Cutter state monitoring method based on multi-source information fusion

Abstract

The invention discloses a tool state monitoring method based on multi-source information fusion. The method comprises the steps that vibration signals, sound signals, current signals and temperature signals of a tool in different states are collected; preprocessing such as denoising, standardization and resampling is carried out on each modal signal characteristic so as to eliminate interference and unify a data reference; for vibration and sound high-frequency dynamic signals, a time-frequency graph is generated through continuous wavelet transform, local impact features are extracted in combination with a 2D convolutional neural network, and collaborative capture of time-frequency domain global information and time domain local details is achieved; and for current and temperature low-frequency slowly-varying signals, converting one-dimensional time sequence data into a time sequence-feature matrix, and finally inputting fusion features into a full-connection classifier to complete multi-source-domain combined diagnosis. According to the method, the cutter wear state can be reflected more accurately through customized feature extraction and intelligent fusion strategies for different signal physical attributes, and the monitoring robustness and accuracy are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of tool condition monitoring for CNC gear milling machines, and in particular to a tool condition monitoring method based on multi-source information fusion. Background Technology

[0002] With the advancement of industrial automation, CNC milling cutters are playing an increasingly crucial role in manufacturing, as their performance directly determines machining quality and efficiency. However, tool fault diagnosis faces a dual challenge: on the one hand, the cutting process is affected by multiple factors such as material properties, cutting parameters, and environmental conditions, resulting in diverse forms of wear and fault manifestations, increasing the complexity of diagnosis; on the other hand, strong vibrations and noise in the working environment can easily mask the true fault characteristics, greatly hindering signal analysis. Against this backdrop, multi-sensor information fusion (MSIF) technology has become an important means to overcome bottlenecks. Originating in the early 1970s, this technology collects and integrates information from multiple sources and in multiple formats to generate complete and accurate integrated information. It was initially applied in the military field for target tracking and pose detection. With the development of information fusion theory, statistics, and artificial intelligence technologies, it has gradually become a research hotspot in areas such as complex industrial monitoring and equipment fault diagnosis. In machining process monitoring, MSIF effectively overcomes the limitations of single sensors due to its information redundancy, complementarity, and synergy. Compared to single sensors, multi-source information is richer and more reliable, providing theoretical and technical support for analyzing the causes of machining conditions and improving quality levels. Currently, this technology has been applied in milling, turning, drilling, and other machining processes, ensuring stable operation by fusing multimodal signals such as vibration and sound. However, existing research still has many limitations. The depth of heterogeneous fusion of multimodal data is insufficient, with most studies remaining at a shallow fusion stage such as feature stitching. The physical correlation between vibration and current signals is not fully explored, making it difficult to form deep semantic correlations across modes. The model's generalization ability is significantly constrained by the working conditions. When cutting parameters change drastically, the sensor signal distribution is prone to shift, leading to a sharp drop in the accuracy of models trained based on fixed working conditions. There is also a lack of dynamic working condition adaptive mechanisms. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention proposes a tool condition monitoring method based on multi-source information fusion, starting from the principle of multi-sensor information fusion for condition monitoring. This method achieves deep correlation through "modal feature extraction + cross-modal attention fusion".

[0004] This invention is achieved through the following techniques: A tool condition monitoring method based on multi-source information fusion includes the following steps: Step S1: Collect vibration signals, sound signals, current signals and temperature signals of the tool under different conditions using different sensors to provide the raw data basis for subsequent analysis; Step S2: Perform preprocessing operations according to the differences in physical characteristics of different modal signals; Step S3: For high-frequency dynamic signals of vibration and sound, a time-frequency diagram is generated through continuous wavelet transform, preserving the joint time-frequency domain features; Step S4: Use a 2D convolutional neural network (2D CNN) to extract features from the time-frequency graph generated in S3; Step S5: For low-frequency slowly varying signals of current and temperature, construct a "time series-feature matrix" using the sliding window method—calculate statistics such as mean, variance, and rate of rise within the window to convert one-dimensional time series data into structured two-dimensional features; Step S6: The two-dimensional features of current and temperature are modeled using a bidirectional long short-term memory network, i.e., a bidirectional LSTM, to model the forward and backward temporal dependencies, and an attention mechanism is introduced to focus on the trend change points of key wear stages, and to extract trend features that reflect the cumulative changes in load. Step S7: Introduce a cross-modal attention mechanism to assign dynamic weights to high-frequency features and low-frequency features respectively. Learn the importance of each mode under different cutting conditions through a fully connected layer, suppress the weights of noise-interfering modes, and realize adaptive information interaction and fusion of the two features. Step S8: Input the fused feature vector into a fully connected classifier, and output the classification result of the tool wear state through the softmax activation function to complete the joint diagnosis of multi-source domain information and achieve accurate identification of tool state.

[0005] In step S1, the specific steps are as follows: Step S1.1: Build a data acquisition system, design acquisition points, and acquire signals of four different modes of tool status: vibration, sound, temperature, and current from the CNC milling machine tool.

[0006] In step S2, the specific steps are as follows: Step S2.1: Apply wavelet thresholding to the vibration signal to eliminate mechanical interference. Perform wavelet decomposition on the vibration signal, selecting the Symlet8 wavelet basis. conduct N Layer decomposition (usually = ), to obtain approximate coefficients and detail coefficient The decomposition formula is as follows: ; ; in These are the scaling function and the wavelet function, respectively. j The decomposition level is specified. The original signal is decomposed into approximate coefficients using a decomposition formula, reflecting low-frequency trends and detail coefficients, and reflecting high-frequency noise and impact characteristics. Subsequently, the detail coefficients are thresholded using a general threshold: ,in The standard deviation of noise. N The signal length. The coefficients after thresholding. Reconstructed signal This effectively eliminates mechanical interference while retaining key characteristics related to tool wear in the vibration signal. The formula is: ; Step S2.2: Suppress ambient noise by bandpass filtering the sound signal. Determine the dominant frequency range of the sound during tool operation, avoid ambient noise, and use a Butterworth bandpass filter for filtering. Its transfer function is: ; in For the center frequency, For quality factors, The cutoff frequency is used. This filter performs a convolution operation on the audio signal, preserving the effective cutting sound signal within the main frequency range and suppressing environmental noise interference. Step S2.3: Perform Kalman filtering on the current signal to smooth grid fluctuations. Establish the state equation and observation equation as follows: Equations of state: ,in This is process noise.

[0007] Observation equation: ,in To filter out noise, the filtering process includes prediction and updating, ultimately outputting a smoothed current signal sequence. Step S2.4: For possible instantaneous jump abnormal values ​​of the temperature sensor, use 3 The criteria are tested and corrected. Temperature series. T(t) With window W Calculate local mean and standard deviation If the temperature at a certain moment T(t) Satisfy | T (t) - |>3 If the value is identified as an outlier, it will be corrected. The correction formula is: ; in k The window radius; Step S2.5: Perform z-score normalization on the preprocessed vibration, sound, current, and temperature signals to eliminate dimensional differences. ,in, For the original signal i Data points, and These are the mean and standard deviation of the signal, respectively. The values ​​are standardized. Resampling is achieved through timestamp alignment to unify the data sampling rate. Downsampling is used for vibration and sound signals with high sampling rates, while linear interpolation is used to complete current and temperature signals with low sampling rates. Time-domain alignment of multimodal signals is achieved through timestamp matching to ensure that features at the same time can be correlated for analysis.

[0008] In step S3, the specific steps are as follows: Step S3.1: Perform a Continuous Wavelet Transform (CWT) on the preprocessed vibration and sound signals to uncover their implicit time-frequency localization features. As high-frequency dynamic signals, the instantaneous impact of vibration signals, such as the intermittent contact between the tool and the workpiece, and the frequency drift of sound signals, such as the shift in the dominant frequency of cutting noise due to increased wear, both exhibit significant time-varying characteristics. The CWT, by convolving the signal with a series of wavelet basis functions generated by scaling and translating the mother wavelet, can achieve refined decomposition of the signal at different time and frequency scales while maintaining good time-domain and frequency-domain localization capabilities. The formula for the Continuous Wavelet Transform (CWT) is: ; in: It is the preprocessed signal. It is the complex conjugate form of the wavelet function, defined as: , It is the scale factor (which controls the frequency resolution). It is the translation factor (which controls the time resolution). It is the mother wavelet function. Result Reflecting signals at different scales and time The magnitude of the component corresponds to the energy intensity at that location in the time-frequency domain, thereby constructing a time-frequency diagram that can intuitively reflect the time-frequency distribution characteristics of the signal, providing structured time-frequency information support for subsequent 2D CNN extraction of local impact features; Step S3.2: Use the `imshow` function to plot the time-frequency graph, displaying the transform coefficients as a two-dimensional image. The x-axis represents time, and the y-axis represents frequency. The plotting method is as follows: ; The extracted energy features, statistical features, spectral entropy, etc., are combined to form a feature vector. ; Step S3.3: Try different scale ranges and color mappings. Set the scale range according to the signal's main frequency characteristics to ensure that the target frequency band is in the center region of the image. Adjust the color threshold using the clim function to suppress background noise interference and ensure that the frequency features are clearly visible. Step S3.4: When the resolution requirement is high, increasing the number of scales can improve the frequency resolution, but the amount of computation will also increase. The calculation should be dynamically adjusted according to the diagnostic accuracy requirements and real-time requirements.

[0009] In step S4, the specific steps are as follows: Step S4.1: Network input layer design. The time-frequency map is standardized to the [0,1] interval and then input into the network. The input size is set to 256×256×1 according to the resolution of the time-frequency map. To preserve the spatial correlation of the time-frequency domain, the input layer does not perform compression processing and directly transmits the original time-frequency distribution features. The high-frequency impact region is represented by local high gray value patches, and the low-frequency trend region is represented by continuous texture, providing clear feature anchors for subsequent convolution operations. Step S4.2, shallow feature extraction, adopts a stacked structure of "convolution + batch normalization + activation + pooling" to focus on basic visual features such as edges and textures in the time-frequency map, corresponding to the primary signal pattern of the tool state: Convolutional Block 1: A 3×3 convolutional layer with 64 filters and a stride of 1. Padding is set to "same" to maintain size. This layer calculates the weighted sum of local pixels through a sliding window to extract the edge contours of high-frequency impacts in the time-frequency map. After accelerating convergence with Batch Normalization, the nonlinear expression is enhanced by the ReLU activation function, followed by a 2×2 max-pooling layer with a stride of 2 to compress the feature map to 128×128×64, preserving significant edge features. The second convolutional block consists of a 3×3 convolutional layer with 128 filters, which further captures the texture features formed by edge combinations. After batch normalization and ReLU activation, it is compressed to 64×64×128 through 2×2 max pooling. The third convolutional block consists of a 3×3 convolutional layer with 256 filters, which extracts more complex local patterns, corresponding to the evolution trend of frequency components during tool wear; the feature map size after pooling is 32×32×256. Step S4.3: After obtaining basic visual features through shallow feature extraction, a deep network is used to further mine abstract semantic features strongly related to the tool state. By expanding the receptive field of the filter and increasing the number of channels, higher-order features strongly related to the wear state in the time-frequency graph are mined. The fourth convolutional block employs a 5×5 convolutional layer with 512 filters, a stride of 1, and "same" padding to cover a wider receptive field across the time-frequency region, capturing cross-time-frequency correlation features. After batch normalization and ReLU activation, global average pooling (GAP) is used to compress the spatial dimension, preserving key semantic information in the 32×32×512 feature map and outputting a 1×1×512 feature. The 5th convolutional block introduces a 1×1 convolutional layer with 256 filters to compress the channel dimension and reduce redundant features. At the same time, it suppresses overfitting by using Dropout with rate=0.5. The final output is a 256-dimensional feature vector, which contains high-order patterns in the time-frequency graph that are directly related to the tool wear state. Step S4.4: Targeted design of feature extraction. For the vibration signal time-frequency map: focus on capturing the shape changes of the high-frequency impact area, extracting patch edges through shallow 3×3 convolutions and capturing the spatial correlation of patch distribution through deep 5×5 convolutions. For the sound signal time-frequency map: focus on the texture continuity of the low-frequency band, balancing local details and global trends through multi-scale convolutions. An attention mechanism is introduced to complement the time-frequency spatial features extracted by 2D CNN and the temporal local impact features extracted by 1D CNN, jointly constructing a multi-dimensional feature system of "temporal instantaneous features - time-frequency spatial patterns" to improve the ability to discriminate tool wear states.

[0010] In step S5, the specific steps are as follows: Step S5.1: Input the preprocessed current signal and temperature signal ; For each signal sequence, a fixed-length sliding window is used for segmentation. The window length is... L The settings are based on signal characteristics and diagnostic requirements, and are combined with the sampling frequency. fs To convert to points, use the following formula: ; To control the time resolution, the sliding step size S Pick L 20% of: ; Step S5.2: Divide each signal into several non-overlapping subsequences, as shown in the following formula: ; Total Each window calculates basic statistics, trend features, morphological features, and other statistics, converting one-dimensional time series data into structured two-dimensional features.

[0011]

[0012] In step S6, the specific steps are as follows: Step S6.1: Input the "time series-feature matrix" constructed with sliding windows, with dimensions T×F, where T is the number of windows and F is the number of statistical features for each window; Step S6.2: Model using a bidirectional LSTM. Input the matrix into a bidirectional LSTM network. Input bidirectional LSTM network: ; ; ; Output: The sequence of hidden states at all time steps ,in H This represents the number of hidden units in a single-layer LSTM. Step S6.3: Introduce an attention mechanism to assign different weights to the hidden state sequence and focus on the key time window where wear and tear changes are obvious.

[0013] Calculate attention weights: ; ; Learnable parameters It is a randomly initialized context vector; Weighted context vector: ; Output feature vectors focusing on key wear stages ; Step S6.4: Finally, output the features and context vector. c This represents the global trend of current or temperature modes, reflecting the evolution of the signal over time, with a focus on key turning points.

[0014] In step S7, the specific steps are as follows: Step S7.1: Denote the time-frequency graph features of the high-frequency modes extracted in S4 as... The feature vector of the low-frequency mode extracted from S6 is denoted as... ; Step S7.2: Map the two types of features to the same dimension d using a fully connected layer: ; right and Each is assigned a learnable weight This indicates that the importance of each mode is reduced under the current operating condition, achieving adaptive mode fusion and weakening the interference of modes with low signal-to-noise ratio on the discrimination results. The formula is as follows: ; ; Step S7.3: Input the fused features from S7.2 into the fully connected layer for further information integration. The adaptively fused feature vector is: , where K is the dimension of the fusion feature.

[0015] In step S8, the fused feature vector is input into a fully connected classifier, and the classification result of the tool wear state is output through the softmax activation function to complete the joint diagnosis of multi-source domain information and achieve accurate identification of tool state.

[0016] The present invention has the following benefits: To address the challenge of diagnosing multimodal heterogeneous signals, this invention innovatively employs a "divide and conquer" differentiated processing and cross-modal adaptive fusion strategy to construct an end-to-end multi-source information joint diagnostic framework. In the feature extraction stage, signals are processed differently based on their physical characteristics: high-frequency signals (vibration, sound) are transformed into time-frequency maps through continuous wavelet transform, and spatial-spectral patterns are extracted using a 2D CNN; low-frequency signals (current, temperature) are transformed into two-dimensional data through sliding window statistical features, and then a BiLSTM+attention mechanism is used to capture long-term time-series dependencies and key abrupt changes, accurately extracting load accumulation trend features and avoiding information loss caused by single-model processing. The cross-modal attention mechanism dynamically allocates weights for high-frequency and low-frequency features, suppresses noise interference modes, and achieves deep interactive fusion of heterogeneous information, solving the key challenge of multi-source information fusion. This invention integrates multi-dimensional complementary information, improving diagnostic accuracy and robustness, enhancing the ability to identify key wear points, adapting to complex working conditions, providing support for predictive maintenance, reducing unexpected downtime, improving processing quality, reducing costs, and promoting the intelligent upgrading of the manufacturing process. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the framework of the present invention; Figure 2 This is a schematic diagram of the data acquisition points. Figure 3 This is a schematic diagram of the data acquisition system. Figure 4 This is a schematic diagram of a 2D CNN convolutional network. Figure 5 This is an accuracy graph of the present invention on the CNC tool status dataset of the Industrial University. Figure 6 This is a curve showing the loss value of the present invention on the CNC tool status dataset of the Industrial University. Figure 7 This is a confusion matrix diagram showing the recognition accuracy of each category on the test set for this invention; Figure 8 This is a t-SNE visualization of the dataset features extracted by this invention. Please note the numbers 1234567 in the figure. Figure 9 This is a t-SNE visualization feature map of the dataset feature extraction method of this invention; numbers 1234567 in the figure can be noted.

[0018] Notes: 1: Normal knife; 2: Slight chipping; 3: Moderate chipping; 4: Severe chipping; 5: Slight mechanical fracture; 6: Moderate mechanical fracture; 7: Severe mechanical fracture Detailed Implementation

[0019] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0020] This invention provides a tool condition monitoring method based on multi-source signal fusion, such as... Figure 1 As shown, it includes the following steps: S1: Collect vibration signals, sound signals, current signals and temperature signals of the cutting tool under different conditions through different sensors; S2: Perform preprocessing operations according to the differences in physical characteristics of different modal signals; S3: For high-frequency dynamic signals of vibration and sound, a time-frequency diagram is generated through continuous wavelet transform; S4: Use a 2D convolutional neural network to extract features from the time-frequency graph generated in S3; S5: For low-frequency slowly varying signals of current and temperature, a "time series-feature matrix" is constructed using the sliding window method. Within the window, statistics such as mean, variance, and rate of rise are calculated, converting one-dimensional time series data into structured two-dimensional features. S6: The two-dimensional features of current and temperature are modeled using a bidirectional long short-term memory network, i.e., a bidirectional LSTM, which is used to model the forward and backward temporal dependencies. An attention mechanism is introduced to focus on the trend change points of key wear stages and extract trend features that reflect the cumulative changes in load. S7: Introduces a cross-modal attention mechanism to assign dynamic weights to high-frequency and low-frequency features respectively. Through a fully connected layer, it learns the importance of each mode under different cutting conditions, suppresses the weights of noise-affected modes, and realizes adaptive information interaction and fusion of the two features. S8: Input the fused feature vector into the fully connected classifier, and output the classification result of tool wear state through the softmax activation function to complete the joint diagnosis of multi-source domain information and achieve accurate identification of tool state.

[0021] Based on the above scheme, this embodiment applies the multi-source signal fusion tool condition monitoring method as follows: The experimental data for this example comes from the CNC tool condition machining process test dataset from the University of Technology of China. The experimental platform is as follows: Figure 2 As shown. The experimental cutter head has an outer diameter of 426mm, a module of 25, and the cutter material is cemented carbide, model Z98a20. Milling experiments were conducted using cutters in different states. The construction principle of the data acquisition box is as follows. Figure 3 As shown, four sensors are used for simultaneous measurement. The sampling frequency for vibration and sound is set to 2000Hz, the sampling frequency for current signal is set to 1000Hz, and the sampling frequency for temperature signal is set to 100Hz; the sampling duration is 50s.

[0022] Wavelet thresholding is used to denoise the vibration signal to eliminate mechanical interference. Wavelet decomposition is performed on the vibration signal, selecting the Symlet8 wavelet basis. conduct N =5-level decomposition, obtaining approximate coefficients and detail coefficient The decomposition formula is as follows: ; ; Thresholding is applied to the detail coefficients using a general threshold: Coefficients after thresholding Reconstructed signal This effectively eliminates mechanical interference while preserving key characteristics of the vibration signal related to tool wear.

[0023] Ambient noise is suppressed by bandpass filtering of the sound signal. The dominant frequency range of the sound during tool operation is determined, and ambient noise is avoided. A Butterworth bandpass filter is selected for filtering, with the following transfer function: The filter performs convolution operations on the sound signal, retaining the effective cutting sound signal within the main frequency range and suppressing environmental noise interference. Kalman filtering is applied to the current signal to smooth power grid fluctuations. State equations and observation equations are established. , The filtering process includes prediction and updating, and finally outputs a smoothed current signal sequence. Temperature sensors may experience transient jumps in values; 3 The criteria are tested and corrected. Temperature series. T(t) With window W Calculate local mean and standard deviation If the temperature at a certain moment T(t) Satisfy | T(t) - |>3 If the value is identified as an outlier, it will be corrected. The correction formula is: ; The processed vibration, sound, current, and temperature signals are respectively z-score normalized to eliminate dimensional differences. ,in, For the original signal i Data points, and These are the mean and standard deviation of the signal, respectively. Resampling is achieved through timestamp alignment to unify the data sampling rate. Downsampling is used for vibration and sound signals with high sampling rates, while linear interpolation is used to complete current and temperature signals with low sampling rates. Time-domain alignment of multimodal signals is achieved through timestamp matching to ensure that features at the same time can be correlated for analysis.

[0024] Continuous wavelet transform is performed on the preprocessed vibration and sound signals to uncover their implicit time-frequency localization features: ; result Reflecting signals at different scales and time The components on the time-frequency domain are represented by their amplitudes, which correspond to the energy intensity at that location in the time-frequency domain. This constructs a time-frequency graph that intuitively reflects the time-frequency distribution characteristics of the signal, providing structured time-frequency information support for subsequent 2D CNN extraction of local impulse features. The `imshow` function is used to plot the time-frequency graph, displaying the transform coefficients as a two-dimensional image, with the x-axis representing time and the y-axis representing frequency. The plotting method is as follows: ; The extracted energy features, statistical features, spectral entropy, etc., are combined to form a feature vector. The color threshold is adjusted using the clim function to suppress background noise interference and ensure clear visibility of frequency features. When high resolution is required, increasing the number of scales can improve frequency resolution, but it will also increase the computational load. The adjustment should be made dynamically according to the diagnostic accuracy and real-time requirements.

[0025] The time-frequency map is normalized to the [0,1] interval and then input into the network. The input size is set to 256×256×1 according to the resolution of the time-frequency map. To preserve the spatial correlation in the time-frequency domain, the input layer does not perform compression processing and directly transmits the original time-frequency distribution features. High-frequency impact regions are represented as local high-grayscale patches, and low-frequency trend regions are represented as continuous textures, providing clear feature anchors for subsequent convolutional operations. The convolutional network structure is as follows: Figure 4 As shown, a stacked structure of "convolution + batch normalization + activation + pooling" is used to focus on basic visual features such as edges and textures in the time-frequency map, corresponding to the primary signal pattern of the tool state: Convolutional Block 1: A 3×3 convolutional layer with 64 filters and a stride of 1. Padding is set to "same" to maintain size. This layer calculates the weighted sum of local pixels through a sliding window to extract the edge contours of high-frequency impacts in the time-frequency map. After accelerating convergence with Batch Normalization, the nonlinear expression is enhanced by the ReLU activation function, followed by a 2×2 max-pooling layer with a stride of 2 to compress the feature map to 128×128×64, preserving significant edge features. The second convolutional block consists of a 3×3 convolutional layer with 128 filters, which further captures the texture features formed by edge combinations. After batch normalization and ReLU activation, it is compressed to 64×64×128 through 2×2 max pooling. The third convolutional block consists of a 3×3 convolutional layer with 256 filters, which extracts more complex local patterns, corresponding to the evolution trend of frequency components during tool wear; the feature map size after pooling is 32×32×256. After obtaining basic visual features through shallow feature extraction, a deep network is used to further mine abstract semantic features strongly related to the tool state. By expanding the receptive field of the filter and increasing the number of channels, higher-order features strongly related to the wear state in the time-frequency graph are mined. The fourth convolutional block employs a 5×5 convolutional layer with 512 filters, a stride of 1, and "same" padding to cover a wider receptive field across the time-frequency region, capturing cross-time-frequency correlation features. After batch normalization and ReLU activation, global average pooling (GAP) is used to compress the spatial dimension, preserving key semantic information in the 32×32×512 feature map and outputting a 1×1×512 feature. The 5th convolutional block introduces a 1×1 convolutional layer with 256 filters to compress the channel dimension and reduce redundant features. At the same time, it suppresses overfitting by using Dropout with rate=0.5. The final output is a 256-dimensional feature vector, which contains high-order patterns in the time-frequency graph that are directly related to the tool wear state. For vibration signal time-frequency maps: the focus is on capturing shape changes in high-frequency impact areas. Shallow 3×3 convolutions are used to extract patch edges, while deep 5×5 convolutions capture the spatial correlation of patch distribution. For sound signal time-frequency maps: the focus is on texture continuity in the low-frequency band, using multi-scale convolutions to balance local details and global trends. An attention mechanism is introduced to complement the time-frequency spatial features extracted by 2D CNNs and the temporal local impact features extracted by 1D CNNs, jointly constructing a multi-dimensional feature system of "temporal instantaneous features - time-frequency spatial patterns," thereby improving the ability to discriminate tool wear states.

[0026] Input preprocessed current signal and temperature signal For each signal sequence, a fixed-length sliding window is used for segmentation. The window length is... L The settings are based on signal characteristics and diagnostic requirements, and are combined with the sampling frequency. fs To convert to points, use the following formula: ; To control the time resolution, the sliding step size S Pick L 20% of: Each signal is divided into several non-overlapping subsequences, as shown in the following formula: ; Total Each window calculates basic statistics, trend characteristics, morphological characteristics, and other statistical measures. One-dimensional time-series data is transformed into structured two-dimensional features. The input is a "time-feature matrix" constructed using sliding windows, with dimensions T×F, where T is the number of windows and F is the number of statistical features per window. This matrix is ​​then modeled using a bidirectional LSTM network. Input bidirectional LSTM network: ; ; ; Hidden state sequence at all time steps An attention mechanism is introduced to assign different weights to the hidden state sequence, focusing on the key time window where wear and tear changes are obvious.

[0027] Calculate attention weights: ; ; Learnable parameters It is a randomly initialized context vector; Weighted context vector: ; Output feature vectors focusing on key wear stages Context vector c This represents the global trend of current or temperature modes, reflecting the evolution of the signal over time, with a focus on key turning points.

[0028] The time-frequency plot features of the high-frequency modes extracted from S4 are denoted as... The feature vector of the low-frequency mode extracted from S6 is denoted as... The two types of features are mapped to the same dimension d through a fully connected layer. ; right and Each is assigned a learnable weight This indicates that the importance of each mode is reduced under the current operating condition, achieving adaptive mode fusion and weakening the interference of modes with low signal-to-noise ratio on the discrimination results. The formula is as follows: ; ; The aforementioned fused features are then input into the fully connected layer for further information integration. The adaptively fused feature vector is: , where K is the dimension of the fusion feature.

[0029] The fused feature vector is input into a fully connected classifier, and the classification result of tool wear state is output through the softmax activation function, thus completing the joint diagnosis of multi-source domain information and realizing accurate identification of tool state.

Claims

1. A tool condition monitoring method based on multi-source information fusion, characterized in that, Includes the following steps: S1: Vibration signals, sound signals, current signals and temperature signals of the tool under different conditions are collected by different sensors to provide the raw data basis for subsequent analysis; S2: Perform preprocessing operations according to the differences in physical characteristics of different modal signals; S3: For high-frequency dynamic signals of vibration and sound, a time-frequency diagram is generated through continuous wavelet transform, preserving the joint time-frequency domain characteristics; S4: Feature extraction is performed on the time-frequency graph generated in S3 using a 2D convolutional neural network (2D CNN); S5: For low-frequency, slowly varying signals of current and temperature, a "time-feature matrix" is constructed using the sliding window method. Within the window, statistics such as mean, variance, and rate of rise are calculated, converting one-dimensional time-series data into structured two-dimensional features. S6: The two-dimensional features of current and temperature are modeled using a bidirectional long short-term memory network, i.e., a bidirectional LSTM, which is used to model the forward and backward temporal dependencies. An attention mechanism is introduced to focus on the trend change points of key wear stages and extract trend features that reflect the cumulative changes in load. S7: Introduces a cross-modal attention mechanism to assign dynamic weights to high-frequency and low-frequency features respectively. Through a fully connected layer, it learns the importance of each mode under different cutting conditions, suppresses the weights of noise-affected modes, and realizes adaptive information interaction and fusion of the two features. S8: Input the fused feature vector into the fully connected classifier, and output the classification result of tool wear state through the softmax activation function to complete the joint diagnosis of multi-source domain information and achieve accurate identification of tool state.

2. The tool condition monitoring method based on multi-source information fusion as described in claim 1, characterized in that: In step S2, preprocessing operations are performed separately for the differences in physical characteristics of different modal signals, mainly including the following steps: S2.1: Wavelet threshold denoising is applied to the vibration signal to eliminate mechanical interference. Wavelet decomposition is performed on the vibration signal, selecting the Symlet8 wavelet basis. conduct N Layer decomposition (usually) ), to obtain approximate coefficients and detail coefficient The decomposition formula is as follows: ； ； in These are the scaling function and the wavelet function, respectively. j The decomposition level is determined by the decomposition formula, which decomposes the original signal into approximate coefficients that reflect low-frequency trends and detail coefficients, as well as high-frequency noise and impulse characteristics. The detail coefficients are then thresholded using a general threshold: ,in The standard deviation of noise. N Signal length; coefficients after thresholding. Reconstructed signal This effectively eliminates mechanical interference while retaining key characteristics related to tool wear in the vibration signal. The formula is: ； S2.2: Suppress ambient noise by bandpass filtering the sound signal. Determine the dominant frequency range of the sound during tool operation, avoid ambient noise, and use a Butterworth bandpass filter for filtering. Its transfer function is: ； in For the center frequency, For quality factors, The cutoff frequency is used. This filter performs a convolution operation on the audio signal, preserving the effective cutting sound signal within the main frequency range and suppressing environmental noise interference. S2.3: Perform Kalman filtering on the current signal to smooth power grid fluctuations. Establish the state equation and observation equation as follows: Equations of state: ,in This is process noise; Observation equation: ,in To filter out noise, the filtering process includes prediction and updating, ultimately outputting a smoothed current signal sequence. S2.4: For the transient jump values ​​that may occur in the temperature sensor, adopt... The criteria are tested and corrected; the temperature series is processed. T(t) With window W Calculate local mean and standard deviation ; If the temperature at a certain moment T(t) Satisfy | T(t) - |>3 If the value is identified as an outlier, it will be corrected. The correction formula is: ； in k The radius of the window; S2.5: Perform z-score normalization on the preprocessed vibration, sound, current, and temperature signals to eliminate dimensional differences. ,in, For the original signal i Data points, and These are the mean and standard deviation of the signal, respectively. The value is the standardized value; Resampling is achieved through timestamp alignment to unify the data sampling rate. For vibration and sound signals with high sampling rates, downsampling is used; for current and temperature signals with low sampling rates, linear interpolation is used to complete the data. Time-domain alignment of multimodal signals is achieved through timestamp matching to ensure that features at the same time can be correlated and analyzed.

3. The tool condition monitoring method based on multi-source information fusion according to claim 1, characterized in that: In step S3, for high-frequency dynamic signals of vibration and sound, a time-frequency diagram is generated through continuous wavelet transform, preserving the joint time-domain and frequency-domain features. This mainly includes the following steps: S3.1: A Continuous Wavelet Transform (CWT) is applied to the preprocessed vibration and sound signals to uncover their implicit time-frequency localization features. As high-frequency dynamic signals, the instantaneous impact of vibration signals, such as the intermittent contact between the tool and the workpiece, and the frequency drift of sound signals, such as the shift in the dominant frequency of cutting noise caused by increased wear, both exhibit significant time-varying characteristics. The CWT, by convolving the signal with a series of wavelet basis functions generated by scaling and translating the mother wavelet, can achieve refined decomposition of the signal at different time and frequency scales while maintaining good time-domain and frequency-domain localization capabilities. The formula for the Continuous Wavelet Transform (CWT) is: ； in: It is the preprocessed signal. It is the complex conjugate form of the wavelet function, defined as: , It is the scale factor (which controls the frequency resolution). It is the translation factor (which controls the time resolution). It is the mother wavelet function. Result Reflecting signals at different scales and time The magnitude of the component corresponds to the energy intensity at that location in the time-frequency domain, thereby constructing a time-frequency diagram that can intuitively reflect the time-frequency distribution characteristics of the signal, providing structured time-frequency information support for subsequent 2D CNN extraction of local impact features; S3.2: Use the `imshow` function to plot a time-frequency graph, displaying the transform coefficients as a two-dimensional image. The x-axis represents time, and the y-axis represents frequency. The plotting method is as follows: ； The extracted energy features, statistical features, spectral entropy, etc., are combined to form a feature vector. ; S3.3: Try different scale ranges and color mappings, set the scale range according to the signal's main frequency characteristics, and ensure that the target frequency band is in the center region of the image; adjust the color threshold through the clim function to suppress background noise interference and ensure that the frequency features are clearly visible; S3.4: When high resolution is required, increasing the number of scales can improve frequency resolution, but it will also increase the computational load. The calculation should be dynamically adjusted according to the diagnostic accuracy and real-time requirements.

4. The tool condition monitoring method based on multi-source information fusion according to claim 1, characterized in that: In step S4, a 2D convolutional neural network (2D CNN) is used to extract features from the time-frequency graph generated in S3. This mainly includes the following steps: S4.1: Network input layer design. The time-frequency map is normalized to the [0,1] interval before being input into the network. The input size is set to 256×256×1 according to the resolution of the time-frequency map. In order to preserve the spatial correlation of the time-frequency domain, the input layer does not perform compression processing and directly transmits the original time-frequency distribution features. The high-frequency impact region is represented by local high gray value patches, and the low-frequency trend region is represented by continuous texture, providing clear feature anchors for subsequent convolution operations. S4.2: Shallow feature extraction, employing a stacked structure of "convolution + batch normalization + activation + pooling", focuses on basic visual features such as edges and textures in the time-frequency map, corresponding to the primary signal patterns of the tool state: The first convolutional block consists of a 3×3 convolutional layer with 64 filters and a stride of 1. Padding is set to "same" to maintain size. This layer calculates the weighted sum of local pixels through a sliding window to extract the edge contours of high-frequency impacts in the time-frequency map. After accelerating convergence through BatchNormalization, the nonlinear expression is enhanced by the ReLU activation function, followed by a 2×2 max-pooling layer with a stride of 2 to compress the feature map to 128×128×64, preserving significant edge features. The second convolutional block consists of a 3×3 convolutional layer with 128 filters, which further captures the texture features formed by edge combinations. After batch normalization and ReLU activation, it is compressed to 64×64×128 by 2×2 max pooling. The third convolutional block consists of a 3×3 convolutional layer with 256 filters, extracting more complex local patterns that correspond to the evolution trend of frequency components during tool wear; the feature map size after pooling is 32×32×256. S4.3: After obtaining basic visual features through shallow feature extraction, a deep network is used to further mine abstract semantic features strongly related to the tool state. By expanding the receptive field of the filter and increasing the number of channels, higher-order features strongly related to the wear state in the time-frequency graph are mined. The fourth convolutional block uses a 5×5 convolutional layer with 512 filters, a stride of 1, and "same" padding to cover a wider receptive field in the time-frequency region and capture cross-time-frequency correlation features. After batch normalization and ReLU activation, global average pooling (GAP) is used to compress the spatial dimension, retaining key semantic information in the 32×32×512 feature map and outputting a 1×1×512 feature. The 5th convolutional block introduces a 1×1 convolutional layer with 256 filters to compress the channel dimension and reduce redundant features. At the same time, it suppresses overfitting by using Dropout with rate=0.

5. The final output is a 256-dimensional feature vector, which contains high-order patterns in the time-frequency graph that are directly related to the tool wear state. S4.4: Targeted design for feature extraction. For vibration signal time-frequency maps: focus on capturing shape changes in high-frequency impact areas, extracting patch edges through shallow 3×3 convolutions and capturing the spatial correlation of patch distribution through deep 5×5 convolutions. For sound signal time-frequency maps: focus on texture continuity in the low-frequency band, balancing local details and global trends through multi-scale convolutions. An attention mechanism is introduced to complement the time-frequency spatial features extracted by 2D CNN and the temporal local impact features extracted by 1D CNN, jointly constructing a multi-dimensional feature system of "temporal instantaneous features - time-frequency spatial patterns," improving the ability to discriminate tool wear states.

5. The tool condition monitoring method based on multi-source information fusion according to claim 1, characterized in that: In step S5, for low-frequency, slowly varying signals of current and temperature, a "time-feature matrix" is constructed using the sliding window method—statistics such as mean, variance, and rate of rise are calculated within the window to convert one-dimensional time-series data into structured two-dimensional features. This mainly includes the following steps: S5.1: Input the pre-processed current signal and temperature signal For each signal sequence, a fixed-length sliding window is used for segmentation. The window length is... L The settings are based on signal characteristics and diagnostic requirements, and are combined with the sampling frequency. fs To convert to points, use the following formula: ； To control the time resolution, the sliding step size S Pick L 20% of: ; S5.2: Divide each signal into several non-overlapping subsequences, as shown in the following formula: ； Total Each window calculates basic statistics, trend characteristics, morphological characteristics, and other statistical measures within it, transforming one-dimensional time series data into structured two-dimensional features. 。 6. The tool condition monitoring method based on multi-source information fusion according to claim 1, characterized in that: In step S6, the two-dimensional features of current and temperature are modeled using a bidirectional long short-term memory network, i.e., a bidirectional LSTM, to model the forward and backward temporal dependencies. An attention mechanism is introduced to focus on the trend abrupt change points of key wear stages, and trend features reflecting cumulative load changes are extracted. This mainly includes the following steps: S6.1: Input a "time series-feature matrix" constructed with sliding windows, with dimensions T×F, where T is the number of windows and F is the number of statistical features for each window; S6.2: Model using a bidirectional LSTM. Input this matrix into a bidirectional LSTM network, and... Input bidirectional LSTM network: ； ； ； Output: The sequence of hidden states at all time steps ,in H This represents the number of hidden units in a single-layer LSTM. S6.3: Introduce an attention mechanism to assign different weights to the hidden state sequence, focusing on the key time window where wear and tear changes are obvious; Calculate attention weights: ； ； Learnable parameters It is a randomly initialized context vector; Weighted context vector: ； Output feature vectors focusing on key wear stages ; S6.4: Final output features, context vector c This represents the global trend of current or temperature modes, reflecting the evolution of the signal over time, with a focus on key turning points.

7. The tool condition monitoring method based on multi-source information fusion according to claim 1, characterized in that: In step S7, a cross-modal attention mechanism is introduced to assign dynamic weights to high-frequency and low-frequency features respectively. The importance of each mode under different cutting conditions is learned through a fully connected layer, and the weights of noise-affected modes are suppressed to achieve adaptive information interaction and fusion of the two features. The main steps include: S7.1: The time-frequency plot features of the high-frequency modes extracted in S4 are denoted as... The feature vector of the low-frequency mode extracted from S6 is denoted as... ; S7.2: Map the two types of features to the same dimension d using a fully connected layer: ； right and Each is assigned a learnable weight This indicates that the importance of each mode is reduced under the current operating condition, achieving adaptive mode fusion and weakening the interference of modes with low signal-to-noise ratio on the discrimination results. The formula is as follows: ； ； S7.3: Input the features fused from S7.2 into the fully connected layer to further integrate the information. The adaptively fused feature vector is: , where K is the dimension of the fusion feature.

Images