Tool wear monitoring method and system based on multi-domain feature behavior law representation

Abstract

The present invention discloses a tool wear monitoring method and system based on multi-domain feature behavior law characterization, which relates to the technical field of thin-walled parts cutting process monitoring, with the monotonicity of the feature vector as the first indicator, the normalized mutual information between the feature vector and the tool wear vector as the second indicator, and the ReLU function of the Spearman correlation coefficient between the feature vector and the tool wear vector as the third indicator; according to the above indicators, the behavior indicators corresponding to each feature vector are determined; the behavior characteristics of the signal channel are characterized, the channel behavior indicators of each signal channel are compared, and the model input vector is determined; the tool state recognition model is trained according to the input vector. The present invention can accurately determine the input signal channel and signal feature vector of the tool state recognition model by quantitatively characterizing the evolution law of the multi-domain feature behavior of the cutting signal and the evolution law of the channel behavior, thereby solving the problem of difficulty in ensuring the accuracy of tool wear monitoring caused by the random influence of tool-work vibration in the milling of thin-walled parts.

Description

Technical Field

[0001] The present invention relates to the technical field of thin-walled part cutting process monitoring, and in particular to a tool wear monitoring method and system based on multi-domain feature behavior law characterization. Background Art

[0002] Due to their advantages in structure, specific stiffness, specific strength, etc., thin-walled workpieces made of difficult-to-process materials, such as titanium alloys and high-temperature alloys, are widely used in key components of aircraft engines, such as impellers, blades, integral blade disks, and casings. However, such workpiece materials have low thermal conductivity, high cutting forces, and rapid tool wear during the cutting process. With the development of artificial intelligence technology, cutting data-driven online tool status monitoring technology has provided a possible solution to this problem.

[0003] Under reasonable process parameters, the cutting process of general conventional workpieces can remain relatively stable. As the tool becomes blunt, the signal characteristics and tool wear show an obvious nonlinear correlation. By extracting multi-domain features related to the evolution of tool state from the sensor signal and combining them with pattern recognition methods such as machine learning and deep learning, real-time monitoring of tool wear can be achieved. For example, the prior art CN114905336A discloses a variable working condition tool wear monitoring method and system based on decoupling of cutting force components, which monitors the tool state by calculating the cutting force increment caused by tool wear and the theoretical cutting force, thereby realizing the identification of tool wear failure under variable load conditions. The prior art CN106514434B discloses a data-based milling tool wear monitoring method, which extracts characteristic coefficients that can characterize milling tool wear from the three-phase output current signal of the spindle drive motor, thereby enabling real-time monitoring of milling tool wear during machining.

[0004] However, in the milling of thin-walled components made of difficult-to-process materials such as aircraft engine impellers and blades, the forced vibration and self-excited vibration of the tool and workpiece caused by the weak rigidity of the thin-wall system are difficult to avoid. The influence of random tool-workpiece vibration makes the behavioral evolution mechanism of signal characteristics significantly different from that of conventional workpiece cutting, and the correlation between multi-domain characteristics and wear amount is greatly reduced, resulting in difficulty in reliably identifying tool wear in thin-walled parts milling. At the same time, the non-normal distribution of the feature vectors of thin-walled parts cutting data makes it difficult to quantitatively characterize its feature evolution mode, making it difficult to consider the behavioral laws of multi-domain characteristics in the existing thin-walled parts milling tool state monitoring, and can only improve the accuracy of tool wear identification through primitive means such as model expansion and feature space complication. The complexity of the model and feature space causes the training process to consume a lot of computing resources, and the complex recognition model is also difficult to generalize to different working conditions, resulting in the inability of the state monitoring algorithm to adapt to engineering practice. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a tool wear monitoring method and system based on the characterization of multi-domain characteristic behavior laws. By quantitatively characterizing the evolution laws of the multi-domain characteristic behaviors of cutting signals and the evolution laws of channel behaviors, the input signal channel and signal characteristic vector of the tool state recognition model can be accurately determined, thereby solving the problem of difficulty in ensuring the accuracy of tool wear monitoring caused by the random influence of tool-workpiece vibration in the milling of thin-walled parts.

[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 tool wear monitoring method based on multi-domain feature behavior law characterization, comprising:

[0008] Determine an index for evaluating the characteristic behavior, wherein the monotonicity of the characteristic vector is used as a first index, the normalized mutual information between the characteristic vector and the tool wear vector is used as a second index, and the ReLU function of the Spearman correlation coefficient between the characteristic vector and the tool wear vector is used as a third index;

[0009] Determine the behavior index corresponding to each feature vector according to the Sigmoi d activation function of the first index, the second index and the third index; calculate the cumulative average value of the behavior index of each feature vector to characterize the behavior characteristics of the signal channel;

[0010] The channel behavior indicators of each signal channel are compared to determine the model input vector; a lightweight gated recurrent unit network tool state recognition model is trained according to the input vector to monitor the tool wear state.

[0011] As a further implementation method, for each cutting signal channel of each tool, the monotonicity of each eigenvector, the normalized mutual information between each eigenvector and the tool wear vector, and the Spearman correlation coefficient between each eigenvector and the tool wear vector are cyclically calculated.

[0012] As a further implementation method, the performance index of the feature vector that is negatively correlated with the tool wear vector is directly set to zero through the ReLU function.

[0013] As a further implementation, according to the first index, the second index and the third index corresponding to each characteristic vector, a characteristic behavior index corresponding to the cutting tool characteristic vector is obtained;

[0014] The characteristic behavior indexes of all cutting tool characteristic vectors are averaged to obtain the behavior index corresponding to each characteristic vector.

[0015] As a further implementation method, the channel behavior indicators of each signal channel are compared, and the signal channels with the top two behavior indicators are selected as input signal channels. A feature vector with the largest characteristic behavior indicator is selected in the time domain, frequency domain, and time-frequency domain of each input signal channel to form a model input vector.

[0016] As a further implementation, the tool state recognition model has four layers in total.

[0017] As a further implementation, the first and second layers include a number of bidirectional gated recurrent units, and the third and fourth layers include a number of fully connected layers of neurons.

[0018] In a second aspect, an embodiment of the present invention further provides a tool wear monitoring system based on multi-domain characteristic behavior law characterization, comprising:

[0019] The indicator determination module is configured to: determine an indicator for evaluating the characteristic behavior, wherein the monotonicity of the characteristic vector is used as a first indicator, the normalized mutual information between the characteristic vector and the tool wear vector is used as a second indicator, and the ReLU function of the Spearman correlation coefficient between the characteristic vector and the tool wear vector is used as a third indicator;

[0020] The signal channel behavior feature characterization module is configured to: determine the behavior index corresponding to each feature vector according to the Sigmoid activation function of the first index, the second index and the third index; calculate the cumulative average value of each feature vector behavior index to characterize the behavior characteristics of the signal channel;

[0021] The tool state recognition model building module is configured to: compare the channel behavior indicators of each signal channel to determine the model input vector; train the lightweight gated recurrent unit network tool state recognition model according to the input vector to monitor the tool wear state.

[0022] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the tool wear monitoring method based on multi-domain feature behavior law characterization is implemented.

[0023] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the tool wear monitoring method based on the characterization of multi-domain feature behavior laws.

[0024] The beneficial effects of the present invention are as follows:

[0025] (1) The present invention can accurately determine the input signal channel and signal feature vector of the tool state recognition model by quantitatively characterizing the evolution law of the multi-domain characteristic behavior of the cutting signal and the evolution law of the channel behavior, thereby solving the problem of difficulty in ensuring the accuracy of tool wear monitoring caused by the random influence of tool-workpiece vibration in the milling of thin-walled parts.

[0026] (2) The present invention solves the problem of difficulty in characterizing the correlation between eigenvectors caused by the non-normal distribution of eigenvectors of thin-walled workpiece cutting signals by calculating the Spearman correlation between the eigenvectors and the tool wear vector. At the same time, by directly setting the negatively correlated eigenvector behavior indicators to zero, the amount of calculation is greatly reduced, effectively improving the calculation efficiency while ensuring the characterization accuracy.

[0027] (3) The lightweight gated recurrent unit network tool state recognition model established in the present invention only includes a 4-layer network structure, and the maximum number of network units in each layer is only 64, and the minimum is 1; compared with the existing tool state monitoring model based on deep learning, the model structure is simple, the model parameters that need to be trained are relatively small, and the model training can be completed with very little computing resources, making the state monitoring algorithm easily adaptable to the actual thin-wall workpiece processing engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings in the specification, 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.

[0029] Figure 1 is a flow chart of the present invention according to one or more embodiments;

[0030] Figure 2 is a multi-signal channel multi-domain characteristic behavior index of each milling tool according to one or more embodiments of the present invention;

[0031] Figure 3 is a multi-domain feature vector behavior characterization indicator according to one or more embodiments of the present invention;

[0032] Figure 4 is a multi-signal channel behavior characterization indicator according to one or more embodiments of the present invention;

[0033] Figure 5 is a structural diagram of a lightweight gated recurrent unit network tool state recognition model according to one or more embodiments of the present invention;

[0034] Figure 6 It is a comparison chart of tool wear monitoring results for thin-walled part milling and actual tool wear according to one or more embodiments of the present invention. DETAILED DESCRIPTION

[0035] Embodiment 1:

[0036] This embodiment provides a tool wear monitoring method based on multi-domain feature behavior law representation, including:

[0037] Determine an index for evaluating the characteristic behavior, wherein the monotonicity of the characteristic vector is used as a first index, the normalized mutual information between the characteristic vector and the tool wear vector is used as a second index, and the ReLU function of the Spearman correlation coefficient between the characteristic vector and the tool wear vector is used as a third index;

[0038] Determine the behavior index corresponding to each feature vector according to the Sigmoid activation function of the first index, the second index and the third index; calculate the cumulative average value of each feature vector behavior index to characterize the behavior characteristics of the signal channel;

[0039] The channel behavior indicators of each signal channel are compared to determine the model input vector; a lightweight gated recurrent unit network tool state recognition model is trained according to the input vector to monitor the tool wear state.

[0040] Specifically, Figure 1 As shown, the following steps are included:

[0041] S1. Input the signal feature matrix obtained from the j-th tool cutting signal, which can be expressed as F j =(f 1j ,f 2j ,…,f ij ,…f nj )∈R m×n , where f ij represents the i-th eigenvector in the feature space of the cutting signal of the j-th tool, m represents the size of the cutting signal sample space, and n represents the dimension of the feature space.

[0042] S2. For each cutting signal channel of each tool, the monotonicity of each feature vector is calculated cyclically as the first indicator for evaluating the feature behavior. The feature vector f ij Monotonicity MO ij It can be expressed by formula (1):

[0043] MO ij =|corr(rank(f ij ),rank(t ij ))| (1)

[0044] In formula (1), corr represents the calculation of Spearman rank correlation coefficient, rank represents the calculation of Spearman rank, and t ij Representation and eigenvector f ij The corresponding time series vector.

[0045] S3. For each cutting signal channel of each tool, loop to calculate each feature vector f ij and tool wear vector w j The normalized mutual information NI ij , as the second indicator to characterize the characteristic behavior, its expression is:

[0046]

[0047] In formula (2), I ij (f ij ;w j ) represents the calculation of the eigenvector f ij and the wear vector w of the jth tool j The mutual information between them, H(·) represents the information entropy of the calculated vector.

[0048] S4. For each cutting signal channel of each tool, loop to calculate each feature vector f ij and tool wear vector w j The Spearman correlation coefficient of the tool wear vector is directly set to zero through the ReLU function, which is used as the third indicator for evaluating the characteristic behavior. The calculation steps are expressed as

[0049] C ij =ReLU(r ij ) (3)

[0050] In formula (3), r ij is the feature vector f ij and tool wear vector w j The Spearman correlation coefficient is:

[0051]

[0052] In formula (4), d ij (u) Denotes the eigenvector f ij The uth feature data and tool wear vector w in j The grade difference between the u-th tool wear values ​​in .

[0053] The distribution of thin-walled cutting signal feature vectors does not necessarily satisfy the normal distribution. In this case, the traditional Pearson correlation coefficient is difficult to characterize the correlation between the signal feature vector and the tool wear. In this step, the Spearman correlation is calculated to solve the correlation characterization problem in thin-walled cutting tool condition monitoring.

[0054] In addition, the expression of the ReLU function in formula (3) is:

[0055] ReLU(x)=max(0,x) (5)

[0056] According to equations (3) and (5), this step can directly set the performance index of the characteristic vector that is negatively correlated with the tool wear vector to zero, reducing unnecessary subsequent calculations and greatly improving the efficiency of characteristic index calculation.

[0057] S5. Based on each feature vector f in the feature space of the jth cutting tool ij The corresponding first indicator MO ij , the second indicator NI ij and the third indicator C ij , we can get the cutting tool feature vector f ij The corresponding characteristic behavior index P ij for:

[0058] P ij =σ(MO ij +NI ij +C ij ) (6)

[0059] In formula (6), σ(·) represents the Sigmoid activation function.

[0060] Based on this, the characteristic behavior index of all cutting tool feature vectors is calculated to get the behavior index corresponding to each feature vector.

[0061]

[0062] In formula (7), P i represents the characteristic behavior index of the i-th eigenvector, and p is the number of cutting tools used in thin-walled parts processing; thus, the quantitative characterization of the multi-domain characteristic vector behavior in each signal channel is achieved.

[0063] S6. Calculate the cumulative average of each characteristic vector behavior index of each signal channel when each tool is cutting to characterize the behavior characteristics of the signal channel

[0064]

[0065] In formula (8), P c Represents the channel behavior characterization index of the cth signal channel. So far, the quantitative characterization of the behavior laws of multiple signal channels has been achieved.

[0066] After completing the quantitative characterization of the characteristic behavior law and the channel behavior law, lightweight monitoring and identification of thin-walled part milling tools can be achieved. The tool state monitoring method based on multi-domain feature evolution characterization includes the following steps:

[0067] S7. Compare the channel behavior index P of each signal channel c , select the top two signal channels with the highest behavior index as the input signal channels, and select a characteristic behavior index P in the time domain, frequency domain and time-frequency domain of each input signal channel ij The largest eigenvector is used to form the model input vector, which is expressed as:

[0068]

[0069] In formula (9), f t (1) ,f f (1) ,f tf (1) They represent the eigenvectors with the largest time domain, frequency domain and time-frequency domain characteristic behavior indicators in the signal channel with the first-ranked behavior indicator, respectively. t (2) ,f f (2) ,f tf (2) They respectively represent the eigenvectors with the largest characteristic behavior indicators in the time domain, frequency domain, and time-frequency domain in the signal channel with the second-ranked behavior indicator.

[0070] S8. Based on the input vector obtained in S7, a lightweight gated recurrent unit network tool state recognition model is trained. The lightweight gated recurrent unit network tool state recognition model has a total of 4 layers; among them, the first layer contains 16 bidirectional gated recurrent units, the second layer contains 32 bidirectional gated recurrent units, the third layer is a fully connected layer containing 64 neurons, and the fourth layer is a fully connected layer containing 1 neuron.

[0071] Among them, the expression of the gated recurrent unit is

[0072]

[0073]

[0074]

[0075]

[0076] Where W * ,U * ,b * , (*∈{h,r,z}) is the weight matrix and bias vector that can be learned, z i is the update gate, r i Represents the reset gate, and Represent the current state and the state at the previous moment respectively. Candidate status.

[0077] The expression of the fully connected layer is:

[0078] output = ReLU(WX+b) (14)

[0079] In formula (14), W,b is the learnable weight matrix and bias vector, and X is the output vector of the previous layer.

[0080] In this embodiment, the model parameters in the training process are set as follows: the batch size is 32, the learning rate is 0.001, the number of training rounds is 300, and the RMSprop algorithm is used to optimize the training process.

[0081] S9. Collect process signals in real time during the milling of thin-walled parts, repeat S1 to S7 based on the collected cutting signals, and input the feature vector into the lightweight tool state recognition model trained in S8 to achieve lightweight online monitoring of the tool state.

[0082] The process signal may be a combination of any two signal channels of milling force, workpiece acceleration, cutting noise, and acoustic emission signals.

[0083] Embodiment 2:

[0084] In order to verify the feasibility of Example 1, a DMU70V five-axis machining center was used to carry out milling experiments on aviation titanium alloy thin-walled parts and milling tool wear measurement experiments.

[0085] The workpiece is a Ti6Al4V rectangular thin plate with a geometric size of 150*100*5mm; the milling tool is a toothed two-edge end mill with a tool diameter of 14mm; side milling is used to process thin-walled workpieces, with a spindle speed of 8000r / min, a feed rate of 1280mm / min, a radial cutting depth of 0.2mm, and an axial cutting depth of 4mm.

[0086] In order to verify the above monitoring method, three identical cutting tools were used to carry out milling experiments with 100 passes in sequence.

[0087] In order to improve the experimental efficiency, the tool flank wear was measured by electron microscope after every 10 passes. Based on the measured results, the flank wear corresponding to each pass was obtained by nonlinear interpolation and used as a data label. During the cutting process, the vibration signals of the workpiece in the machine tool feed direction, vertical feed direction, and axial direction were collected in real time by a three-axis accelerometer. x , v y , v z At the same time, the milling force f in the three directions of tangential, radial and axial of the tool is collected in real time through the Kistler rotary force measuring tool holder x , fy , f z , and the milling moment m in the axial direction z .

[0088] For the cutting signals of each channel of each tool during cutting, according to Table 2 in Reference 1 (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), p1 to p2 in the table are extracted in sequence. 11 11 time domain features of the , and extract the maximum value of the signal as a supplement, a total of 12 time domain features; and the table p 12 ~p 23 11 frequency domain features expressed; and, based on the db1 wavelet basis function, 3 layers of wavelet packet decomposition are performed on each channel cutting signal to obtain 8 wavelet packet energies, that is, 8 time-frequency domain features, so that a total of 32 multi-domain features can be obtained, that is, the signal feature matrix F in step S1 is obtained. j , where the 1st to 12th eigenvectors are time domain features, the 13th to 24th eigenvectors are frequency domain features, and the 25th to 32nd eigenvectors are time-frequency domain features. The signal sample space size m is equal to the number of passes, 100, and the feature space dimension n is equal to the number of multi-domain features, 32.

[0089] Therefore, for the cutting signal of each tool, the first index, the second index and the third index corresponding to each feature vector of each signal channel can be calculated according to steps S2 to S4 respectively. According to formula (6) in step S5, the characteristic behavior index of the multi-domain feature vector of each channel of each tool can be obtained, such as Figure 2 As shown. Therefore, according to formula (7) in step S5, the behavior characterization index corresponding to each feature vector can be obtained, and its value is as follows Figure 3 Based on this, according to step S6, a quantitative characterization index of the signal channel behavior can be obtained. Figure 4 What is displayed are the calculated channel behavior indicators of each signal channel.

[0090] According to step S7, based on Figure 4 , select f x and m z As the input signal channel, according to Figure 3 Select the 5th, 16th, and 26th features as f x The input features of the channel, select the 5th, 14th, and 26th features as mz The input features of the channel can be used to obtain the model input vector in step S7.

[0091] In this embodiment, the cutting signals of tool 1 and tool 2 are used as training data. Based on step S8, the establishment and training of the tool state recognition model of the lightweight gated recurrent unit network can be completed. Figure 5 The structure and number of parameters of the trained model are shown. Since two channels are input and three feature vectors are input for each channel, according to step S8, the batch size is set to 32. In this embodiment, the sequence length is set to 15, and the input data size is [32, 15, 6].

[0092] According to step S8, the first layer of the model is 16 bidirectional gated recurrent units, then the first layer output data is [32, 15, 32], and the input and output sizes and parameter numbers of the remaining layers can be obtained in turn according to step S8. Compared with the parameters of the thin-walled milling tool state recognition model in document 2 (Wang R, Song Q, Peng Y, et al. A milling tool wear monitoring method with sensing generalization capability [J]. Journal of Manufacturing Systems, 2023, 68: 25-41) (see Table 3 in document 2), the lightweight recognition model established in this embodiment has a total number of parameters reduced by more than 5.6 times, which improves the calculation efficiency and effectively reduces the consumption of computing resources.

[0093] After the model training is completed, in this embodiment, the cutting signal of the tool 3 is used as the test data, and according to step S9, the lightweight online monitoring of the tool state can be realized. Figure 6 The monitoring results are used, and the root mean square error (RMSE) between the monitoring results and the actual tool wear is calculated in turn as the accuracy evaluation index. It is found that the root mean square error RMSE between the monitoring results and the actual results is 9.98, which is almost at the same level as the recognition accuracy of thin-walled milling tool wear in document 2. However, since this embodiment takes into account the multi-domain characteristic behavior laws of thin-walled milling signals through quantitative characterization of the characteristic evolution laws, this embodiment can achieve accurate monitoring of the milling tool wear state only through a lightweight tool state recognition model. The behavior law characterization process of the proposed method has high computational efficiency, and the model training process consumes less resources, which is easier to promote to the actual thin-walled component processing engineering of aviation.

[0094] Embodiment three:

[0095] This embodiment provides a tool wear monitoring system based on multi-domain feature behavior law representation, including:

[0096] The indicator determination module is configured to: determine an indicator for evaluating the characteristic behavior, wherein the monotonicity of the characteristic vector is used as a first indicator, the normalized mutual information between the characteristic vector and the tool wear vector is used as a second indicator, and the ReLU function of the Spearman correlation coefficient between the characteristic vector and the tool wear vector is used as a third indicator;

[0097] The signal channel behavior feature characterization module is configured to: determine the behavior index corresponding to each feature vector according to the Sigmoid activation function of the first index, the second index and the third index; calculate the cumulative average value of each feature vector behavior index to characterize the behavior characteristics of the signal channel;

[0098] The tool state recognition model building module is configured to: compare the channel behavior indicators of each signal channel to determine the model input vector; train the lightweight gated recurrent unit network tool state recognition model according to the input vector to monitor the tool wear state.

[0099] Embodiment 4:

[0100] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the tool wear monitoring method based on multi-domain feature behavior law characterization described in Embodiment 1 is implemented.

[0101] Embodiment five:

[0102] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the tool wear monitoring method based on multi-domain feature behavior law characterization described in the first embodiment is implemented.

[0103] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A tool wear monitoring method based on multi-domain feature behavior law characterization, characterized in that: include: Determine an index for evaluating the characteristic behavior, wherein the monotonicity of the characteristic vector is used as a first index, the normalized mutual information between the characteristic vector and the tool wear vector is used as a second index, and the ReLU function of the Spearman correlation coefficient between the characteristic vector and the tool wear vector is used as a third index; Determine the behavior index corresponding to each feature vector according to the Sigmoid activation function of the first index, the second index and the third index; calculate the cumulative average value of each feature vector behavior index to characterize the behavior characteristics of the signal channel; Compare the channel behavior indicators of each signal channel to determine the model input vector; train a lightweight gated recurrent unit network tool state recognition model based on the input vector to monitor the tool wear state; For each cutting signal channel of each tool, the monotonicity of each feature vector is calculated cyclically as the first indicator for evaluating the feature behavior. The feature vector f ij Monotonicity MO ij It can be expressed by formula (1): MO ij =|corr(rank(f ij ),rank(t ij ))| (1) In formula (1), corr represents the calculation of Spearman rank correlation coefficient, rank represents the calculation of Spearman rank, and t ij Representation and eigenvector f ij The corresponding time series vector; For each cutting signal channel of each tool, each feature vector f is calculated cyclically ij and tool wear vector w j The normalized mutual information NI ij , as the second indicator to characterize the characteristic behavior, its expression is: In formula (2), I ij (f ij ;w j ) represents the calculation of the eigenvector f ij and the wear vector w of the jth tool j The mutual information between them, H(·) represents the information entropy of the calculated vector; The feature vector is the feature vector in the feature space of the tool cutting signal.

2. The tool wear monitoring method based on multi-domain feature behavior law characterization according to claim 1 is characterized in that: For each cutting signal channel of each tool, the monotonicity of each eigenvector, the normalized mutual information between each eigenvector and the tool wear vector, and the Spearman correlation coefficient between each eigenvector and the tool wear vector are cyclically calculated.

3. The tool wear monitoring method based on multi-domain feature behavior law characterization according to claim 1 or 2 is characterized in that: The performance index of the feature vector negatively correlated with the tool wear vector is directly set to zero through the ReLU function.

4. The tool wear monitoring method based on multi-domain feature behavior law characterization according to claim 1 is characterized in that: According to the first index, the second index and the third index corresponding to each characteristic vector, a characteristic behavior index corresponding to the cutting tool characteristic vector is obtained; The characteristic behavior indexes of all cutting tool characteristic vectors are averaged to obtain the behavior index corresponding to each characteristic vector.

5. The tool wear monitoring method based on multi-domain feature behavior law characterization according to claim 1 or 4 is characterized in that: Compare the channel behavior indicators of each signal channel, select the top two signal channels as input signal channels, and select a feature vector with the largest characteristic behavior indicator in the time domain, frequency domain and time-frequency domain of each input signal channel to form the model input vector.

6. The tool wear monitoring method based on multi-domain feature behavior law characterization according to claim 1 is characterized in that: The tool state recognition model has four layers in total.

7. The tool wear monitoring method based on multi-domain feature behavior law characterization according to claim 6 is characterized in that: The first and second layers contain several bidirectional gated recurrent units, and the third and fourth layers contain fully connected layers of several neurons.

8. A tool wear monitoring system based on multi-domain characteristic behavior law characterization, based on a tool wear monitoring method based on multi-domain characteristic behavior law characterization as claimed in any one of claims 1 to 7, characterized in that: include: The indicator determination module is configured to: determine an indicator for evaluating the characteristic behavior, wherein the monotonicity of the characteristic vector is used as a first indicator, the normalized mutual information between the characteristic vector and the tool wear vector is used as a second indicator, and the ReLU function of the Spearman correlation coefficient between the characteristic vector and the tool wear vector is used as a third indicator; The signal channel behavior feature characterization module is configured to: determine the behavior index corresponding to each feature vector according to the Sigmoid activation function of the first index, the second index and the third index; calculate the cumulative average value of each feature vector behavior index to characterize the behavior characteristics of the signal channel; The tool state recognition model building module is configured to: compare the channel behavior indicators of each signal channel to determine the model input vector; train the lightweight gated recurrent unit network tool state recognition model according to the input vector to monitor the tool wear state.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the tool wear monitoring method based on multi-domain feature behavior law characterization as described in any one of claims 1-2, 4, 6-7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the tool wear monitoring method based on multi-domain feature behavior law characterization as described in any one of claims 1-2, 4, 6-7 is implemented.

Images