A tool wear condition monitoring method based on multi-algorithm fusion

Abstract

This paper presents a tool wear condition monitoring method based on multi-algorithm fusion, belonging to the field of metal cutting monitoring technology. To address the technical problem of decreased prediction accuracy due to uneven weight distribution in existing tool wear monitoring methods, a CNN-BiLSTM-Attention model is constructed. An adaptive scaling factor is introduced at the attention mechanism layer to make the attention distribution more closely match the characteristics of the data itself. This effectively improves the feature weight imbalance problem existing in traditional attention mechanisms, balances the weight ratio of hidden features at each time step, avoids the over-amplification or weakening of key wear features, and improves the prediction accuracy of tool wear condition. It is mainly used for tool wear condition monitoring.

Description

Technical Field

[0001] This invention relates to the field of metal cutting monitoring technology, and more specifically to a method for monitoring tool wear conditions. Background Technology

[0002] With the development of aerospace, mold manufacturing, and automotive industries, the requirements for part precision are constantly increasing. Common part machining processes include turning, milling, and drilling. Among these processes, discrete-edge cutting tools (especially in milling applications) not only meet the requirements of high precision and high efficiency but also effectively address common challenges in milling, such as chip handling and chatter suppression. Therefore, they are often used to machine complex components such as aero-engine blades, impellers, and integral structural parts. Discrete-edge end mills are a special type of end mill with discrete chip flutes. In addition to integrating the advantages of low cutting force, strong chip removal capability, and low cutting temperature of wave-edge cutting tools, they also have advantages that wave-edge cutting tools do not possess, such as short manufacturing time and simple grinding. Compared with traditional cutting tools, they are better able to meet the actual production needs in today's complex machining conditions.

[0003] Existing methods for monitoring tool wear are mainly divided into two categories: direct monitoring and indirect monitoring. Direct monitoring requires removing the tool from the machining equipment for inspection, interrupting the normal machining process. Furthermore, the results are highly dependent on the operator's experience, leading to human error and subjectivity. This makes real-time monitoring of tool wear impossible, reducing production efficiency and potentially causing excessive tool wear due to delayed monitoring. Consequently, this affects the dimensional accuracy and surface quality of the machined workpiece, increasing production costs.

[0004] Indirect monitoring methods use multiple sensors to collect signals of tool wear changes. By monitoring the correlation between tool condition changes and machining process parameters, it can be used to determine whether the tool is in a normal working state. However, most of these methods suffer from a decrease in the accuracy of monitoring results due to uneven weight distribution. Summary of the Invention

[0005] To overcome the technical problem of decreased prediction accuracy caused by uneven weight distribution in existing tool wear monitoring methods, this invention provides a tool wear state monitoring method based on multi-algorithm fusion.

[0006] This invention is achieved through the following technical solution:

[0007] A tool wear condition monitoring method based on multi-algorithm fusion includes the following steps:

[0008] Step S1: Construct a tool wear dataset. Each data sample includes a tool dynamic time-domain signal and a tool wear value corresponding to the dynamic signal. The dynamic time-domain signal is a time-series signal collected during a single cutting process, including tool X, Y, and Z three-axis vibration signals and cutting force signals.

[0009] Step S2: After preprocessing the data in the original dataset, it is divided into a training set, a validation set, and a test set;

[0010] Step S3: Construct a CNN-BiLSTM-Attention model, including an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer, an attention mechanism layer, a fully connected layer, and an output layer; train the CNN-BiLSTM-Attention model using the training set to obtain the trained CNN-BiLSTM-Attention tool wear status monitoring model.

[0011] Step S4: Input the dynamic time-domain signal of the tool to be predicted into the trained CNN-BiLSTM-Attention model, and the model outputs the tool wear prediction value.

[0012] Further, step S2 includes the following steps:

[0013] (1) After filtering and smoothing the dynamic time domain signal, remove null values, outliers and duplicate values ​​from the data;

[0014] (2) Use wavelet denoising algorithm to denoise the dynamic time domain signal within the target dynamic time domain signal interval;

[0015] (3) Extract the multi-domain features of the dynamic time-domain signal, wherein the multi-domain features are composed of time-domain features, frequency-domain features and time-frequency-domain features concatenated in sequence;

[0016] (4) Reduce the multi-domain features to a preset dimension to obtain the preset dimension features of each data sample; construct a modeling dataset with the preset dimension features corresponding to each data sample and the corresponding tool wear value label, and divide the modeling dataset into a training set, a validation set and a test set.

[0017] Furthermore, the method for determining the target dynamic time-domain signal interval in step S2 is as follows:

[0018] Calculate the preset quantiles of the amplitude sequences of the vibration signal and the cutting force signal in a single cutting operation; then use the preset quantiles of the vibration signal amplitude sequence as the vibration signal threshold and the preset quantiles of the cutting force signal amplitude sequence as the cutting force signal threshold; determine the intervals where the amplitudes of both the vibration signal and the cutting force signal are higher than their respective thresholds as the target dynamic time domain signal intervals.

[0019] Furthermore, the time-domain features are calculated directly based on the dynamic time-domain signal, including the average value, standard deviation, peak value, root mean square, and impulse factor.

[0020] The frequency domain features are obtained by performing a Fourier transform on the dynamic time domain signal to convert the dynamic time domain signal into a frequency domain representation, and then calculating the maximum frequency, minimum frequency, average frequency, centroid frequency, frequency deviation, and mean square frequency.

[0021] The process of obtaining the time-frequency domain features is as follows: wavelet packet transform is performed on the dynamic time-domain signal to divide the dynamic time-domain signal into multiple non-overlapping sub-frequency bands; time-domain features and frequency-domain features are extracted in each sub-frequency band and arranged in the order of time-domain features first and frequency-domain features last to obtain the features of each sub-frequency band; then the features of each sub-frequency band are arranged according to the sub-frequency band number to obtain the time-frequency domain features.

[0022] Further, in step S2, principal component analysis is used to perform dimensionality reduction on the multi-domain features. The first ten principal components are used as the new feature space to reduce the multi-domain features to ten dimensions, thus obtaining the ten-dimensional reduced features of each data sample. The ten-dimensional reduced features of each data sample and the corresponding tool wear value labels are used to form a modeling dataset, and the modeling dataset is divided into a training set, a validation set, and a test set.

[0023] Furthermore, the convolutional neural network layer includes an input layer, two one-dimensional convolutional layers, one average pooling layer, one fully connected layer, and an output layer.

[0024] Furthermore, the kernel size of the two one-dimensional convolutional layers is 3; the kernel size of the average pooling layer is 2, and the stride is 1.

[0025] Furthermore, in the CNN-BiLSTM-Attention model, the attention layer takes the complete bidirectional temporal feature sequence output by the bidirectional long short-term memory network layer as input and performs the following operations sequentially:

[0026] (1) Calculate the basic importance score of the hidden state feature vector at each time step in the complete bidirectional temporal feature sequence; the calculation formula is:

[0027] ,

[0028] In the formula, This represents the fused hidden state feature vector of the p-th layer and the t-th time step of the BiLSTM. The basic importance score, with values ​​ranging from [0, +∞]; Represents the L2 norm; This represents the fused hidden state feature vector of the p-th layer and the t-th time step of the BiLSTM.

[0029] (2) Assign scores to each basic importance level Mapping to the (0, 1) interval yields the hidden state feature vectors at each time step. Adaptive scaling factor The calculation formula is:

[0030] ,

[0031] In the formula, Indicates the adaptive scaling factor; This represents a learnable scaling factor used to control the adaptive scaling factor. Steepness;

[0032] (3) Based on the adaptive scaling factor Calculate the hidden state feature vector at time step t. The attention score is calculated using the following formula:

[0033] ,

[0034] In the formula, Represents the hidden state feature vector at time step t. Attention score; tanh represents the hyperbolic tangent function; This indicates that the output of the hyperbolic tangent function tanh is mapped to a weight vector of scalar scores. In Indicates the transpose symbol; The weight matrix representing the attention mechanism; Indicates the bias term;

[0035] (4) Calculate the hidden state feature vector at time step t. Normalized attention weights The calculation formula is:

[0036] ,

[0037] In the formula, This represents the attention score at time step z.

[0038] (5) Based on the normalized attention weights, perform a weighted summation operation on the hidden state feature vectors at all time steps within the current training batch to obtain a comprehensive output feature containing complete global context information. The calculation formula is:

[0039] ,

[0040] In the formula, The length of the complete bidirectional temporal feature sequence; Used to identify the importance of the hidden state features at time step t;

[0041] (6) The integrated output features The input is fed into a fully connected layer, which then synthesizes and outputs the features. This is mapped to the final predicted tool wear value;

[0042] The formula for the fully connected layer is expressed as follows:

[0043] ,

[0044] In the formula, This is the weight matrix of the fully connected layer; For bias terms; This is the final output of the model, i.e., the predicted tool wear value.

[0045] Furthermore, the CNN-BiLSTM-Attention model is considered converged when any of the following conditions are met:

[0046] Condition 1: The model's loss value on the validation set has not decreased for N consecutive training epochs, where N represents the preset early stopping epoch threshold;

[0047] Condition 2: The model training rounds reach the preset maximum number of iterations M, where M is the preset maximum training rounds threshold.

[0048] Furthermore, M=80, N=10.

[0049] The beneficial effects of this invention are:

[0050] This invention constructs a CNN-BiLSTM-Attention model comprising a convolutional neural network layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a fully connected layer. In the attention mechanism layer, a learnable scaling factor is introduced to map the basic importance scores. Compared with the traditional fixed-weight attention mechanism, the attention weights optimized by the adaptive scaling factor in this invention can effectively improve the problem of unbalanced feature weight distribution in the traditional attention mechanism, balance the weight ratio of hidden features at each time step, avoid over-amplification or weakening of key wear features, and improve the accuracy of tool working state prediction. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the CNN-BiLSTM-Attention model structure in one embodiment of the method of the present invention;

[0053] Figure 2 This is a schematic diagram of the tool wear condition monitoring execution process in one embodiment of the method of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] This application provides a tool wear state monitoring method based on multi-algorithm fusion, wherein the multi-algorithm fusion refers to the fusion of Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM) network, and Attention Mechanism; the method includes the following steps:

[0056] Step S100: Construct a tool wear dataset. Each data sample includes a tool dynamic time domain signal and a tool wear value corresponding to the dynamic signal. The dynamic time domain signal is the time-series raw signal collected in a single cutting process, specifically including the tool's X, Y, and Z three-axis vibration signals and cutting force signal.

[0057] In this embodiment, a discrete-edge end mill is taken as the research object. During the CNC machining process, dynamic time-domain signals of the tool in the X, Y, and Z directions are collected in real time. During each cutting process, multiple sets of dynamic time-domain signal sequences collected by sensors correspond to a data sample. After each cutting is completed, the wear value of the tool flank is measured synchronously, and the wear value is used as the label corresponding to the data sample to construct the original dataset.

[0058] Step S200: After preprocessing the data in the original dataset, it is divided into a training set, a validation set, and a test set. Specifically, this includes the following steps:

[0059] The training set is mainly used to train the CNN-BiLSTM-Attention model; the validation set is used to evaluate whether the CNN-BiLSTM-Attention model is overfitting during training; the test set is not involved in training and parameter tuning, but is only used for the final performance test after the model training is completed, reflecting the model's performance on unknown data.

[0060] (1) After filtering and smoothing the dynamic time domain signal, identify null values, outliers and duplicate values ​​in the data;

[0061] (2) The wavelet denoising algorithm is used to denoise the dynamic time domain signal within the target dynamic time domain signal interval; the method for determining the target dynamic time domain signal interval is as follows:

[0062] Calculate the preset quantiles of the amplitude sequences of the vibration signal and the cutting force signal in a single cut (i.e., each data sample); use the preset quantiles of the vibration signal amplitude sequence as the vibration signal threshold and the preset quantiles of the cutting force signal amplitude sequence as the cutting force signal threshold; determine the interval where the amplitudes of both the vibration signal and the cutting force signal are higher than their respective thresholds as the target cutting interval, and eliminate invalid signal segments such as air cutting, feed, and retraction.

[0063] In this embodiment, the preset quantile is set to 75%.

[0064] (3) Extract the multi-domain features of the dynamic time-domain signal, wherein the multi-domain features are composed of time-domain features, frequency-domain features and time-frequency-domain features concatenated in sequence.

[0065] The time-domain features are calculated directly based on the dynamic time-domain signal and include the average value, standard deviation, peak value, root mean square, and impulse factor.

[0066] The frequency domain features are obtained by performing a Fourier transform on the dynamic time domain signal to convert the dynamic time domain signal into a frequency domain representation, and then calculating the maximum frequency, minimum frequency, average frequency, centroid frequency, frequency deviation, and mean square frequency.

[0067] The process of obtaining the time-frequency domain features is as follows: Wavelet packet transform (WPT) is performed on the dynamic time-domain signal to divide it into multiple non-overlapping sub-bands; time-domain and frequency-domain features are extracted from each sub-band, and arranged in order of time-domain features first, followed by frequency-domain features, to obtain the features of each sub-band; then, the features of each sub-band are arranged according to their sub-band numbers to obtain the time-frequency domain features. Assuming the dynamic time-domain signal is divided into Num non-overlapping sub-bands, the time-frequency domain features can be represented as: the first sub-band feature, the second sub-band feature, ..., the Num-th sub-band feature. Each sub-band feature can be represented as: mean, standard deviation, peak value, root mean square, impulse factor, maximum frequency, minimum frequency, average frequency, centroid frequency, frequency deviation, and mean square frequency.

[0068] For any data sample, the multi-domain features can be represented as: mean, standard deviation, peak value, root mean square, impulse factor, maximum frequency, minimum frequency, average frequency, centroid frequency, frequency deviation, mean square frequency, first sub-band feature, second sub-band feature, ..., Nth sub-band feature.

[0069] (4) The multi-domain features are reduced in dimensionality using principal component analysis (PCA). The first ten principal components are used as the new feature space to reduce the multi-domain features to ten dimensions, thus obtaining the ten-dimensional reduced features of each data sample. The modeling dataset is constructed using the ten-dimensional reduced features of each data sample and the corresponding tool wear value labels, and the modeling dataset is divided into training set, validation set and test set.

[0070] In this embodiment of the application, the training set accounts for 70%, and the validation set and test set each account for 15%.

[0071] Step S300: Construct a CNN-BiLSTM-Attention model, including an input layer, a convolutional neural network (CNN) layer, a bidirectional long short-term memory (BiLSTM) layer, an attention mechanism layer, a fully connected layer, and an output layer; train the CNN-BiLSTM-Attention model using the training set to obtain the trained CNN-BiLSTM-Attention tool wear status monitoring model. For the internal execution flow of the CNN-BiLSTM-Attention model, please refer to [link / reference needed]. Figure 1 .

[0072] (1) Convolutional Neural Network Structure and Execution Process

[0073] In this embodiment, the convolutional neural network includes an input layer, two one-dimensional convolutional layers, one pooling layer, one fully connected layer, and an output layer. The two one-dimensional convolutional layers are cascaded sequentially, each with a kernel size of 3. An average pooling layer with a kernel size of 2 and a stride of 1 is placed after the two convolutional layers.

[0074] The 2-layer convolutional + 1-layer pooling structure designed in this application embodiment can reduce information loss and is more suitable for processing temporal features.

[0075] The input to the CNN is a preprocessed dynamic temporal signal feature sequence, with an overall dimensionality format of [batch size, feature dimension, sequence length]. In this embodiment, the batch size is set to 128, the feature dimension to 10, and the sequence length to 24. Each sample corresponds to a two-dimensional temporal feature matrix, where rows represent temporal steps (i.e., sequence length) and columns represent feature dimensions. The convolution kernel slides across the two-dimensional data matrix, performing convolution operations. The calculation formula is as follows:

[0076] (1)

[0077] In the formula, Indicates the convolutional layer at the 1st... The feature output value at the kth time step and the kth output convolutional channel; This represents the first time sequence feature matrix within the input of the convolutional layer. The time step, the first The feature values ​​corresponding to each input feature channel; Represents the input time-series feature matrix; Indicates the window offset index of the convolution kernel in the time dimension (in the direction of the input matrix); uppercase. This represents the total number of channels, or columns, of the input two-dimensional matrix; lowercase. This represents the channel number, or column index, of the input two-dimensional time-series feature matrix; uppercase. Indicates the convolution kernel; lowercase. Indicates the number of output channels; Indicates the first The bias corresponding to the output channel.

[0078] The pooling layer employs average pooling, and the "smoothing effect" can reduce the impact of local noise on the features, making the output features more stable.

[0079] The fully connected layer is mainly used to integrate the local features extracted by the convolutional and pooling layers. In the fully connected layer, each neuron establishes a connection with all neurons in the previous layer to achieve the association of global information.

[0080] The mathematical expression of the execution process of the fully connected layer is shown in Equation (2):

[0081] (2)

[0082] in, It is the number of input neurons, i.e., the dimension of the output features of the pooling layer after flattening. This represents the output of the i-th neuron; Indicates the connection of the first Input neuron and the first The weights of each output neuron; Indicates the first The input value of each neuron, Indicates the first Bias terms for each neuron.

[0083] After the above processing, the output layer outputs time-series feature data [batch size = 128, feature dimension = 64, sequence length = 24].

[0084] (2) Structure and execution process of bidirectional long short-term memory network

[0085] The time-series feature data is converted into a dimensional feature format and then input into a two-layer stacked BiLSTM network to obtain a complete bidirectional time-series feature sequence.

[0086] The dual-layer stacked BiLSTM network consists of two cascaded layers of bidirectional long short-term memory (BiLSTM) networks. The bidirectional temporal feature sequence output by each BiLSTM layer is composed of multiple time-step hidden vectors.

[0087] In this embodiment, the temporal feature data output by the CNN layer is converted to [batch size=128, sequence length=24, feature dimension=64] by dimensionality transpose.

[0088] Specifically, the execution flow inside a BiLSTM network can be represented by the following formula:

[0089] (3)

[0090] (4)

[0091] (5)

[0092] In the formula, This represents the hidden state of the forward LSTM at time step t. This represents the hidden state of the forward LSTM at time step t-1; This represents the hidden state of the inverse LSTM at time step t. This represents the hidden state at time step (t-1) of the LSTM. This represents the input feature at time step t, which comes from the temporal feature data output by the CNN; , This represents the forward LSTM gating weights, which control the influence of the input on the previous state. , This represents the inverse LSTM gating weights, which control the influence of the input on the next state. , The fusion weights represent the forward and backward hidden states; , These represent the cell states at time step t of the forward and backward LSTM, respectively, used for long-term memory of key feature changes during tool wear, capturing wear trends and filtering noise; Let represent the fused hidden state feature vector of the p-th layer and the t-th time step of the BiLSTM.

[0093] The feature dimension format of the complete bidirectional temporal feature sequence output by BiLSTM is [batch size = 128, sequence length = 24, feature dimension = 128].

[0094] This step transforms the temporal feature data output from the CNN layer through dimensionality transposition to adapt it to the BiLSTM input format. The transposed temporal feature data is then input into a two-layer stacked BiLSTM network. The bottom layer first performs bidirectional encoding on the input temporal feature data, mining the basic contextual correlation features of the temporal feature data and outputting the hidden state sequence at each time step. The top layer takes the hidden state sequence at each time step output by the bottom layer as input and further performs deep bidirectional temporal feature fusion, completing the feature information transfer and deepening expression layer by layer, and finally outputting a complete bidirectional temporal feature sequence that integrates global temporal contextual dependency information.

[0095] (3) The structure and execution process of the attention mechanism

[0096] Because neural networks (CNNs) typically assign equal weights to hidden state features across all time steps when processing temporal data, crucial temporal information strongly correlated with tool wear is often overlooked, making it difficult for the model to focus on truly important cutting state changes. To address this issue, this application introduces an Additive Attention mechanism. Through a learnable weight matching mechanism, it adaptively assigns importance weights to each time step, guiding the model to focus on the temporal segments most critical to tool wear prediction. This attention mechanism enables the model to differentiate its focus across different cutting moments, improving the efficiency of extracting wear-sensitive features and thus optimizing prediction performance.

[0097] The complete bidirectional temporal feature sequence is input into the attention layer, and the following operations are performed sequentially:

[0098] (1) Calculate the basic importance score of the hidden state feature vector at each time step in the complete bidirectional temporal feature sequence. The calculation formula is as follows:

[0099] (6)

[0100] In the formula, This represents the fused hidden state feature vector of the p-th layer and the t-th time step of the BiLSTM. The basic importance score, with values ​​ranging from [0, +∞]; It represents the L2 norm (Euclidean modulus).

[0101] (2) The basic importance score Mapping to the (0, 1) interval yields the adaptive scaling factor. In this implementation, the Sigmoid function is used to convert the modulus... Mapping to an adaptive scaling factor with values ​​in the range (0, 1) To enhance the distinguishability of features;

[0102] The adaptive scaling factor The expression is:

[0103] (7)

[0104] In the formula, This represents the adaptive scaling factor, obtained from the Sigmoid function, used to enhance the discriminative power of attention scores; This represents a learnable scaling factor that controls the adaptive scaling factor. The steepness.

[0105] (3) Based on the adaptive scaling factor Calculate the hidden state feature vector at time step t. The attention score is calculated using the following formula:

[0106] (8)

[0107] In the formula, Represents the hidden state feature vector at time step t. Attention score; tanh represents the hyperbolic tangent function; This indicates that the output of the hyperbolic tangent function tanh is mapped to a weight vector of scalar scores. In Indicates the transpose symbol; The weight matrix represents the attention mechanism, with the input dimension being the hidden state feature vector. The dimension of the output dimension is The dimension; This indicates the bias term.

[0108] (4) Calculate the hidden state feature vector at time step t. Normalized attention weights The calculation formula is:

[0109] (9)

[0110] In the formula, This represents the attention score at the z-th time step.

[0111] (5) Based on the normalized attention weights, perform weighted summation of features at all time steps within the current training batch to achieve adaptive fusion of temporal features at different time steps, and obtain comprehensive output features containing complete global context information. The calculation formula is:

[0112] (10)

[0113] In the formula, The length of the complete bidirectional temporal feature sequence; Used to identify the importance of the hidden state features at time step t.

[0114] (6) The integrated output features The input is fed into a fully connected layer, which then synthesizes and outputs the features. This is mapped to the final predicted tool wear value.

[0115] The formula for the fully connected layer is expressed as follows:

[0116] (11)

[0117] In the formula, This is the weight matrix of the fully connected layer, with dimension representing the combined output features. The dimension; For bias terms, dimensions and weight matrices same; This is the final output of the model, i.e., the predicted tool wear value.

[0118] The CNN-BiLSTM-Attention model is considered converged when any of the following conditions are met:

[0119] Condition 1: The model's loss value on the validation set has not decreased for N consecutive training epochs, where N represents the preset early stopping epoch threshold;

[0120] Condition 2: The model training rounds reach the preset maximum number of iterations M, where M is the preset maximum training rounds threshold.

[0121] In this embodiment, N=10 and M=80.

[0122] Step S400: Input the dynamic time-domain signal of the tool to be predicted into the trained CNN-BiLSTM-Attention model, and the model outputs the tool wear prediction value.

[0123] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0124] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A tool wear state monitoring method based on multi-algorithm fusion, characterized in that, Includes the following steps: Step S1: Construct a tool wear dataset. Each data sample includes a tool dynamic time-domain signal and a tool wear value corresponding to the dynamic signal. The dynamic time-domain signal is a time-series signal collected during a single cutting process, including tool X, Y, and Z three-axis vibration signals and cutting force signals. Step S2: After preprocessing the data in the original dataset, it is divided into a training set, a validation set, and a test set; Step S3: Construct a CNN-BiLSTM-Attention model, including an input layer, a convolutional neural network layer, a bidirectional long short-term memory network layer, an attention mechanism layer, a fully connected layer, and an output layer; train the CNN-BiLSTM-Attention model using the training set to obtain the trained CNN-BiLSTM-Attention model. Step S4: Input the dynamic time-domain signal of the tool to be predicted into the trained CNN-BiLSTM-Attention model, and the model outputs the tool wear prediction value.

2. The tool wear state monitoring method based on multi-algorithm fusion according to claim 1, characterized in that, Step S2 includes the following steps: (1) After filtering and smoothing the dynamic time domain signal, remove null values, outliers and duplicate values ​​from the data; (2) Use wavelet denoising algorithm to denoise the dynamic time domain signal within the target dynamic time domain signal interval; (3) Extract the multi-domain features of the dynamic time-domain signal, wherein the multi-domain features are composed of time-domain features, frequency-domain features and time-frequency-domain features concatenated in sequence; (4) Reduce the multi-domain features to a preset dimension to obtain the preset dimension features of each data sample; construct a modeling dataset with the preset dimension features corresponding to each data sample and the corresponding tool wear value label, and divide the modeling dataset into a training set, a validation set and a test set.

3. The tool wear state monitoring method based on multi-algorithm fusion according to claim 2, characterized in that, The method for determining the target dynamic time-domain signal interval in step S2 is as follows: Calculate the preset quantiles of the amplitude sequences of the vibration signal and the cutting force signal in a single cutting operation; then use the preset quantiles of the vibration signal amplitude sequence as the vibration signal threshold and the preset quantiles of the cutting force signal amplitude sequence as the cutting force signal threshold; determine the intervals where the amplitudes of both the vibration signal and the cutting force signal are higher than their respective thresholds as the target dynamic time domain signal intervals.

4. The tool wear state monitoring method based on multi-algorithm fusion according to claim 2, characterized in that, The time-domain features are calculated directly based on the dynamic time-domain signal, including the average value, standard deviation, peak value, root mean square, and impulse factor. The frequency domain features are obtained by performing a Fourier transform on the dynamic time domain signal to convert the dynamic time domain signal into a frequency domain representation, and then calculating the maximum frequency, minimum frequency, average frequency, centroid frequency, frequency deviation, and mean square frequency. The process of obtaining the time-frequency domain features is as follows: wavelet packet transform is performed on the dynamic time-domain signal to divide the dynamic time-domain signal into multiple non-overlapping sub-frequency bands; time-domain features and frequency-domain features are extracted in each sub-frequency band and arranged in the order of time-domain features first and frequency-domain features last to obtain the features of each sub-frequency band; then the features of each sub-frequency band are arranged according to the sub-frequency band number to obtain the time-frequency domain features.

5. The tool wear state monitoring method based on multi-algorithm fusion according to claim 2, characterized in that, Step S2 uses principal component analysis to perform dimensionality reduction on the multi-domain features. The first ten principal components are used as the new feature space to reduce the multi-domain features to ten dimensions, resulting in ten-dimensional reduced features for each data sample. The ten-dimensional reduced features corresponding to each data sample and the corresponding tool wear value labels are used to form a modeling dataset, which is then divided into a training set, a validation set, and a test set.

6. The tool wear condition monitoring method based on multi-algorithm fusion according to claim 1, characterized in that, The convolutional neural network layer includes an input layer, two one-dimensional convolutional layers, one average pooling layer, one fully connected layer, and an output layer.

7. The tool wear condition monitoring method based on multi-algorithm fusion according to claim 6, characterized in that, The kernel size of the two one-dimensional convolutional layers is 3; the kernel size of the average pooling layer is 2, and the stride is 1.

8. The tool wear condition monitoring method based on multi-algorithm fusion according to claim 7, characterized in that, In the CNN-BiLSTM-Attention model, the attention layer takes the complete bidirectional temporal feature sequence output by the bidirectional long short-term memory network layer as input and performs the following operations sequentially: (1) Calculate the basic importance score of the hidden state feature vector at each time step in the complete bidirectional temporal feature sequence; The calculation formula is: ， In the formula, This represents the fused hidden state feature vector of the p-th layer and the t-th time step of the BiLSTM. The basic importance score, with values ​​ranging from [0, +∞]; Represents the L2 norm; (2) Assign scores to each basic importance level Mapping to the (0, 1) interval yields the hidden state feature vectors at each time step. Adaptive scaling factor The calculation formula is: ， In the formula, Indicates the adaptive scaling factor; This represents a learnable scaling factor used to control the adaptive scaling factor. Steepness; (3) Based on the adaptive scaling factor Calculate the hidden state feature vector at time step t. The attention score is calculated using the following formula: ， In the formula, Represents the hidden state feature vector at time step t. Attention score; tanh represents the hyperbolic tangent function; This indicates that the output of the hyperbolic tangent function tanh is mapped to a weight vector of scalar scores. In Indicates the transpose symbol; The weight matrix representing the attention mechanism; Indicates the bias term; (4) Calculate the hidden state feature vector at time step t. Normalized attention weights The calculation formula is: ， In the formula, This represents the attention score at time step z; (5) Based on the normalized attention weights, perform a weighted summation operation on the hidden state feature vectors at all time steps within the current training batch to obtain a comprehensive output feature containing complete global context information. The calculation formula is: ， In the formula, The length of the complete bidirectional temporal feature sequence; Used to identify the importance of the hidden state features at time step t; (6) The integrated output features The input is fed into a fully connected layer, which then synthesizes and outputs the features. This is mapped to the final predicted tool wear value; The formula for the fully connected layer is expressed as follows: ， In the formula, This is the weight matrix of the fully connected layer; For bias terms; This is the final output of the model, i.e., the predicted tool wear value.

9. The tool wear condition monitoring method based on multi-algorithm fusion according to claim 8, characterized in that, The CNN-BiLSTM-Attention model is considered converged when any of the following conditions are met: Condition 1: The model's loss value on the validation set has not decreased for N consecutive training epochs, where N represents the preset early stopping epoch threshold; Condition 2: The model training rounds reach the preset maximum number of iterations M, where M is the preset maximum training rounds threshold.

10. The tool wear condition monitoring method based on multi-algorithm fusion according to claim 9, characterized in that, M=80, N=10.

Images