A high-quality fusion method for multi-domain features of cutting signals and performance evaluation of fused features

Abstract

The present invention discloses a method for high-quality fusion of multi-domain features of cutting signals and evaluation of fusion feature performance, which relates to the technical field of mechanical processing process monitoring. The method comprises calculating the Pearson correlation coefficient between each feature vector and the identification target vector based on the cutting signal feature set and the identification target vector to obtain a multi-domain correlation coefficient set; calculating the fusion feature component corresponding to each feature vector based on the automatic superposition function to obtain a multi-domain fusion feature component set; calculating the average vector of each fusion feature component matrix in the multi-domain fusion feature component set to obtain a multi-domain fusion matrix; calculating the Pearson correlation coefficient between each dimension of the multi-domain fusion matrix and the identification target vector to obtain a fusion correlation vector between the multi-domain fusion matrix and the identification target vector; and outputting the final multi-domain fusion matrix after calculation. The present invention can improve the performance of fusion features and, by establishing performance evaluation indicators, realizes an evaluation method that can pre-judge the performance of fused features before training a recognition model.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical machining process monitoring, and in particular to a method for high-quality fusion of multi-domain features of cutting signals and fusion feature performance evaluation. Background Art

[0002] The extraction and fusion of sensor information features are key technologies for determining the accuracy of online monitoring and identification of cutting processes, and have been a hot topic of research in the field of intelligent cutting process monitoring in recent years. Research has shown that multi-domain feature fusion, encompassing time, frequency, and time-frequency domain information, is an important means of improving cutting process monitoring accuracy. Consequently, multi-domain feature fusion of cutting signals is widely used in monitoring precision and ultra-precision machining processes.

[0003] CN110153801B discloses a tool wear identification method based on multi-feature fusion. It extracts the time domain, frequency domain, and entropy features of the cutting force signal and the cutting vibration signal to form a joint multi-feature vector, and fuses the multi-domain features based on singular value decomposition to identify the tool wear status. CN111644900B discloses a real-time tool breakage monitoring method based on spindle vibration feature fusion. Based on the time domain and frequency domain characteristics of the milling force signal, it achieves a precise definition of the tool breakage characteristics by fusing the frequency domain amplitude of the vibration displacement response in the feed direction and the perpendicular feed direction.

[0004] However, existing feature fusion methods mostly fuse multi-domain features based on feature screening or data dimensionality reduction. Feature screening often relies on manual experience, and the screening results are subject to great uncertainty, while data dimensionality reduction requires changing the original data structure. The resulting fused features have poor performance, making it difficult to fit the monitoring target, and requiring a high level of complexity in the recognition model. In addition, although there are currently feasible screening criteria for multiple features before fusion, there is still a lack of accurate evaluation methods for the performance of the fused features. This results in the need to repeatedly train the recognition model using different feature fusion methods and compare the recognition accuracy of different fusion methods to determine the pros and cons of the feature fusion method. This evaluation method, which relies on model training, consumes a lot of time and computing resources. The inefficiency of fusion methods and the difficulty in pre-evaluating the performance of fused features have become key bottlenecks restricting the application and development of cutting process monitoring technology. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a high-quality fusion method of multi-domain features of cutting signals and a fusion feature performance evaluation method to solve the problems of large information loss in the fusion process, poor fusion feature performance, and difficulty in pre-evaluation of fusion feature performance in the existing cutting signal multi-feature fusion method. The multi-domain feature high-quality fusion method can improve the fusion feature performance capability, and at the same time, by establishing performance evaluation indicators, an evaluation method is realized that can pre-judge the performance of the fused features before training the recognition model.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a high-quality fusion method for multi-domain features of cutting signals, comprising:

[0008] According to the cutting signal feature set and the identification target vector, the Pearson correlation coefficient between each feature vector and the identification target vector is calculated to obtain a multi-domain correlation coefficient set;

[0009] Calculate the fusion feature component corresponding to each feature vector according to the automatic superposition function to obtain a multi-domain fusion feature component set;

[0010] The average vector of each fused feature component matrix in the multi-domain fusion feature component set is calculated horizontally to obtain the corresponding intra-domain fusion vector. The intra-domain fusion vectors are sequentially spliced ​​to obtain the multi-domain fusion matrix. The intra-domain fusion vector is obtained by superimposing the representation of each feature vector to automatically eliminate weakly correlated feature components in the original features.

[0011] The Pearson correlation coefficient of each dimension of the multi-domain fusion matrix and the identification target vector is calculated respectively to obtain the fusion correlation vector of the multi-domain fusion matrix and the identification target vector; the norm is calculated by fusion correlation vector, and the final multi-domain fusion matrix is ​​output.

[0012] As a further implementation, the automatic superposition function is:

[0013]

[0014] Among them, β represents the fusion level, c i represents the Pearson correlation coefficient, f i represents the feature vector.

[0015] As a further implementation method, the degree of elimination of weakly correlated feature components in the original feature data is adjusted by modifying the fusion level.

[0016] As a further implementation method, the value of the fusion level β is determined by taking the average absolute value of the correlation of each element of the fusion correlation vector equal to 0.5 as the critical point; when the l1 norm of the fusion correlation vector is greater than or equal to 0.5q, it is considered that the critical condition is met; where q is the number of domains.

[0017] As a further implementation method, the correlation coefficients of the data of each dimension in the Pearson correlation coefficient are horizontally spliced ​​to obtain a fused correlation vector; when the norm of the fused correlation vector meets the critical condition, the final multi-domain fusion matrix is ​​output.

[0018] As a further implementation, the cutting signal feature set includes a time domain feature matrix, a frequency domain feature matrix, and a time-frequency domain feature matrix;

[0019] The identification target vector is obtained based on the observation data.

[0020] In a second aspect, an embodiment of the present invention further provides a method for evaluating the performance of high-quality fusion features of cutting signals using multi-domain features, including:

[0021] According to the fusion matrix obtained by different feature fusion methods, the Pearson correlation coefficient, normalized mutual information and monotonicity of each fusion vector and the identification target vector are calculated to obtain the feature representation vector;

[0022] The fusion feature performance index is obtained according to the feature representation vector, and the fusion feature performance indexes of different feature fusion methods are ranked to form an evaluation scheme for the fusion feature performance.

[0023] As a further implementation, the fusion feature performance index PI is expressed as:

[0024]

[0025] Among them, σ(·) represents the activation function, CR x represents the Pearson correlation coefficient, NI x represents the normalized mutual information, MO x Represents the monotonicity value, and the value of q is an integer.

[0026] As a further implementation method, the correlation, mutual information and monotonicity of the fused features between the fused features and the identified target vector are calculated, and the average is accumulated and calculated, and then mapped to the [0,1] interval through the activation function.

[0027] As a further implementation method, multiple fusion feature matrices are obtained based on different feature fusion methods, and the obtained fusion feature performance indicators PI are sorted in descending order, that is, the performance ranking of the fusion features is obtained.

[0028] The beneficial effects of the present invention are as follows:

[0029] (1) Based on the physical laws of tool wear and equipment performance degradation during the cutting process, the present invention establishes a feature fusion function to automatically eliminate weakly correlated feature components according to the performance of the original features, thereby autonomously generating a multi-domain feature fusion method for fusion feature vectors; it avoids the influence of human factors in feature screening, effectively improves the performance of fusion features, and ultimately achieves the purpose of high-quality fusion of multi-domain features.

[0030] (2) The multi-domain feature fusion method of the present invention directly performs superposition operations on the original data features based on the established feature fusion function, avoiding the mapping transformation calculations in traditional dimensionality reduction fusion methods such as principal component analysis (PCA), distributed domain embedding (TSNE), and unified manifold approximation and projection (UMPA), effectively improving the computational efficiency of the multi-domain feature fusion process.

[0031] (3) The fusion feature performance evaluation method of the present invention avoids the tedious training process of the recognition model and greatly improves the efficiency of the fusion feature performance evaluation; at the same time, by introducing a nonlinear activation function, the fusion feature performance evaluation method of the present invention integrates multiple evaluation indicators such as feature monotonicity, mutual information, and correlation with the recognition target, ensuring the robustness of the evaluation method and being able to comprehensively reflect the feature performance capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0033] Figure 1 is a flow chart of a method for high-quality fusion of multi-domain features of cutting signals according to one or more embodiments of the present invention;

[0034] Figure 2 is a flow chart of a method for evaluating cutting signal fusion feature performance according to one or more embodiments of the present invention;

[0035] Figure 3 It is the first two-dimensional data distribution diagram of the fusion results of the traditional dimensionality reduction fusion method and the multi-domain feature high-quality fusion method of the present invention;

[0036] Figure 4 It is a comparison chart of the regression curve of the fusion result obtained based on multiple fusion methods and the change trend of the identified target;

[0037] Figure 5 The fusion feature performance indicators and ranking diagrams of different fusion methods. DETAILED DESCRIPTION

[0038] Example 1:

[0039] This embodiment provides a high-quality fusion method for multi-domain features of cutting signals, including:

[0040] According to the cutting signal feature set and the identification target vector, the Pearson correlation coefficient between each feature vector and the identification target vector is calculated to obtain a multi-domain correlation coefficient set;

[0041] Calculate the fusion feature component corresponding to each feature vector according to the automatic superposition function to obtain a multi-domain fusion feature component set;

[0042] The average vector of each fused feature component matrix in the multi-domain fusion feature component set is calculated horizontally to obtain the corresponding intra-domain fusion vector. The intra-domain fusion vectors are sequentially spliced ​​to obtain the multi-domain fusion matrix. The intra-domain fusion vector is obtained by superimposing the representation of each feature vector to automatically eliminate weakly correlated feature components in the original features.

[0043] The Pearson correlation coefficient of each dimension of the multi-domain fusion matrix and the identification target vector is calculated respectively to obtain the fusion correlation vector of the multi-domain fusion matrix and the identification target vector; the norm is calculated by fusion correlation vector, and the final multi-domain fusion matrix is ​​output.

[0044] Specifically, such as Figure 1 As shown, the following steps are included:

[0045] S1. Input the extracted cutting signal feature set F = {F t ,F f ,F tf}, F∈R m×l , the feature set contains the time domain feature matrix F t , frequency domain feature matrix F f And the time-frequency domain feature matrix F tf , where each feature matrix is ​​recorded as:

[0046]

[0047]

[0048]

[0049] Among them, f i Represents the i-th eigenvector in the multi-domain; m represents the sample capacity of the signal feature set, that is, there are m signal samples in total; j, k, and n represent the number of features of each signal sample in the time domain, frequency domain, and time-frequency domain, respectively, and their sum is the number of all features l, that is, the feature dimension.

[0050] S2. By recording observation data t p, depending on the applicable scenario, the observation data can be the tool flank wear width, equipment performance degradation index, etc., thereby obtaining the identification target vector t:

[0051] t=(t1,t2,...,t p ,...,t m ) T (4)

[0052] Among them, t p Represents the identification target value corresponding to the p-th signal sample.

[0053] S3. Calculate each eigenvector f in turn i Pearson correlation coefficient with the identification target vector t:

[0054]

[0055] Among them, cov represents the covariance between the calculated feature vector and the target vector. and σ t Represent the cutting process feature vector f i And the standard deviation of the identification target vector t, from which the multi-domain correlation coefficient set C, C∈R 1×l :

[0056] C={C t ,C f ,C tf} (6)

[0057] Among them, the time domain correlation coefficient vector C t =(c1,c2,…c i ,…,c j ) T ,C t ∈R 1×j , frequency domain correlation coefficient vector Time-frequency domain correlation coefficient vector C tf =(c j+k+1 ,c j+k+2 ,…,c j+k+n ) T ,C f ∈R 1×n .

[0058] S4. Calculate each eigenvector f in turn by the automatic superposition function described in formula (7) i The corresponding fusion feature component a i :

[0059]

[0060] Where β represents the fusion level, which is an odd number greater than 1, and the initial parameter is β = 3. Thus, we can get the multi-domain fusion feature component set A = {A t ,A f ,A tf}, A∈R m×l , where the time domain fusion feature component matrix A t =(a1,a2,…a i ,…,a j ) T , frequency domain fusion feature component matrix A f =(a j+1 ,a j+2 ,…,a j+k ) T , time-frequency domain fusion feature component matrix A tf =(a j+k+1 ,a j+k+2 ,…,a j+k+n ) T .

[0061] S5. Calculate the time domain, frequency domain, and time-frequency domain fusion feature component matrix A respectively. t 、A f 、A tf The average vector of the fusion feature components in the domain is calculated horizontally to obtain the fusion vector in the time domain, the fusion vector in the frequency domain, and the fusion vector in the time-frequency domain. The multi-domain fusion matrix A can be obtained by sequentially splicing the fusion vectors in each domain * :

[0062]

[0063] So far, through the formula (7) in S4 and the formula (8) in S5, it is possible to calculate the value of each eigenvector f i The expressive power of the two algorithms is superimposed to obtain the intra-domain fusion vector, thereby automatically eliminating the weakly correlated feature components in the original features.

[0064] S6. Calculate the multi-domain fusion matrix A separately * The dimensional data A in x * Pearson correlation coefficient CR with the identification target vector t x , horizontally splicing the correlation coefficients of each dimension data, we can get the multi-domain fusion matrix A * The fusion correlation vector CR with the identification target vector t * =(CR1,CR2,…,CR x ,…,CR q ).

[0065] This embodiment includes three fusion domains: time domain, frequency domain, and time-frequency domain, that is, q=3.

[0066] According to Equation (7) in S3, the degree of elimination of weakly correlated feature components in the original feature data can be adjusted by modifying the fusion level β. A larger fusion level β indicates that more weakly correlated feature components are eliminated. It is generally believed that when the Pearson correlation coefficient of two variables is above 0.5, there is at least a moderate correlation between the variables.

[0067] Therefore, the value of the fusion level β is determined by taking the average absolute value of the correlation of each element of the fusion correlation vector equal to 0.5 as the critical point. * The critical condition is considered to be reached when ||1 ​​is greater than or equal to 0.5q, where q is the number of domains. Therefore, ||CR * When ||1<0.5, set β in S4=β+2 and repeat S4 and S5. * When ||1≥0.5, the final multi-domain fusion matrix A is output * .

[0068] This embodiment is based on the physical laws of tool wear and equipment performance degradation during the cutting process. By establishing a feature fusion function, it realizes a multi-domain feature fusion method that automatically eliminates weakly correlated feature components according to the expressive ability of the original features, thereby autonomously generating a fused feature vector. It avoids the influence of human factors in feature screening, effectively improves the expressive ability of the fused features, and ultimately achieves the goal of high-quality fusion of multi-domain features.

[0069] This embodiment directly performs superposition operations on the original data features based on the established feature fusion function, avoiding the mapping transformation calculations in traditional dimensionality reduction fusion methods such as principal component analysis (PCA), distributed domain embedding (TSNE), and unified manifold approximation and projection (UMPA), effectively improving the computational efficiency of the multi-domain feature fusion process.

[0070] Example 2:

[0071] This embodiment provides a method for evaluating the performance of high-quality fusion features of cutting signals using multi-domain features, including:

[0072] According to the fusion matrix obtained by different feature fusion methods, the Pearson correlation coefficient, normalized mutual information and monotonicity of each fusion vector and the identification target vector are calculated to obtain the feature representation vector;

[0073] The fusion feature performance index is obtained according to the feature representation vector, and the fusion feature performance indexes of different feature fusion methods are ranked to form an evaluation scheme for the fusion feature performance.

[0074] Among them, different feature fusion methods include three traditional dimensionality reduction fusion methods: principal component analysis (PCA), distributed domain embedding (TSNE), and unified manifold approximation and projection (UMPA), as well as the multi-domain feature high-quality fusion method described in Example 1.

[0075] Specifically, such as Figure 2 As shown, the following steps are included:

[0076] S1. Input the fusion matrix A obtained by different feature fusion methods * =(A1 * ,A2 * ,…,A x * ,…,A q * )∈R m×q , calculate the fusion matrix A in sequence * Each fusion vector A in x * Pearson correlation coefficient CR with the identification target vector t x , and calculate the normalized mutual information NI with the identification target vector t x , and the fusion vector A x * Monotonicity MO x .

[0077] Among them, the Pearson correlation coefficient can be obtained by formula (5) in embodiment 1, and the fusion vector A x * The corresponding normalized mutual information NI x It can be expressed as:

[0078]

[0079] Among them, H(·) represents the calculated information entropy, I x (A x * ; t) represents the fusion vector A x * The mutual information with the identification target vector t is expressed as:

[0080]

[0081] Where A x u represents the xth characteristic element of the uth signal sample in the sample space, P(A x u ) and P(t p ) represents the fusion vector A x * and the marginal probability distribution function of the target vector t, P(Ax u ,t p ) is the fusion vector A x * and the joint probability distribution function of the identification target vector t.

[0082] Fusion vector A x * Monotonicity MO x They can be obtained by formula (11):

[0083] MO x =|corr(rank(A x * ),rank(s x ))| (11)

[0084] In the formula, corr(·) represents the calculation of Spearman's rank correlation coefficient, rank(·) represents the calculation of Spearman's rank, that is, the rank of each vector element in the vector to which it belongs, s x Represents characteristic element A x * The corresponding sampling time series vector.

[0085] S2. For each fusion vector A x * The Pearson correlation coefficient CR can be obtained x , normalized mutual information NI x And the monotonicity value MO x , therefore, we can get the feature representation vector P = (CR1, NI1, MO1, CR2, NI2, MO2, ..., CR q ,NI q ,MO q )∈R 1×3q , then the fusion feature performance index PI can be defined as:

[0086]

[0087] Where σ(·) represents the sigmoid function, which is expressed as:

[0088]

[0089] S3. Based on different feature fusion methods, multiple fusion feature matrices A can be obtained * , repeat S7 and S8, and sort the obtained fusion feature performance index PI in descending order, that is, the ranking of the fusion feature performance is obtained, thereby realizing the evaluation of the fusion feature performance.

[0090] This embodiment avoids the tedious training process of the recognition model and greatly improves the efficiency of the fusion feature performance evaluation. Furthermore, by introducing a nonlinear activation function, the fusion feature performance evaluation method in this embodiment integrates multiple evaluation indicators such as feature monotonicity, mutual information, and correlation with the recognition target, ensuring the robustness of the evaluation method and comprehensively reflecting the feature performance capabilities.

[0091] Example 3:

[0092] In order to verify the feasibility of the high-quality fusion of multi-domain features of cutting signals and the fusion feature performance evaluation method in Example 1 and Example 2, the C1 milling tool wear data and x-direction cutting force signal data in the "PHM 2010 Tool Wear Dataset" published by the International Fault Diagnosis and Health Management Association were used to verify the high-quality fusion method of multi-domain features and the fusion feature performance evaluation method.

[0093] The experimental data set contains cutting force signals of 315 tool passes. According to Table 2 in the literature (Lei Y, He Z, Zi Y, et al. Fault diagnosis of rotating machinery based on multiple ANFIS combination with GAs [J]. Mechanical Systems & Signal Processing, 2007, 21 (5): 2280-2294.), 11 time domain features and 12 frequency domain features (excluding the last frequency domain feature in the literature) can be extracted from the cutting force signal. By performing a three-layer wavelet packet decomposition on the cutting force signal, 8 wavelet packet energies, i.e., 8 time-frequency domain features, can be obtained.

[0094] From this, we can get the time domain feature matrix F in equations (1) to (3): t , frequency domain feature matrix F f And the time-frequency domain feature matrix F tf , where m=315, j=11, k=12, and n=8.

[0095] In addition, in the experimental data set, the cutting force signal of each tool pass corresponds to a measured flank wear width, from which the identification target vector t in Equation (4) can be obtained.

[0096] Thus, according to S3, the multi-domain correlation coefficient set C can be obtained. In this embodiment, the calculated C t =(-0.005,0.956,0.958,0.956,0.955,0.247,0.696,0.339,0.412,0.699,0.388), C f=(-0.889,-0.229,0.980,0.871,-0.621,0.691,0.986,0.927,-0.608,-0.636,0.943,0.978), C tf =(0.957,0.933,0.924,0.951,0.949,0.908,0.935,0.928).

[0097] Based on the obtained multi-domain correlation coefficient set C and the time domain, frequency domain, and time-frequency domain feature matrices, the fused feature components corresponding to each feature vector can be calculated according to S4, and then the multi-domain fused feature component set A = {At, Af, Atf} is obtained. In this embodiment, when β = 13, the critical condition requirements can be met.

[0098] The feature component matrix of each domain is fused separately, and the average value is accumulated horizontally to obtain the fusion vector in each domain. The fusion vectors in the domain are sequentially spliced ​​to obtain the multi-domain fusion matrix A. * At this point, the automatic elimination of weakly correlated features is achieved under the action of the multi-domain correlation coefficient set C and the automatic superposition function, that is, the high-quality fusion of multi-domain features.

[0099] Figure 3 The distribution diagram of the first two dimensions of the fusion results obtained using three traditional dimensionality reduction fusion methods: principal component analysis (PCA), distributed domain embedding (TSNE), and unified manifold approximation and projection (UMPA), as well as the high-quality fusion method of multi-domain features in this embodiment; Figure 4 The graph is a graph of nonlinear regression of the fusion results obtained by different fusion methods and mapped to the sample space, as well as the changing trend of the identification target (tool flank wear width) in this embodiment.

[0100] pass Figure 3 and Figure 4 It can be seen that the second dimension data of the results of the multi-domain feature high-quality fusion method proposed in this embodiment tends to increase with the first dimension, and the distribution is more in line with the physical law of cumulative tool wear. This shows that it is easier to learn the tool wear trend from the feature data fused by the present invention, and the complexity requirement of the recognition model is lower.

[0101] In addition to the qualitative comparison, the cutting signal fusion feature performance evaluation method proposed in this embodiment can quantitatively compare the performance differences of different fusion methods.

[0102] According to S1 in the second embodiment, the fusion matrix A obtained by PCA, TSNE, UMPA and the high-quality fusion method of multi-domain features in this embodiment is input respectively. *According to formula (9) to formula (11), the Pearson correlation coefficient, normalized mutual information, and monotonicity of the fusion vector between the fusion vector and the identification target vector can be calculated for the fusion matrix obtained by different fusion methods. Finally, according to S2 and S3 in Example 2, the fusion feature performance index PI of each fusion method and the corresponding ranking can be obtained, as shown in the following example: Figure 5 As shown in Figure 3, the feasibility of the cutting signal fusion feature performance evaluation method in this embodiment is demonstrated. It can be found that the fusion feature performance obtained by using the multi-domain feature high-quality fusion method proposed in this embodiment is significantly better than the other three traditional dimensionality reduction fusion methods.

[0103] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A high-quality fusion method for multi-domain features of cutting signals, characterized by: include: According to the cutting signal feature set and the identification target vector, the Pearson correlation coefficient between each feature vector and the identification target vector is calculated to obtain a multi-domain correlation coefficient set; The fusion feature component corresponding to each feature vector is calculated according to the automatic superposition function to obtain a multi-domain fusion feature component set; the automatic superposition function is: Among them, β represents the fusion level, c i represents the Pearson correlation coefficient, f i represents the eigenvector; The average vector of each fused feature component matrix in the multi-domain fusion feature component set is calculated horizontally to obtain the corresponding intra-domain fusion vector. The intra-domain fusion vectors are sequentially spliced ​​to obtain the multi-domain fusion matrix. The intra-domain fusion vector is obtained by superimposing the representation of each feature vector to automatically eliminate weakly correlated feature components in the original features. The Pearson correlation coefficient of each dimension of the multi-domain fusion matrix and the identification target vector is calculated respectively to obtain the fusion correlation vector of the multi-domain fusion matrix and the identification target vector; the norm is calculated by fusion correlation vector, and the final multi-domain fusion matrix is ​​output.

2. A high-quality fusion method for multi-domain features of cutting signals according to claim 1, characterized in that: The degree of elimination of weakly correlated feature components in the original feature data is adjusted by modifying the fusion level.

3. The high-quality fusion method of multi-domain features of cutting signals according to claim 2 is characterized in that: The value of the fusion level β is determined by taking the average absolute value of the correlation of each element of the fusion correlation vector equal to 0.5 as the critical point; when the fusion correlation vector l 1-norm greater than or equal to 0.5 q When , the critical condition is considered to be reached; q is the number of domains.

4. The high-quality fusion method of multi-domain features of cutting signals according to claim 3 is characterized in that: The correlation coefficients of each dimension of the Pearson correlation coefficient are horizontally spliced ​​to obtain a fused correlation vector; when the norm of the fused correlation vector meets the critical condition, the final multi-domain fusion matrix is ​​output.

5. The high-quality fusion method of multi-domain features of cutting signals according to claim 1 is characterized in that: The cutting signal feature set includes a time domain feature matrix, a frequency domain feature matrix and a time-frequency domain feature matrix; The identification target vector is obtained based on the observation data.

Images