Cutter wear monitoring method based on physical guide gating Mama network

Abstract

The invention provides a cutter wear monitoring method based on a physical guide gating Mama network. The method comprises the steps of time-frequency domain feature extraction, sensitive feature screening, construction of a physical guide gating Mama network by adopting an implicit physical model, establishment of a mapping relation between sample data and a tool wear value and the like. According to the method, the wear degree can be represented timely and accurately, a quantitative basis is provided for preventive replacement before failure, and productivity reduction and surface quality degradation caused by excessive wear are effectively inhibited, so that the comprehensive cost caused by equipment abnormity and defective products is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing, and particularly relates to a tool wear monitoring method based on a physically guided gated Mamba network. BACKGROUND

[0002] As a core component of the cutting system, the health status of the tool directly determines the machining efficiency of the machine tool and the quality of the workpiece. With the rapid development of intelligent manufacturing, enterprises have increasingly high demands for precision and reliability in the cutting process. However, tool wear inevitably occurs during machining. When the tool wear reaches a certain degree, it will seriously affect the workpiece machining quality and machine tool machining efficiency. Therefore, real-time monitoring of tool wear state is of great significance for timely adjustment of processing parameters and tool replacement.

[0003] Existing researches are generally divided into mechanism modeling and data-driven methods. The mechanism-based method relies on cutting mechanics, friction and wear, and heat conduction models to reveal the evolution mechanism, and has good physical interpretation, but it often depends on prior assumptions and is difficult to meet the application scenarios of complex working conditions and materials. The data-driven method learns the information related to tool wear from vibration, force, acoustic emission and other multi-source signals through deep learning algorithms, and establishes a mapping relationship between input data and wear value, thereby realizing tool wear monitoring. However, the data-driven method is usually highly sensitive to data quality and distribution stability, and the deep learning algorithm modeling process is a black box operation, which has low interpretability and reliability. Although the combination of numerical modeling and data-driven methods to a large extent alleviates the limitations of mechanism modeling and data-driven methods, the introduction of physical information usually depends on empirical physical formulas and does not have universality. In addition, learning tool wear-related information through complex neural networks often leads to high model complexity, making it difficult to meet the real-time and deployability requirements in industrial applications.

[0004] Therefore, it is urgent to develop a tool wear monitoring method that combines physical knowledge and data modeling capabilities and has high real-time performance and lightweight deployability. SUMMARY

[0005] The purpose of the present application is to provide a tool wear monitoring method based on a physically guided gated Mamba network to solve the problems in the prior art.

[0006] The technical solution adopted to achieve the purpose of the present application is as follows: a tool wear monitoring method based on a physically guided gated Mamba network, comprising the following steps:

[0007] S1) Multi-source sensing signal acquisition. Collecting force signals, vibration signals and acoustic emission signals generated during milling.

[0008] S2) denoising, truncating and feature extraction of the original signal to generate a standardized original feature set.

[0009] S3) correlation study on the extracted original feature set, evaluating the effectiveness and correlation of each feature in representing the wear information by calculating the comprehensive sensitivity index between the feature and the tool wear.

[0010] S4) according to the comprehensive sensitivity index value from large to small, screening the top-k features with the largest comprehensive sensitivity index value to form a sensitive feature set for model training and testing.

[0011] S5) using Gated Mamba to extract tool wear related information in the input features and establishing long-term dependencies between features.

[0012] S6) using two multi-layer perceptrons to construct the tool wear physical mechanism and tool wear monitoring value output respectively.

[0013] Further, the step S2) comprises the following sub-steps:

[0014] S2.1) cutting-in and cutting-out rejection. The transient part of 2.5% at the beginning and end of the signal is removed, and the steady-state machining signal is retained.

[0015] S2.2) multi-domain feature extraction. Time domain, frequency domain and time-frequency domain features are extracted from the sensor signals after truncation and rejection. A total of 32 features are obtained for each sensor signal.

[0016] S2.3) feature normalization. The extracted features are normalized using the maximum and minimum normalization method to ensure that all feature amplitudes are within the same range:

[0017]

[0018] In the formula: represents the feature i extracted based on the sensor signal s, and respectively represent the maximum and minimum values in the corresponding feature.

[0019] Further, the comprehensive sensitivity index in step 3 is specifically:

[0020]

[0021]

[0022]

[0023] In the formula: Ic and Im respectively represent the absolute Pearson correlation coefficient and mutual information. denotes the feature i extracted from sensor signal s at the jth pass, and N denotes the total number of passes. denotes the average value of the feature i extracted from sensor signal s. and denote the true wear value and the average wear value of the jth pass, respectively. denotes the joint probability distribution between the feature and the wear value, and denote the marginal probability distribution of the feature and the wear value, respectively. and denote the weights of Ic and Im, which are set to 0.8 and 0.2, respectively.

[0024] Further, the sensitive feature set in step 4 is constructed as follows:

[0025] The comprehensive sensitive index values calculated in step 3 are sorted in descending order, and the top-k features in the order are selected for model training and testing. To avoid data leakage, the selected feature categories are only based on the training set, i.e., the test set selects the same features as the training set.

[0026] Further, the step 5 adopts the gated Mamba network to extract information related to tool wear from the input features, and the process can be represented as:

[0027] S5.1) Input projection:

[0028] For the lth layer of the gated Mamba, the time series at the t-1 time point and the hidden node are taken as inputs to update the state at the t time point, and the corresponding formula can be expressed as:

[0029]

[0030] In the formula: denotes the updated state vector at the t time point, and are the weights and biases.

[0031] S5.2) State aggregation:

[0032] The layer normalization is used to aggregate the state vector at the t time point and the hidden node at the t-1 time point , and the aggregation process can be represented as:

[0033]

[0034] In the formula: ds denotes the hidden node at the t-1 time point where j denotes the j-th feature dimension, and LayerNorm denotes the layer normalization.

[0035] S5.3) Gating computation:

[0036] The gating vector is computed using state aggregation for regulating the signal flow:

[0037]

[0038]

[0039]

[0040] where , , denote the reset, update, and state gate vectors, respectively. and are the sigmoid and tanh activation functions, respectively. , , and are the weights for the reset, update, and state gates, , , and are the corresponding biases.

[0041] S5.4) State space update:

[0042] The state space vector is updated by fusing the hidden node and the reset gate :

[0043]

[0044] where A and B are the transformation matrices for the hidden node and the reset gate, respectively. is the state space vector that is dynamically updated during the model training process.

[0045] S5.5) Candidate state update:

[0046] Subsequently, the state gate vector is linearly projected for constructing the auxiliary candidate state:

[0047]

[0048] where C denotes the state gate vector transformation matrix.

[0049] S5.6) Gated memory update:

[0050] The gating memory is updated by fusing the state space vector and the state gate candidate vector for the next time point hidden node:

[0051]

[0052] S5.7) Output:

[0053] The gating memory is updated by fusing the state space vector and the state gate candidate vector for the next time point hidden node: and the state vector Update the t time series:

[0054]

[0055] where D denotes the gating memory transformation matrix, is the t time series, which is used as the input for the t+1 time point.

[0056] For multiple gating Mamba units, the l-th layer t time point update can be expressed as:

[0057]

[0058] where L denotes the total gating Mamba layer, denotes the mapping of the l-th layer t time point. Initially, is the initial input feature. is composed of an all-zero vector. The final output of the model can be expressed as:

[0059]

[0060] where T denotes the total time step.

[0061] Further, the two multi-layer perceptrons used in step 6 to construct the tool wear physical mechanism and tool wear monitoring value output can be specifically expressed as:

[0062]

[0063]

[0064] where l denotes the linear layer, denotes the ReLU activation function, and is the data-driven multi-layer perceptron weight, denotes the implicit nonlinear mapping, and denotes the first and second order derivatives of the gating Mamba output feature, and is the first and second order derivatives of the original input feature.

[0065] Further, the training process of the gated Mamba network is based on historical data, and in the training process, the model update evaluation is carried out through the following loss function:

[0066]

[0067] In the formula, , , , respectively represent the data-driven loss, the physical consistency loss, the knowledge loss and the smoothing loss, and are corresponding weights.

[0068] Further, the data-driven loss, the physical consistency loss, the knowledge loss and the smoothing loss are specifically as follows:

[0069]

[0070]

[0071]

[0072]

[0073] In the formula, N represents the total number of passes, and respectively represent the actual wear value and the model predicted wear value of the i-th pass, represents the predicted tool wear rate.

[0074] The application also provides a tool wear monitoring system based on the physically guided gated Mamba network of any one of the above-mentioned methods, comprising a multi-source sensor, a processing unit, a workbench, a label value measurement system, a multi-channel data acquisition card and a data display module.

[0075] The workbench is used to support the milling workpiece. The cutting tool is used to process the milling workpiece. The multi-source sensor comprises a magnetic suction type force sensor, a vibration sensor and an acoustic emission sensor. The multi-source sensor is attached to the surface of the workbench and the milling workpiece. The multi-source sensor is connected with the multi-channel data acquisition card. The multi-channel data acquisition card and the data display module are both connected with the processing unit.

[0076] In the model construction stage, the multi-source sensor collects multi-source physical signals in the processing process in real time for feature extraction and model training. The multi-source physical signal collection covers the whole life cycle of the tool. The multi-channel data acquisition card realizes synchronous acquisition and digital conversion of multi-channel signals, and transmits them to the processing unit for preprocessing and feature extraction. The processing unit runs data preprocessing, feature screening, gated Mamba network training and verification algorithm. After each milling, the tool flank wear condition is measured by the label value measurement system as the true label for model calibration. The trained tool wear monitoring model is deployed in the processing unit.

[0077] In the subsequent real-time monitoring stage, the multi-source sensor collects multi-source physical signals in the processing process in real time. The multi-channel data acquisition card realizes synchronous acquisition and digital conversion of multi-channel signals, and transmits them to the processing unit. The processing unit runs the tool wear monitoring model to output the wear value in real time. The wear value is visualized and output through the data display module.

[0078] The technical effects of the present application are self-evident:

[0079] A. Introduce wear monotonicity and physical consistency constraints (implicit physical branch and regularization term) in the learning process, strengthen the description of wear evolution law and improve the explainability and stability of the results;

[0080] B. Introduce gated Mamba timing unit, model long-term sequence dependence of sensitive features, and realize dynamic screening and retention of effective degradation information through state aggregation, gated memory and state space update operations, so as to obtain more stable prediction output under complex working conditions;

[0081] C. Take multi-domain feature extraction and sensitivity evaluation as the entrance, combine the joint optimization of gated Mamba and multiple losses to form an end-to-end process, which effectively improves the accuracy and stability of wear prediction while keeping the computational overhead controllable. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 The flowchart of the tool wear monitoring method based on the physical guided gated Mamba network;

[0083] Figure 2 The structure diagram of the gated Mamba network;

[0084] Figure 3 The framework diagram of the physical guided gated Mamba;

[0085] Figure 4 The tool wear multi-source sensor data acquisition device diagram;

[0086] Figure 5 The data preprocessing and multi-domain feature extraction diagram;

[0087] Figure 6 For sensor feature distribution map;

[0088] Figure 7 For milling tool wear test set of experimental data set monitoring results schematic; test subset C1 see Figure 7 (a), test subset C4 see Figure 7 (b), test subset C6 see Figure 7 (c).

[0089] In the figure: cutting tool 1, support milling workpiece 2, workbench 3, label value measurement system 4, multi-channel data acquisition card 8, data display module 9. DETAILED DESCRIPTION

[0090] The application will be further described below in conjunction with examples, but should not be understood as limiting the above-mentioned subject matter of the application only to the following examples. Various substitutions and modifications can be made according to ordinary technical knowledge and conventional means in the art without departing from the above-mentioned technical ideas of the application, and all should be included in the protection scope of the application.

[0091] Example 1:

[0092] The embodiment provides a tool wear monitoring method based on a physical guided gating Mamba network, comprising the following steps:

[0093] S1) Multi-source sensing signal acquisition. Acquire force signals, vibration signals and acoustic emission signals generated during the milling process.

[0094] S2) Denoising, truncation and feature extraction of original signals to generate a standardized original feature set. The step S2) comprises the following substeps:

[0095] S2.1) Cutting in and cutting out rejection. Rejection of transient part of 2.5% at the beginning and end of the signal, and retention of steady-state machining signal.

[0096] S2.2) Multi-domain feature extraction. Time domain, frequency domain and time-frequency domain feature extraction are carried out on the sensor signals after truncation and rejection processing. A total of 32 features are obtained for each sensor signal.

[0097] S2.3) Feature normalization processing. The extracted features are normalized by using the maximum and minimum normalization method to ensure that all feature amplitudes are within the same range:

[0098]

[0099] In the formula: represents the feature i extracted based on the sensor signal s, and respectively represent the maximum and minimum values in the corresponding feature.

[0100] S3) Correlation study on the extracted original feature set, evaluate the effectiveness and correlation of each feature in representing the wear information by calculating the comprehensive sensitivity index between the feature and the tool wear. The comprehensive sensitivity index is specifically:

[0101]

[0102]

[0103]

[0104] where Ic and Im represent the absolute Pearson correlation coefficient and mutual information, respectively. represents the feature i extracted from the sensor signal s at the jth pass, and N represents the total number of passes. represents the average value of the feature i extracted from the sensor signal s. and are the true wear value and the average wear value of the jth pass, respectively. represents the joint probability distribution between the feature and the wear value, and are the marginal probability distributions of the feature and the wear value, respectively. and are the weights of Ic and Im, which are set to 0.8 and 0.2, respectively.

[0105] S4) According to the comprehensive sensitivity index value from large to small, filter the top-k features with the largest comprehensive sensitivity index value to form the sensitive feature set for model training and testing. The sensitive feature set is constructed as follows: the comprehensive sensitivity index values calculated in step 3 are sorted in descending order, and the top-k features are selected for model training and testing. To avoid data leakage, the selected feature categories are only based on the training set, i.e. the test set selects the same features as the training set.

[0106] S5) Use gated Mamba to extract tool wear related information in the input features and establish long-term dependencies between features. Step S5) specifically includes the following sub-steps:

[0107] S5.1) Input projection:

[0108] For the lth layer of gated Mamba, the time series at t-1 time point and hidden nodes are taken as input to update the state at t time point, which can be expressed as:

[0109]

[0110] where: denotes the state vector updated at time t, and are weights and biases.

[0111] S5.2) State aggregation:

[0112] The state vector at time t is aggregated using layer normalization and the hidden nodes at time t-1 The aggregation process can be represented as:

[0113]

[0114] where ds denotes the feature dimension of the hidden nodes at time t-1 j denotes the j-th feature dimension, and LayerNorm denotes the layer normalization process.

[0115] S5.3) Gating computation:

[0116] The gating vectors are computed using the state aggregation to regulate the signal flow:

[0117]

[0118]

[0119]

[0120] where: , , denote the reset, update, and state gate vectors, respectively. and are the sigmoid and tanh activation functions, respectively. , , and are the weights for the reset, update, and state gates, , , and are the corresponding biases.

[0121] S5.4) State space update:

[0122] The state space vector is updated by fusing the hidden nodes and the reset gates

[0123]

[0124] where A and B are the transformation matrices for the hidden nodes and reset gates, respectively.​ a state space vector dynamically updated during the model training process.

[0125] S5.5) Candidate state update:

[0126] Subsequently, a linear projection of the state gate vector is performed to construct an auxiliary candidate state:

[0127]

[0128] where C denotes the state gate vector a transformation matrix.

[0129] S5.6) Gated memory update:

[0130] The gated memory is updated by fusing the state space vector and the state gate candidate vector for the hidden nodes at the next time point:

[0131]

[0132] S5.7) Output:

[0133] The updated gated memory and the state vector update the t time sequence:

[0134]

[0135] where D denotes the gated memory transformation matrix, is the t time sequence, which is used as the input at the t+1 time point.

[0136] For multiple gated Mamba units, the update at the t time point of the lth layer can be expressed as:

[0137]

[0138] where L denotes the total number of gated Mamba layers, denotes the mapping at the t time point of the lth layer. Initially, is the initial input feature. is composed of an all-zero vector. The final output of the model can be expressed as:

[0139]

[0140] where T denotes the total number of time steps.

[0141] S6) Two multi-layer perceptrons are used to construct the tool wear physical mechanism and the tool wear monitoring value output. Specifically, the two multi-layer perceptrons used to construct the tool wear physical mechanism and the tool wear monitoring value output can be expressed as:

[0142]

[0143]

[0144] where l denotes linear layer, denotes ReLU activation function, and is the data-driven multi-layer perceptron weight, denotes implicit nonlinear mapping, and denotes the first and second order time derivative of the gated Mamba output feature, and is the first and second order derivative of the original input feature.

[0145] The embodiment solves the problems of existing hybrid driving tool wear monitoring methods that fuse physical knowledge and data modeling capabilities, such as dependence on empirical formula, low real-time performance, high complexity, and poor accuracy. The embodiment learns information related to tool wear from multi-source sensor data through a gated Mamba network, and improves the explainability and reliability of tool wear monitoring results through an implicit physical modeling mechanism, achieving real-time and accurate perception of tool wear state.

[0146] Embodiment 2:

[0147] The main content of the embodiment is the same as that of embodiment 1, wherein the training process of the gated Mamba network is based on historical data, and model update evaluation is performed through the following loss function during the training process:

[0148]

[0149] wherein, , , , denote data-driven loss, physical consistency loss, knowledge loss, and smoothing loss, respectively, and are the corresponding weights.

[0150] Embodiment 3:

[0151] The main content of the embodiment is the same as that of embodiment 1, wherein the data-driven loss, the physical consistency loss, the knowledge loss, and the smoothing loss are as follows:

[0152]

[0153]

[0154]

[0155]

[0156] N represents the total number of passes, and respectively the actual wear value and the model predicted wear value of the i-th pass, represents the predicted tool wear rate.

[0157] Example 4:

[0158] The main content of this embodiment is the same as any one of examples 1-3, wherein the data set used in this embodiment is the PHM Society 2010 milling tool wear public data set (PHM2010). The data set collects multi-modal sensor information based on a high-speed numerical control milling platform, including cutting forces and vibration signals in x, y, z directions, and acoustic emission (AE-RMS) signals. As shown in Figure 4 is the PHM2010 experimental platform, and during the experiment, six tools of the same type are used for full-life milling machining, and sensor information generated during tool machining is collected through sensors arranged on the workpiece table. The specific parameters of the experimental process are shown in Table 1, wherein tool 1, tool 4, and tool 6 are measured for tool flank wear after each milling by a wear measuring instrument, and the average wear value of the three blades is taken as the final use label value. Therefore, three tools containing complete label values are used as training and testing objects in this embodiment. In order to ensure fairness and evaluation robustness, the present application adopts a leave-one-tool cross-validation strategy of three-fold cross-validation: in each round, the data of two tools are selected as the training set, and the remaining one is used as the independent test set, and the training and verification are completed in three rounds, thereby improving the generalization ability and reliability of the model. The corresponding cross-validation is shown in Table 2.

[0159] Table 1

[0160]

[0161] Table 2

[0162]

[0163] After collecting multi-source sensor signals during milling machining, as shown in Figure 1 the specific implementation of the present application should also include the following steps:

[0164] Step 1: data preprocessing and multi-domain feature extraction

[0165] Sub-step (S1), as shown in Figure 4As shown, force, vibration, and acoustic emission sensors arranged on the worktable collect signals generated during the machining process, including cutting force and vibration signals in the x, y, and z directions, as well as acoustic emission signals. It is worth noting that the raw signals typically contain a significant amount of noise and invalid data. To improve signal representation, wavelet denoising is employed to reduce noise and enhance the signal-to-noise ratio. Furthermore, the signals collected during the tool's entry into and exit from the workpiece often exhibit non-steady-state characteristics. For example... Figure 5 As shown, in order to avoid transient interference, the noise-reduced signal is subjected to cut-in and cut-out portion rejection processing. In this invention, the 2.5% signal segments at the beginning and end of the signal are used as the cut-in and cut-out portions, and are truncated and rejected.

[0166] Sub-step (S2), Figure 5 The preprocessed multi-source sensor signals are subjected to multi-domain feature extraction, including 12 time-domain features such as mean, maximum value, and standard deviation to reflect the average time, impact, and fluctuation during the tool machining process. Secondly, 12 frequency-domain features such as spectral mean, spectral variance, and spectral skewness are extracted to characterize the amplitude modulation sideband and spectral structure changes during the cutting process. The specific time-domain and frequency-domain features are shown in Table 3. Finally, to capture non-stationary and transient information, three-level wavelet packet decomposition is used to obtain the energy of eight sub-bands and their energy ratios, forming eight time-frequency domain features. In summary, 32 different features are extracted from each sensor, and the resulting feature set can be represented as:

[0167]

[0168]

[0169] In the formula: s represents the s-th sensor, This represents the features extracted from the signal of the s-th sensor, which includes 32 features in the time domain, frequency domain, and time-frequency domain. S represents the total number of sensors.

[0170] Table 3

[0171]

[0172] Sub-step (S3), to avoid the impact of differences in sensor signals and feature amplitudes on model training and testing, uses min-max normalization to normalize all features, ensuring that all feature amplitudes are within the same dimension. The normalization process can be expressed as:

[0173]

[0174] In the formula: This represents the i-th feature extracted from sensor s. Features are obtained after normalization. and These are the maximum and minimum values ​​used in the normalization process. In this embodiment, to avoid data leakage, the maximum and minimum values ​​used in both the training and test data are provided by the training data.

[0175] Sub-step (S4): Although the normalized multi-domain features maintain the same magnitude range, not all features can effectively characterize the tool degradation process. Furthermore, not all sensors contain information related to tool wear. In addition, using all features extracted from all sensors as input for model training and testing can significantly increase model complexity. Therefore, this invention constructs a comprehensive sensitivity index to evaluate features by fusing Pearson correlation coefficient and mutual information, expressed by the following formula:

[0176]

[0177]

[0178]

[0179] In the formula: Ic and Im represent the absolute Pearson correlation coefficient and mutual information, respectively. Let i represent the feature extracted from the sensor signal s during the j-th tool pass, and N represent the total number of tool passes. This represents the average value of feature i extracted from sensor signal s. and These are the actual wear value and the average wear value for the j-th tool pass, respectively. This represents the joint probability distribution between features and wear values. and These are the marginal probability distributions of the features and wear values, respectively. and The weights for Ic and Im are set to 0.8 and 0.2 respectively.

[0180] Step 2: Model Training

[0181] Model building

[0182] While the filtered features obtained in step 1 are sufficient to characterize the gradual process of tool wear, they still cannot directly infer the current wear state of the tool. Therefore, this invention employs a gated Mamba network to extract tool wear-related information from the input features to establish a mapping between the input sensing features and the tool wear value. The feature learning process of the gated Mamba network can be represented as follows:

[0183] S5.1) Input projection:

[0184] For the l-th level gated Mamba, the time series at time t-1 is... and hidden nodes As input, used to update the state at time t, the corresponding formula can be expressed as:

[0185]

[0186] In the formula: This represents the state vector updated at time t. and For weights and biases.

[0187] S5.2) State aggregation:

[0188] Layer-normalized aggregation of state vectors at time t and hidden nodes at time point t-1 The polymerization process can be represented as:

[0189]

[0190] In the formula: ds represents the hidden node at time t-1. The feature dimension is denoted by j, where j represents the j-th feature dimension, and LayerNorm represents the layer normalization process.

[0191] S5.3) Gating calculation:

[0192] The gating vector is calculated using state aggregation and used to regulate the signal flow.

[0193]

[0194]

[0195]

[0196] In the formula: , , These represent the reset gate, update gate, and state gate vector, respectively. and These are the sigmoid activation function and the tanh activation function, respectively. , , and To reset the weights of gates, update gates, and state gates, , , and This is the corresponding bias.

[0197] S5.4) State space update:

[0198] By merging hidden nodes and reset door To update the state space vector:

[0199]

[0200] In the formula, A and B are the transformation matrices of the hidden node and the reset gate, respectively. This is the state space vector that is dynamically updated during model training.

[0201] S5.5) Candidate State Update:

[0202] Subsequently, a linear projection is performed on the state gate vector to construct auxiliary candidate states:

[0203]

[0204] In the formula, C represents the state gate vector. Transformation matrix.

[0205] S5.6) Gated memory update:

[0206] The gating memory is updated by fusing the state space vector and the state gate candidate vector, which is used to hide nodes at the next time step.

[0207]

[0208] S5.7) Output:

[0209] By integrating and updating gated memory and state vector Update the t-time series:

[0210]

[0211] In the formula, D represents the gated memory transformation matrix. For time series t, it is used as the input at time point t+1.

[0212] For multiple gated Mamba units, their update at time t in layer l can be expressed as:

[0213]

[0214] In the formula, L represents the total gated Mamba layer. This represents the mapping at time t in the l-th layer. Initially, in the first layer... , which are the initial input features. It consists of vectors of all zeros. The final output of the model can be represented as:

[0215]

[0216] In the formula, T represents the total time steps.

[0217] Finally, the features extracted by Mamba are mapped to tool wear values ​​through a fully connected layer. Simultaneously, the extracted features are differentiated to construct an implicit physical model, improving the stability and reliability of the prediction results. The corresponding processing can be represented as follows:

[0218]

[0219]

[0220] In the formula, l represents a linear layer. Represents the ReLU activation function. and For data-driven multilayer perceptron weights, This represents an implicit nonlinear mapping. and This represents the time first and second derivatives of the output features of the gated Mamba. and To find the first and second derivatives of the original input features.

[0221] Loss function construction

[0222] To guide parameter updates during model training, this invention employs a hybrid loss function, comprising data-driven loss, physical consistency loss, knowledge loss, and smoothing loss. The hybrid loss function is as follows:

[0223]

[0224] In the formula, , , , These represent data-driven loss, physical consistency loss, knowledge loss, and smoothing loss, respectively. and In this embodiment, the corresponding weights are... and All initial values ​​are set to 1. After the initial 1000 epochs, a judgment is made... and The weights are updated based on the size of the values. Specifically, assuming 1000 iterations, If it is established, then The half-decay is 0.5, and a check is performed every 200 iterations thereafter. If the condition is still met, the decay continues. The attenuation condition is similar. The specific loss function for each term is as follows:

[0225]

[0226]

[0227]

[0228]

[0229] In the formula, N represents the total number of tool passes. and Let be the actual wear value and the model-predicted wear value for the i-th time, respectively. This indicates the predicted tool wear rate.

[0230] Step 3: Performance Testing

[0231] To evaluate the performance of the design method in tool wear monitoring, root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were used. S2) The corresponding formulas are as follows:

[0232]

[0233]

[0234]

[0235] In the formula: N represents the total number of cuts made by the tool used for testing from initial operation to failure. and The tool wear value measured during the i-th cutting operation and the wear value estimated by the model are respectively... This represents the average value of the tool wear measurements.

[0236] The tool wear monitoring method based on physically guided gating Mamba described above has RMSE, MAE, and R2 values ​​for three test cases, as shown in Table 4.

[0237] Table 4

[0238]

[0239] To more clearly compare predicted and measured wear, Figure 7 The fitting curves of the present invention on three test subsets are shown: the dashed line represents the actual wear measured by microscopic images, and the solid line represents the estimated output of the tool wear monitoring method based on a physically guided gated Mamba network of the present invention.

[0240] Depend on Figure 7As shown in Table 4, the tool wear monitoring model constructed in this invention exhibits high prediction accuracy and stability across all test subsets. This method can characterize the degree of wear in a timely and accurate manner, providing a quantitative basis for preventative replacement before failure, effectively suppressing the decline in production capacity and surface quality deterioration caused by excessive wear, thereby reducing the overall cost caused by equipment malfunctions and defective products.

[0241] Example 5:

[0242] See Figure 4 This embodiment provides a tool wear monitoring system based on a physically guided gated Mamba network according to any one of the methods described in embodiments 1 to 4, including a multi-source sensor, a processing unit, a worktable 3, a tag value measurement system 4, a multi-channel data acquisition card 8, and a data display module 9.

[0243] The worktable 3 supports the milling workpiece 2. The cutting tool 1 is used to machine the milling workpiece 2. The multi-source sensor includes a magnetic force sensor, a vibration sensor, and an acoustic emission sensor. The multi-source sensor is attached to the worktable 3 and the surface of the milling workpiece 2. The multi-source sensor is connected to a multi-channel data acquisition card 8. Both the multi-channel data acquisition card 8 and the data display module 9 are connected to the processing unit.

[0244] During the model building phase, the multi-source sensors acquire multi-source physical signals in real time during the machining process for feature extraction and model training. The acquisition of multi-source physical signals covers the entire tool lifecycle. A multi-channel data acquisition card 8 enables simultaneous acquisition and digital conversion of multiple signals, which are then transmitted to the processing unit for preprocessing and feature extraction. The processing unit runs data preprocessing, feature filtering, gated Mamba network training, and validation algorithms. After each milling operation, the tool flank wear is measured using a label value measurement system 4 as a real label for model calibration. The trained tool wear monitoring model is then deployed to the processing unit.

[0245] In the subsequent real-time monitoring phase, the multi-source sensors acquire multi-source physical signals during the machining process in real time. The multi-channel data acquisition card 8 enables simultaneous acquisition and digital conversion of multiple signals, and transmits them to the processing unit. The processing unit runs the tool wear monitoring model to perform real-time inference and output wear values. The wear values ​​are then visualized through the data display module 9.

Claims

1. A tool wear monitoring method based on a physically guided gated Mamba network, characterized in that, Includes the following steps: S1) Multi-source sensor signal acquisition; acquiring force signals, vibration signals, and acoustic emission signals generated during the milling process; S2) The original signal is denoised, truncated, and its features are extracted to generate a standardized original feature set; S3) Conduct a correlation study on the extracted original feature set, and evaluate the effectiveness and correlation of each feature in characterizing wear information by calculating the comprehensive sensitivity index between the feature and tool wear; S4) Sort the features by comprehensive sensitivity index value from largest to smallest, and select the top-k features with the highest comprehensive sensitivity index value to form a sensitive feature set for model training and testing; S5) Gated Mamba is used to extract tool wear-related information from the input features and establish long-term dependencies between features; S6) Two multilayer perceptrons are used to construct the physical mechanism of tool wear and output the tool wear monitoring value respectively.

2. The tool wear monitoring method based on a physically guided gated Mamba network according to claim 1, characterized in that, Step S2) includes the following sub-steps: S2.1) Cut-in and cut-out rejection; reject the first and last 2.5% of the transient portion of the signal, and retain the steady-state processing signal; S2.2) Multi-domain feature extraction; Time-domain, frequency-domain, and time-frequency-domain feature extraction are performed on the truncated and eliminated sensor signals; a total of 32 features are obtained for each sensor signal; S2.3) Feature normalization: The extracted features are normalized using the minimum-maximum normalization method to ensure that all feature amplitudes are within the same range. In the formula: This represents the feature i extracted based on sensor signal s. and These represent the maximum and minimum values ​​of the corresponding features, respectively.

3. The tool wear monitoring method based on a physically guided gated Mamba network according to claim 1 is characterized in that: The comprehensive sensitivity indicators in step 3 are as follows: In the formula: Ic and Im represent the absolute Pearson correlation coefficient and mutual information, respectively; Let i represent the feature extracted from the sensor signal s during the j-th tool pass, and N represent the total number of tool passes; This represents the average value of feature i extracted from sensor signal s; and These are the actual wear value and the average wear value for the j-th tool pass, respectively; This represents the joint probability distribution between features and wear values. and These are the marginal probability distributions of the feature and wear value, respectively. and The weights for Ic and Im are set to 0.8 and 0.2 respectively.

4. The tool wear monitoring method based on a physically guided gated Mamba network according to claim 1, characterized in that: The sensitive feature set is constructed in step 4 as follows: The comprehensive sensitivity index values ​​calculated in step 3 are sorted from largest to smallest, and the top-k features are selected for model training and testing. To avoid data leakage, the selected feature categories are based only on the training set, that is, the test set selects the same features as the training set.

5. The tool wear monitoring method based on a physically guided gated Mamba network according to claim 1, characterized in that: In step 5, a gated Mamba network is used to extract tool wear-related information from the input features. The process can be represented as follows: S5.1) Input projection: For the l-th level gated Mamba, the time series at time t-1 is... and hidden nodes As input, used to update the state at time t, the corresponding formula can be expressed as: In the formula: This represents the state vector updated at time t. and For weights and biases; S5.2) State aggregation: Layer-normalized aggregation of state vectors at time t and hidden nodes at time point t-1 The polymerization process can be represented as: In the formula: ds represents the hidden node at time t-1. The feature dimension is denoted by j, where j represents the feature of the j-th dimension, and LayerNorm represents the layer normalization process. S5.3) Gating calculation: The gating vector is calculated using state aggregation and used to regulate the signal flow. In the formula: , , These represent the reset gate, update gate, and state gate vector, respectively. and These are the sigmoid activation function and the tanh activation function, respectively. , , and To reset the weights of gates, update gates, and state gates, , ,and For the corresponding bias; S5.4) State space update: By merging hidden nodes and reset door To update the state space vector: In the formula, A and B are the transformation matrices of the hidden node and the reset gate, respectively; This is the state space vector that is dynamically updated during model training. S5.5) Candidate State Update: Subsequently, a linear projection is performed on the state gate vector to construct auxiliary candidate states: In the formula, C represents the state gate vector. Transformation matrix; S5.6) Gated memory update: The gating memory is updated by fusing the state space vector and the state gate candidate vector, which is used to hide nodes at the next time step. S5.7) Output: By integrating and updating gated memory and state vector Update the t-time series: In the formula, D represents the gated memory transformation matrix. For time series t, it is used as the input at time point t+1; For multiple gated Mamba units, their update at time t in layer l can be expressed as: In the formula, L represents the total gated Mamba layer. This represents the mapping at time t in the l-th layer; initially, in the first layer... , which are the initial input features; Composed of all-zero vectors; the final output of the model can be represented as: In the formula, T represents the total time steps.

6. The tool wear monitoring method based on a physically guided gated Mamba network according to claim 1, characterized in that: Step 6 describes the use of two multilayer perceptrons to construct the physical mechanism of tool wear and the output of tool wear monitoring values, which can be specifically described as follows: In the formula, l represents a linear layer. Represents the ReLU activation function. and For data-driven multilayer perceptron weights, This represents an implicit nonlinear mapping. and This represents the time first and second derivatives of the output features of the gated Mamba. and To find the first and second derivatives of the original input features.

7. The tool wear monitoring method based on a physically guided gated Mamba network according to claim 6, characterized in that: The training process of the gated Mamba network is based on historical data, and the model is updated and evaluated during training using the following loss function: In the formula, , , , These represent data-driven loss, physical consistency loss, knowledge loss, and smoothing loss, respectively. and For the corresponding weights.

8. The tool wear monitoring method based on a physically guided gated Mamba network according to claim 7, characterized in that: The data-driven loss, physical consistency loss, knowledge loss, and smoothing loss are described in detail below: In the formula, N represents the total number of tool passes. and Let be the actual wear value and the model-predicted wear value for the i-th time, respectively. This indicates the predicted tool wear rate.

9. A tool wear monitoring system based on a physically guided gated Mamba network according to any one of claims 1 to 8, characterized in that: It includes a multi-source sensor, a processing unit, a workbench (3), a tag value measurement system (4), a multi-channel data acquisition card (8), and a data display module (9); The worktable (3) is used to support the milling workpiece (2); the cutting tool (1) is used to process the milling workpiece (2); the multi-source sensor includes a magnetic force sensor, a vibration sensor and an acoustic emission sensor; the multi-source sensor is attached to the surface of the worktable (3) and the milling workpiece (2); the multi-source sensor is connected to a multi-channel data acquisition card (8); the multi-channel data acquisition card (8) and the data display module (9) are both connected to the processing unit; During the model building phase, the multi-source sensors collect multi-source physical signals in real time during the machining process for feature extraction and model training; the multi-source physical signal acquisition covers the entire life cycle of the tool; the multi-channel data acquisition card (8) realizes the synchronous acquisition and digital conversion of multiple signals, and transmits them to the processing unit for preprocessing and feature extraction; the processing unit runs data preprocessing, feature filtering, gated Mamba network training and verification algorithms; After each milling operation, the tool flank wear is measured by the label value measurement system (4) as a real label for model calibration; the trained tool wear monitoring model is deployed to the processing unit. In the subsequent real-time monitoring stage, the multi-source sensor collects multi-source physical signals during the processing in real time; the multi-channel data acquisition card (8) realizes the synchronous acquisition and digital conversion of multiple signals and transmits them to the processing unit; the processing unit runs the tool wear monitoring model to perform real-time reasoning and output the wear value; the wear value is visualized and output through the data display module (9).

Images