45, identifying instantaneous abnormalities and long-term accumulated abnormal signals, by setting dynamic thresholds, dynamically adjusting the thresholds according to changes in real-time signals, and detecting signal abnormalities;

[0041] S46. Based on the set dynamic threshold, detect instantaneous anomalies and long-term accumulated abnormal signals, and generate anomaly detection results, each of which indicates whether a corresponding feature is abnormal;

[0042] S47. Output the anomaly detection results and update the model parameters in real time. According to the new signal data and detection results, the adaptive multi-scale decomposition and dynamic reconstruction residual detection mechanism is continuously optimized to adapt to different working conditions and signal feature changes.

[0043] Optionally, the S5 specifically includes:

[0044] S51, using a bio-inspired adaptive filter to process the real-time signal, the output signal of the filter is represented by S f (t), the filter processes the signal based on past signal samples and current weight coefficients, and adapts to signal changes by dynamically adjusting the weight coefficients;

[0045] S52, adaptively adjusting the weight coefficient of the filter according to the change of the real-time signal, wherein the adjustment of the weight coefficient is based on the filtering error e(t) and the learning rate η at the current moment;

[0046] S53, calculating the characteristic value of the signal after filtering, the characteristic value includes the amplitude, frequency and phase of the signal:

[0047]

[0048] Among them, i∈{1,2,…,n} represents the different dimensions of the signal, α k , βk , γ and δ k is the adjustment parameter, N is the number of signal samples, S f (t) is the filtered signal, s f,i (t) is the i-th signal after filtering, s i (t) is the i-th signal in the real-time signal data, f i (t) is the eigenvalue of the i-th signal after filtering, is the potential feature representation reconstructed by the autoencoder, and m is the dimension of the potential feature;

[0049] S54, dynamically adjusting the characteristic value of the filtered signal, and adaptively adjusting the signal processing method according to the change of the real-time signal;

[0050] S55, calculating the rate of change of the signal characteristic value, where the rate of change is used to evaluate the change of the signal characteristic value at different times and capture the dynamic characteristics of the signal;

[0051] S56. Generate an adaptive strategy for signal processing based on the rate of change of the signal characteristic value, and automatically optimize parameter settings through a self-learning mechanism to adapt to different processing tasks and environments.

[0052] Optionally, the S6 specifically includes:

[0053] S61, the preprocessed signal S'(t) = [s'1(t), s'2(t), ..., s' n (t)] for real-time analysis, where s' i (t) is the i-th signal dimension;

[0054] S62. Input the signal into the temporal deep learning model based on the hybrid attention mechanism. First, the signal features are weighted through the self-attention mechanism to capture the key temporal patterns and long-term dependencies:

[0055]

[0056] Among them, H e (t) is the hidden state of the encoder after self-attention processing, W Q,j and W K,j are the weights of the query and key matrices respectively, is the latent feature representation generated by the autoencoder, α j and β j is the weighting coefficient, is the result of convolution processing, ∈ is a constant term to prevent the denominator from being zero;

[0057] S63, the decoder part combines the output of the self-attention mechanism and the reconstructed features generated by the autoencoder Further analysis of the short-term characteristics of the signal generates the clamping state evaluation result R(t):

[0058]

[0059] in, is the reconstructed signal feature output by the decoder, γ j and δ j To adjust the parameters;

[0060] S64, adjusting the parameters of the model in real time based on the generated clamping state evaluation result;

[0061] S65. The model is regularly updated based on new signal data and evaluation results, and the model parameters are adaptively adjusted through continuous optimization.

[0062] Optionally, the S7 specifically includes:

[0063] S71. Based on historical data and current evaluation results, a multi-objective optimization function is constructed. The optimization function combines the latent feature representation reconstructed by the autoencoder. And the filtered signal eigenvalue S f (t), the optimization objectives include the accuracy, robustness and real-time performance of the clamping state, and these objectives are weighted and combined through weight coefficients to generate a comprehensive optimization objective function;

[0064] S72, using fuzzy logic to assist in multi-objective optimization strategy, reasoning and weighing the optimization objectives, the fuzzy logic reasoning process includes accurately fuzzifying the optimization objectives, and generating the optimization strategy according to the preset fuzzy reasoning rules, the reasoning rules are based on historical data and current evaluation results, and are adjusted for specific key factors in the clamping state to generate a specific optimization strategy;

[0065] S73. Automatically adjust relevant parameters of the model according to the generated optimization strategy:

[0066]

[0067] Among them, θ new is the updated parameter, θ old is the current parameter, η is the learning rate, To optimize the gradient of the objective function, γ j and δ j To adjust the parameters, used to precisely control the weight adjustment;

[0068] S74, the iterative optimization process stops when the rate of change of the optimization objective function reaches a convergence condition;

[0069] S75. Apply the optimized parameters to the real-time clamping operation and continuously adjust the optimization strategy based on new signal data and evaluation results.

[0070] The beneficial effects of the present invention are:

[0071] (1) The present invention combines the vibration, temperature, acoustic emission and strain signals collected by multimodal sensors, and uses the unsupervised learning ability of the autoencoder to perform deep feature extraction on multimodal signals to achieve data dimensionality reduction and denoising. Compared with the traditional single signal monitoring method, the present invention can comprehensively process multimodal data, reduce the interference of noise and redundant information on the evaluation results, and significantly improve the monitoring accuracy of the tool clamping state and the robustness of the system.

[0072] (2) The present invention can accurately construct residual signals through the self-encoding reconstruction algorithm, and use these signals to indirectly detect the tool clamping status. Through adaptive multi-scale decomposition and dynamic reconstruction technology, the system can identify instantaneous anomalies and long-term accumulated abnormal signals in real time. Compared with traditional static signal processing methods, the dynamic adjustment capability of the present invention enables it to adapt to real-time changes in processing conditions, capture abnormalities in the clamping state in a timely manner, reduce the problem of large-scale scrapping of workpieces due to improper clamping, and reduce economic losses.

[0073] (3) The present invention realizes the automatic adjustment and optimization of parameters in the clamping state evaluation process by introducing a multi-objective optimization strategy assisted by fuzzy logic, reduces the reliance on manual intervention, and significantly improves production efficiency. At the same time, the intelligent parameter optimization of the system further improves the processing quality, reduces production costs, and greatly improves the overall economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] 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:

[0075] Figure 1 A flowchart of a tool clamping state evaluation method based on self-encoding reconstruction features proposed by the present invention;

[0076] Figure 2 A schematic diagram of the autoencoder proposed in the present invention performing deep feature extraction on a multimodal signal and generating a residual signal;

[0077] Figure 3 Schematic diagram of the application of the fuzzy logic-assisted multi-objective optimization strategy proposed in the present invention in the clamping state assessment. DETAILED DESCRIPTION

[0078] 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.

[0079] refer to Figure 1-3 , a tool clamping state evaluation method based on self-encoding reconstruction features, comprising the following steps:

[0080] S1. Use multi-modal sensors installed near the tool to collect multiple modal signals in real time, including vibration, temperature, acoustic emission, and strain signals;

[0081] S2, preprocessing the collected multimodal signals, and fusing the multimodal signal features to generate a unified feature vector;

[0082] S3, input the generated feature vector into the dynamic sparsity adaptive autoencoder, perform deep feature extraction on the fusion signal through the autoencoder, and dynamically adjust the sparsity parameters to adapt to signal changes under different working conditions;

[0083] S4, compare the reconstructed feature vector output by the autoencoder with the original feature vector, construct a residual signal set, perform multi-scale decomposition and dynamic reconstruction on the residual signal based on the adaptive multi-scale decomposition and dynamic reconstruction residual detection mechanism, and identify instantaneous anomalies and long-term accumulated abnormal signals;

[0084] S5. Adaptive signal processing technology based on bioinspiration, which adaptively adjusts the signal processing method according to the changes of real-time signals to cope with different signal characteristics and noise levels;

[0085] S6. Perform real-time analysis on the preprocessed signal, combine it with the time series deep learning model based on the hybrid attention mechanism, analyze the time series pattern and characteristics of the signal, and generate the clamping state evaluation result;

[0086] S7. Based on historical data and current evaluation results, a multi-objective optimization strategy assisted by fuzzy logic is used to perform fuzzy reasoning and trade-offs between multiple optimization objectives, generate an optimization strategy for the clamping state, and automatically adjust relevant parameters to optimize the clamping operation;

[0087] S8. Through the interaction between the operator and the system, the model parameters are adjusted in real time, and the parameter adjustment behavior is recorded and analyzed. The system automatically optimizes the parameter settings through the self-learning mechanism to adapt to different processing tasks and environments.

[0088] In this implementation, S3 specifically includes:

[0089] S31, input the generated feature vector into the dynamic sparsity adaptive autoencoder, and the feature vector is represented as X = [x1, x2, ..., x n], where n is the feature dimension;

[0090] S32, the adaptive self-encoder includes an encoder and a decoder, and the encoder compresses the input feature vector into a latent space representation:

[0091]

[0092] Among them, Z is the latent space representation, f enc is the encoder function, θ enc is the encoder parameter, W enc is the encoder weight matrix, b enc is the bias vector, α,∈ and δ are small constants, ∥W enc X∥ 2 is the square of the two-norm of the vector after the encoder linear transformation, is the natural logarithm of the square of the eigenvalue;

[0093] S33. Reconstruct the input feature vector from the latent space representation through the decoder:

[0094]

[0095] in, is the reconstructed eigenvector, f dec is the decoder function, θ dec is the decoder parameter, W dec is the decoder weight matrix, b dec is the bias vector, σ is the activation function, λ, μ and ν are adjustment parameters, z i is the i-th component in the latent space representation, sin(W dec Z) is the sine term after linear transformation of the latent space representation, is the cosine term represented in the latent space;

[0096] S34, calculating the reconstruction error between the original feature vector and the reconstructed feature vector:

[0097]

[0098] in, is the reconstruction error, x i is the i-th component in the original eigenvector, is the i-th component in the reconstructed feature vector, β is the adjustment parameter, is the natural logarithm of the square of the error, is the fourth power of the error;

[0099] S35. Use a dynamic sparsity adjustment mechanism to adjust the sparsity parameters of the autoencoder based on the reconstruction error:

[0100]

[0101] Among them, θ spar is the sparsity parameter, η is the learning rate, γ and δ are regularization coefficients, z i is the i-th eigenvalue in the latent space, is the partial derivative of the reconstruction error with respect to the sparsity parameter, k is the specific number of sparsity parameters, log e (1+exp(-||z i ||)) is the logarithmic term of the latent space representation, is the square term of the latent space representation;

[0102] S36. Through multiple iterative training, the sparsity parameters are adaptively adjusted to optimize the reconstruction capability of the autoencoder and adapt to signal changes under different working conditions.

[0103] In this implementation manner, the S4 specifically includes:

[0104] S41, comparing the reconstructed feature vector output by the autoencoder with the original feature vector, and constructing a residual signal set, each residual signal representing the difference between the original feature value and the reconstructed feature value;

[0105] S42, performing adaptive multi-scale decomposition on the constructed residual signal set to generate components of different scales, each component representing a change of the residual signal at a different time scale;

[0106] S43, dynamically reconstructing each component after multi-scale decomposition, and adjusting the reconstructed signal according to the components of multi-scale decomposition to adapt to changes in the signal;

[0107] S44, calculating the instantaneous abnormal value and the long-term cumulative abnormal value of each reconstructed signal, wherein the instantaneous abnormal value indicates the abnormal degree of the signal at the current moment, and the long-term cumulative abnormal value indicates the abnormal degree accumulated over a period of time;

[0108] S45, identifying instantaneous abnormalities and long-term accumulated abnormal signals, by setting dynamic thresholds, dynamically adjusting the thresholds according to changes in real-time signals, and detecting signal abnormalities;

[0109] S46. Based on the set dynamic threshold, detect instantaneous anomalies and long-term accumulated abnormal signals, and generate anomaly detection results, each of which indicates whether a corresponding feature is abnormal;

[0110] S47. Output the anomaly detection results and update the model parameters in real time. According to the new signal data and detection results, the adaptive multi-scale decomposition and dynamic reconstruction residual detection mechanism is continuously optimized to adapt to different working conditions and signal feature changes.

[0111] In this implementation manner, S5 specifically includes:

[0112] S51, using a bio-inspired adaptive filter to process the real-time signal, the output signal of the filter is represented by S f (t), the filter processes the signal based on past signal samples and current weight coefficients, and adapts to signal changes by dynamically adjusting the weight coefficients;

[0113] S52, adaptively adjusting the weight coefficient of the filter according to the change of the real-time signal, wherein the adjustment of the weight coefficient is based on the filtering error e(t) and the learning rate η at the current moment;

[0114] S53, calculating the characteristic value of the signal after filtering, the characteristic value includes the amplitude, frequency and phase of the signal:

[0115]

[0116] Among them, i∈{1,2,…,n} represents the different dimensions of the signal, α k , β k , γ and δ k is the adjustment parameter, N is the number of signal samples, S f (t) is the filtered signal, s f,i (t) is the i-th signal after filtering, s i (t) is the i-th signal in the real-time signal data, f i (t) is the eigenvalue of the i-th signal after filtering, is the potential feature representation reconstructed by the autoencoder, and m is the dimension of the potential feature;

[0117] S54, dynamically adjusting the characteristic value of the filtered signal, and adaptively adjusting the signal processing method according to the change of the real-time signal;

[0118] S55, calculating the rate of change of the signal characteristic value, where the rate of change is used to evaluate the change of the signal characteristic value at different times and capture the dynamic characteristics of the signal;

[0119] S56. Generate an adaptive strategy for signal processing based on the rate of change of the signal characteristic value, and automatically optimize parameter settings through a self-learning mechanism to adapt to different processing tasks and environments.

[0120] In this implementation manner, S6 specifically includes:

[0121] S61, the preprocessed signal S'(t) = [s'1(t), s'2(t), ..., s' n (t)] for real-time analysis, where s' i (t) is the i-th signal dimension;

[0122] S62. Input the signal into the temporal deep learning model based on the hybrid attention mechanism. First, the signal features are weighted through the self-attention mechanism to capture the key temporal patterns and long-term dependencies:

[0123]

[0124] Among them, H e (t) is the hidden state of the encoder after self-attention processing, W Q,j and W K,j are the weights of the query and key matrices respectively, is the latent feature representation generated by the autoencoder, α j and β j is the weighting coefficient, is the result of convolution processing, ∈ is a constant term to prevent the denominator from being zero;

[0125] S63, the decoder part combines the output of the self-attention mechanism and the reconstructed features generated by the autoencoder Further analysis of the short-term characteristics of the signal generates the clamping state evaluation result R(t):

[0126]

[0127] in, is the reconstructed signal feature output by the decoder, γ j and δ j To adjust the parameters;

[0128] S64, adjusting the parameters of the model in real time based on the generated clamping state evaluation result;

[0129] S65. The model is regularly updated based on new signal data and evaluation results, and the model parameters are adaptively adjusted through continuous optimization.

[0130] In this implementation manner, the S7 specifically includes:

[0131] S71. Based on historical data and current evaluation results, a multi-objective optimization function is constructed. The optimization function combines the latent feature representation reconstructed by the autoencoder. And the filtered signal eigenvalue S f (t), the optimization objectives include the accuracy, robustness and real-time performance of the clamping state, and these objectives are weighted and combined through weight coefficients to generate a comprehensive optimization objective function;

[0132] S72, using fuzzy logic to assist in multi-objective optimization strategy, reasoning and weighing the optimization objectives, the fuzzy logic reasoning process includes accurately fuzzifying the optimization objectives, and generating the optimization strategy according to the preset fuzzy reasoning rules, the reasoning rules are based on historical data and current evaluation results, and are adjusted for specific key factors in the clamping state to generate a specific optimization strategy;

[0133] S73. Automatically adjust relevant parameters of the model according to the generated optimization strategy:

[0134]

[0135] Among them, θ new is the updated parameter, θ old is the current parameter, η is the learning rate, To optimize the gradient of the objective function, γ j and δ j To adjust the parameters, used to precisely control the weight adjustment;

[0136] S74, the iterative optimization process stops when the rate of change of the optimization objective function reaches a convergence condition;

[0137] S75. Apply the optimized parameters to the real-time clamping operation and continuously adjust the optimization strategy based on new signal data and evaluation results.

[0138] Embodiment 1:

[0139] In order to verify the feasibility of the present invention in practice, the present invention is applied to the precision machining workshop of a Chinese aviation manufacturing company, which is mainly responsible for the machining of core parts of aircraft engines. These parts have extremely high machining accuracy requirements, and any slight error may cause the parts to be scrapped, thereby affecting the safety of the entire engine. Therefore, the clamping state of the tool is crucial to the stability of the machining process and the final quality of the parts.

[0140] In the traditional machining process, workers mainly rely on experience to judge the clamping status of the tool. This method is significantly subjective and easily affected by the operator's experience level and fatigue state, resulting in insufficient reliability of clamping status monitoring, which in turn affects machining accuracy and production efficiency. To overcome these problems, the workshop introduced a tool clamping status assessment method based on self-encoding reconstruction features.

[0141] During the machining process, the system collects the vibration, temperature, acoustic emission and strain signals of the tool in real time through multimodal sensors installed near the tool and the workpiece, reflecting the working status and clamping quality of the tool. These signals are transmitted to the central control system, and the autoencoder is used for deep feature extraction and noise reduction. By learning the sparse features in the signal, the high-dimensional data is compressed into a low-dimensional space, the noise and redundant information are filtered out, and the residual signal is generated. Subsequently, the residual signal is multi-scale decomposition and dynamically reconstructed to identify the instantaneous abnormalities and accumulated abnormal signals in the tool clamping process, and prevent and correct potential problems. The reconstructed signal enters the fuzzy logic reasoning system, which performs multi-objective optimization analysis based on the preset rule base, processes the uncertainty of the signal, and automatically adjusts the clamping state according to the current working conditions to optimize the machining process. Based on the results of the fuzzy logic reasoning, the system provides adjustment suggestions to the operator, such as re-clamping the tool or adjusting the cutting speed. After the operator makes adjustments according to the suggestions, the system collects the signal again, verifies the adjustment effect, and continuously optimizes the clamping state. Finally, the system evaluates the clamping status during the entire machining process, and further optimizes the machining strategy through feedback loops to ensure that the tool clamping status is always in the best state, thereby improving machining accuracy and efficiency. To verify the effectiveness of this method, the workshop selected 100 identical aircraft engine turbine blades for machining tests, which were divided into two stages for comparison.

[0142] Table 1: Comparison of machining effects of tool clamping state evaluation methods based on self-encoding reconstruction features

[0143] project Traditional methods Evaluation method based on autoencoder reconstruction features Total number of processed blades 100 100 Number of blades with severe machining errors 2 0 Number of blades with slightly substandard surface roughness 3 1 Blade qualification rate 95% 99% Average processing time per blade (hours) 2.0 1.8 Total processing time (hours) 200 180 Extended tool life - +20% Recommended times for tool adjustment none 5 times <![CDATA[Vibration signal (m / s 2 )]]> 15.3 10.2 Temperature signal (℃) 60.0 50.0 Acoustic emission signal (dB) 85 75 Strain signal (μm / m) 1200 950 <![CDATA[Crack area change (mm 2 )]]> 200 180 Blade surface roughness (micrometers) 0.8 0.6 Blade size deviation (mm) 0.05 0.02 Tool replacement frequency (times / 100 blades) 2 times 1 time

[0144] The data in the table show that when using the traditional method, the qualified rate of the blades is 95%, while after using the evaluation method based on self-encoding reconstruction features, the qualified rate of the blades is increased to 99%. In addition, the new method reduces the average processing time of each blade to 1.8 hours, and the total processing time is reduced to 180 hours, reflecting a 10% efficiency improvement. More importantly, the new method significantly extends the service life of the tool by 20%, which not only reduces production costs, but also reduces the frequency of tool replacement.

[0145] This embodiment demonstrates the application of the tool clamping state assessment method based on autoencoder reconstruction features in the aerospace engine parts processing scenario. Through the combination of multimodal signal acquisition, deep feature extraction of autoencoders, fuzzy logic reasoning and multi-objective optimization technology, this method realizes real-time monitoring and intelligent adjustment of the tool clamping state. Experimental data show that the use of this method can significantly improve machining accuracy, reduce machining errors, reduce production costs, and improve production efficiency, providing strong technical support for the field of high-precision manufacturing. The successful application of this method shows that it has broad application prospects in the future field of intelligent manufacturing.

[0146] 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. A tool clamping state evaluation method based on self-encoding reconstruction features, characterized in that: The steps include: S1, through the multimodal sensor installed near the tool, multiple modal signals are collected in real time, including vibration, temperature, acoustic emission, and strain signals; S2. Preprocess the collected multimodal signals and perform multimodal signal feature fusion to generate a unified feature vector; S3. Input the generated feature vector into a dynamic sparsity adaptive autoencoder, perform deep feature extraction on the fusion signal through the autoencoder, and dynamically adjust the sparsity parameters to adapt to signal changes under different working conditions; S4. Compare the reconstructed feature vector output by the autoencoder with the original feature vector to construct a residual signal set. Based on the adaptive multi-scale decomposition and dynamic reconstruction residual detection mechanism, the residual signal is multi-scale decomposition and dynamic reconstruction are performed to identify instantaneous anomalies and long-term accumulated abnormal signals; S5, bio-inspired adaptive signal processing technology, which adaptively adjusts signal processing methods according to real-time signal changes to cope with different signal characteristics and noise levels; S6. Perform real-time analysis on the pre-processed signal, combine it with the time series deep learning model based on the hybrid attention mechanism, analyze the time series pattern and characteristics of the signal, and generate the clamping status evaluation result; S7. Based on historical data and current evaluation results, a multi-objective optimization strategy assisted by fuzzy logic is used to perform fuzzy reasoning and trade-offs between multiple optimization objectives, generate an optimization strategy for the clamping state, and automatically adjust relevant parameters to optimize the clamping operation; S8. Through the interaction between the operator and the system, the model parameters are adjusted in real time, and the parameter adjustment behavior is recorded and analyzed. The system automatically optimizes the parameter settings through the self-learning mechanism to adapt to different processing tasks and environments.

2. The tool clamping state evaluation method based on self-encoding reconstruction features according to claim 1 is characterized in that: The S3 specifically includes: S31. Input the generated feature vector into the dynamic sparsity adaptive autoencoder. The feature vector is represented as X=[x1,x2,…,x n ], where n is the feature dimension; S32, the adaptive autoencoder includes an encoder and a decoder, and the encoder compresses the input feature vector into a latent space representation: Where Z is the latent space representation, f enc is the encoder function, θ enc is the encoder parameter, W enc is the encoder weight matrix, b enc is the bias vector, α,∈ and δ are small constants, ∥W enc X∥ 2 is the square of the two-norm of the vector after the encoder linear transformation, is the natural logarithm of the square of the eigenvalue; S33. Reconstruct the input feature vector from the latent space representation through the decoder: in, is the reconstructed eigenvector, f dec is the decoder function, θ dec is the decoder parameter, W dec is the decoder weight matrix, b dec is the bias vector, σ is the activation function, λ, μ and ν are the adjustment parameters, z i is the i-th component in the latent space representation, sin(W dec Z) is the sine term after linear transformation of the latent space representation, is the cosine term of the latent space representation; S34. Calculate the reconstruction error between the original eigenvector and the reconstructed eigenvector: in, is the reconstruction error, x i is the i-th component in the original eigenvector, is the i-th component in the reconstructed eigenvector, β is the adjustment parameter, is the natural logarithm of the square of the error, is the fourth power of the error; S35. Use a dynamic sparsity adjustment mechanism to adjust the sparsity parameters of the autoencoder based on the reconstruction error: Among them, θ spar is the sparsity parameter, η is the learning rate, γ and δ are regularization coefficients, z i is the i-th eigenvalue in the latent space, is the partial derivative of the reconstruction error with respect to the sparsity parameter, k is the number of specific sparsity parameters, log e (1+exp(-||z i ||)) is the logarithmic term of the latent space representation, is the square term of the latent space representation; S36. Through multiple iterative training, the sparsity parameters are adaptively adjusted to optimize the reconstruction capability of the autoencoder and adapt to signal changes under different working conditions.

3. The tool clamping state evaluation method based on self-encoding reconstruction features according to claim 1 is characterized in that: The S4 specifically includes: S41. Compare the reconstructed feature vector output by the autoencoder with the original feature vector to construct a residual signal set, where each residual signal represents the difference between the original feature value and the reconstructed feature value; S42, performing adaptive multi-scale decomposition on the constructed residual signal set to generate components of different scales, each component representing a change of the residual signal at a different time scale; S43, dynamically reconstructing the components after the multi-scale decomposition, and adjusting the reconstructed signal according to the components of the multi-scale decomposition to adapt to changes in the signal; S44. Calculate the instantaneous outlier value and the long-term cumulative outlier value of each reconstructed signal, where the instantaneous outlier value indicates the degree of abnormality of the signal at the current moment, and the long-term cumulative outlier value indicates the degree of abnormality accumulated over a period of time. S45. Identify instantaneous abnormalities and long-term accumulated abnormal signals, set dynamic thresholds, dynamically adjust the thresholds according to real-time signal changes, and detect signal abnormalities; S46. Based on the set dynamic threshold, detect instantaneous anomalies and long-term accumulated anomaly signals and generate anomaly detection results, each of which indicates whether the corresponding feature is abnormal; S47. Output anomaly detection results and update model parameters in real time. Continuously optimize the adaptive multi-scale decomposition and dynamic reconstruction residual detection mechanism based on new signal data and detection results to adapt to different working conditions and signal feature changes.

4. The tool clamping state evaluation method based on self-encoding reconstruction features according to claim 1 is characterized in that: The S5 specifically includes: S51, using biologically inspired adaptive filter to process real-time signals, the output signal of the filter is expressed as S f Ct), the filter processes the signal based on past signal samples and current weight coefficients, and adapts to signal changes by dynamically adjusting the weight coefficients; S52, adaptively adjusting the weight coefficient of the filter according to the change of the real-time signal, where the adjustment of the weight coefficient is based on the current filtering error e(t) and the learning rate η; S53. Calculate the characteristic value of the filtered signal, which includes the amplitude, frequency and phase of the signal: Among them, i∈{1,2,…,n} represents the different dimensions of the signal, α k , β k , γ, and δ k is the adjustment parameter, N is the number of signal samples, S f (t) is the filtered signal, s f,i (t) is the i-th signal after filtering, s i (t) is the i-th signal in the real-time signal data, f i (t) is the eigenvalue of the i-th signal after filtering, is the potential feature representation reconstructed by the autoencoder, and m is the dimension of the potential feature; S54, dynamically adjusting the characteristic value of the filtered signal, and adaptively adjusting the signal processing method according to the change of the real-time signal; S55. Calculate the rate of change of the signal characteristic value, where the rate of change is used to evaluate the change of the signal characteristic value at different times and capture the dynamic characteristics of the signal; S56. Generate an adaptive strategy for signal processing based on the rate of change of the signal characteristic value, and automatically optimize parameter settings through a self-learning mechanism to adapt to different processing tasks and environments.

5. The tool clamping state evaluation method based on self-encoding reconstruction features according to claim 1 is characterized in that: The S6 specifically includes: S61, the pre-processed signal S'(t) = [s'1(t), s'2{t), ..., s' n (t)] for real-time analysis, where s' i (t) is the i-th signal dimension; S62. Input the signal into a time series deep learning model based on a hybrid attention mechanism. First, the signal features are weighted through the self-attention mechanism to capture key time series patterns and long-term dependencies: Among them, H e (t) is the hidden state of the encoder after self-attention processing, W Q,j and W K,j are the weights of the query and key matrices respectively, is the latent feature representation generated by the autoencoder, α j and β j is the weighting coefficient, is the result of convolution processing, ∈ is a constant term to prevent the denominator from being zero; S63, the decoder combines the output of the self-attention mechanism and the reconstructed features generated by the autoencoder Further analysis of the short-term characteristics of the signal generates the clamping state evaluation result R(t): in, is the reconstructed signal feature output by the decoder, γ j and δ j is the adjustment parameter; S64, adjusting the parameters of the model in real time based on the generated clamping state evaluation results; S65. The model is regularly updated based on new signal data and evaluation results, and the model parameters are adaptively adjusted through continuous optimization.

6. The tool clamping state evaluation method based on self-encoding reconstruction features according to claim 1 is characterized in that: The S7 specifically includes: S71. Based on historical data and current evaluation results, a multi-objective optimization function is constructed, which combines the latent feature representation reconstructed by the autoencoder And the filtered signal eigenvalue S f (t), the optimization objectives include the accuracy, robustness and real-time performance of the clamping state, and these objectives are weighted and combined through weight coefficients to generate a comprehensive optimization objective function; S72. Using fuzzy logic to assist in multi-objective optimization strategy, reasoning and weighing the optimization objectives. The fuzzy logic reasoning process includes accurately fuzzifying the optimization objectives and generating an optimization strategy based on preset fuzzy reasoning rules. The reasoning rules are based on historical data and current evaluation results, and are adjusted for specific key factors in the clamping state to generate a specific optimization strategy. S73. Automatically adjust relevant parameters of the model according to the generated optimization strategy: Among them, θ new is the updated parameter, θ old is the current parameter, η is the learning rate, To optimize the gradient of the objective function, γ j and δ j To adjust the parameters, used to precisely control the weight adjustment; S74, the iterative optimization process stops when the rate of change of the optimization objective function meets the convergence condition; S75. Apply the optimized parameters to the real-time clamping operation and continuously adjust the optimization strategy based on new signal data and evaluation results.

Citation Information

Patent Citations

  • Method and device for determining service life information of numerical control machine tool cutter

    CN115673873A

  • TF-IGLSTM thermal error prediction model and intelligent processing and error control system

    CN115859500A

  • Drilling tool state monitoring method based on multi-sensor fusion

    CN116871978A

  • Tool remaining life prediction method based on weighted trend coefficient and self-coding

    CN117077313A

  • Transformer state evaluation method and device based on automatic encoder and logistic regression

    CN117131331A

Cited By

  • Method and system for evaluating performance of clamping auxiliary device based on reinforcement learning

    CN120558539A

  • Clamping auxiliary device performance evaluation method and system based on reinforcement learning

    CN120558539B

  • Method for evaluating running state of whole-house water purification system of smart home

    CN121435082A