Cutter wear state prediction method based on multi-source feature fusion

Abstract

The invention provides a tool wear state prediction method and device based on multi-source feature fusion, and the method comprises the following steps: obtaining various tool wear state data to construct a multi-source feature data set, and carrying out the preprocessing of the multi-source feature data set; performing deep fusion feature learning on the preprocessed multi-source feature data set by using a multi-modal fusion auto-encoder AE to generate a fusion feature data set; performing feature extraction on the fused feature data set by using a neural network to generate a feature sequence, and inputting the feature sequence into a multi-source gating prediction model WGX for training; and utilizing the trained multi-source gating prediction model WGX to predict the wear state of the tool to be detected. According to the method, the multi-mode fusion AE is used for fusing the multi-source features of the tool wear state, key information is effectively extracted from data, a set of fusion features with more representativeness are generated, a prediction model capable of comprehensively reflecting the tool wear state is constructed, and prediction precision and generalization ability are improved.

Description

Technical Field

[0001] The present invention relates to the field of tool wear state prediction, and in particular to a tool wear state prediction method and device based on multi-source feature fusion. Background Art

[0002] In modern manufacturing, cutting tools are key components of CNC machine tools, and their wear status directly determines the machining accuracy, production efficiency and production cost of workpieces. With the deepening of the concepts of Industry 4.0 and intelligent manufacturing, accurate prediction and real-time monitoring of cutting tool wear status have become an important part of realizing intelligent and automated manufacturing processes.

[0003] At present, the prediction methods of cutting tool wear status are mainly divided into two categories: offline prediction and online prediction. Offline prediction mainly detects the wear amount of the tool after cutting manually, and uses historical data and empirical formulas to predict the future wear status of the tool. Although this method is simple and easy to implement, it has problems such as low prediction accuracy and poor real-time performance, and cannot meet the needs of modern manufacturing for high-precision and high-efficiency processing. The online prediction method collects dynamic working parameters in the cutting process in real time, such as cutting force, cutting temperature, vibration signal, etc., and uses signal processing, feature extraction and artificial intelligence algorithms to predict the current wear status of the tool. This method has the advantages of strong real-time performance and high prediction accuracy, and has become a hot topic in current research.

[0004] However, online prediction methods also face some challenges. First, the cutting process is a complex physical process involving the interaction of multiple parameters and variables. A single sensor signal often cannot fully reflect the wear state of the tool. Second, different processing materials and cutting conditions have different effects on tool wear. It is necessary to comprehensively consider multiple factors to accurately predict the wear state of the tool. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method and device for predicting tool wear status based on multi-source feature fusion. By fusing multi-source features, a prediction model that can comprehensively reflect the tool wear status is constructed, thereby improving prediction accuracy and generalization ability.

[0006] To achieve the above-mentioned objectives, on the one hand, the present invention provides a tool wear state prediction method based on multi-source feature fusion, comprising the following steps: acquiring a variety of tool wear state data to construct a multi-source feature data set, and preprocessing the multi-source feature data set; using a multimodal fusion autoencoder AE to perform deep fusion feature learning on the preprocessed multi-source feature data set to generate a fused feature data set; using a neural network to extract features from the fused feature data set to generate a feature sequence, and inputting the feature sequence into a multi-source gated prediction model WGX for training; and using the trained multi-source gated prediction model WGX to predict the wear state of the tool to be detected.

[0007] The multi-source gating prediction model WGX is a fusion model of the whale optimization algorithm WOA, the gated recurrent unit GRU, the connection fusion layer MLP and XGBoost.

[0008] The plurality of tool wear state data include vibration signal data, sound signal data and temperature signal data.

[0009] The method of using the multimodal fusion autoencoder AE to perform deep fusion feature learning on the preprocessed multi-source feature dataset to generate a fusion feature dataset also includes the following steps: The encoding process from the input layer to the hidden layer is calculated as follows (1): ; The decoding process from the hidden layer to the output layer is calculated as follows (2): ; in, represents the output of the encoding process, represents the output of the decoding process, represents the Sigma activation function;Sigma activation function; , , , represents the weight matrix of the fully connected layer, , represents the bias term, X represents the vibration signal data, Y represents the sound signal data, and Z represents the temperature signal data.

[0010] The step of inputting the feature sequence into the multi-source gated prediction model WGX for training also includes the following steps: using the WOA-GRUA model to model and predict the feature sequence to obtain a first prediction result, wherein the WOA-GRUA model is a fusion model of the gated recurrent unit GRU enhanced by the attention mechanism and the whale optimization algorithm WOA; using the XGBoost to model and predict the feature sequence to obtain a second prediction result; fusing the first prediction result and the second prediction result through the connection fusion layer MLP to generate a third prediction result; and binarizing the third prediction result through a second activation function to generate a final prediction result.

[0011] The WOA-GRUA model is used to model and predict the feature sequence to obtain a first prediction result, including: calculating an attention score, for the output of the gated recurrent unit GRU hidden state at each time step, calculating its value score through a fully connected layer, and using a first activation function to ensure that the score is between -1 and 1; calculating an attention weight coefficient, using a second activation function to convert the score into a probability distribution, ensuring that the sum of all weight coefficients is 1, and each weight coefficient represents the relative importance of the gated recurrent unit GRU hidden state output at the corresponding time step in the sequence; calculating the output of the attention layer, and using the weight coefficient and the corresponding gated recurrent unit GRU hidden state output to realize the output of the attention layer.

[0012] The method of fusing the first prediction result and the second prediction result through the connection fusion layer MLP to generate a third prediction result also includes the following steps: using a splicing function to splice the first prediction result and the second prediction result to generate a fused feature vector; and inputting the fused feature vector into the connection fusion layer MLP to generate the third prediction result.

[0013] The calculation formula of the fused feature vector is as follows (8): , Among them, F is the fused feature vector, A is the first prediction result, B is the second prediction result, and concat is the concatenation function.

[0014] The connection fusion layer MLP includes: The first layer receives the fused feature vector F, and the calculation formula is as follows (9): , in, is the output of the first layer, is the weight matrix of the first layer, is the bias term of the first layer, is the activation function; The second layer, the calculation formula is as follows (10): (10), in, is the output of the second layer, is the weight matrix of the second layer, is the bias term of the second layer; The third layer outputs the third prediction result, and the calculation formula is as follows (11): (11), Wherein, y is the third prediction result, W out is the weight matrix of the third layer, b out is the bias term of the third layer, The method of using the trained multi-source gating prediction model WGX to predict the wear state of the tool to be detected also includes: using accuracy and ROC curve as evaluation criteria, analyzing the wear state of the tool to be detected according to the prediction result, and providing a wear trend and remaining life prediction.

[0015] On the other hand, the present invention also provides a tool wear state prediction device based on multi-source feature fusion, including: a multi-source feature data set construction module, used to acquire a variety of tool wear state data to construct a multi-source feature data set, and pre-process the multi-source feature data set; a fusion feature data set formation module, used to use a multimodal fusion autoencoder AE to perform deep fusion feature learning on the pre-processed multi-source feature data set to generate a fusion feature data set; a prediction model construction module, used to use a neural network to extract features from the fused feature data set, generate a feature sequence, and input the feature sequence into a multi-source gated prediction model WGX for training; a prediction module, used to use the trained multi-source gated prediction model WGX to predict the wear state of the tool to be detected.

[0016] The prediction model construction model includes: a first prediction result generation module, which is used to model and predict the feature sequence using the WOA-GRUA model to obtain a first prediction result, wherein the WOA-GRUA model is a model of the gated recurrent unit GRU enhanced by the attention mechanism and the whale optimization algorithm WOA fusion; a second prediction result generation module, which is used to model and predict the feature sequence using the XGBoost to obtain a second prediction result; a third prediction result generation module, which is used to fuse the first prediction result and the second prediction result through the connection fusion layer MLP to generate a third prediction result; a final prediction result generation module, which is used to classify the third prediction result through a second activation function to generate a final prediction result.

[0017] The first prediction result generation module is also used for the WOA-GRUA model to obtain the first prediction result using two gating devices of the gated recurrent unit GRU enhanced by the attention mechanism, including: an attention score calculation unit, which is used to calculate the value score of the output of the hidden state of the gated recurrent unit GRU at each time step through a fully connected layer, and use a first activation function to ensure that the score is between -1 and 1; an attention weight coefficient calculation unit, which is used to convert the score into a probability distribution using a second activation function to ensure that the sum of all weight coefficients is 1, and each weight coefficient represents the relative importance of the hidden state output of the gated recurrent unit GRU at the corresponding time step in the sequence; an attention layer output calculation unit, which is used to use the weight coefficient and the corresponding gated recurrent unit GRU hidden state output to realize the output of the attention layer.

[0018] The third prediction result generation module also includes: a splicing unit, which is used to splice the first prediction result and the second prediction result using a splicing function to generate a fused feature vector; and a generation unit, which is used to input the fused feature vector into the connection fusion layer MLP to generate the third prediction result.

[0019] It can be seen from the above scheme that the advantages of the present invention are: The autoencoder AE model is used to fuse the multi-source features of the tool wear status, effectively extract key information from the data, and generate a set of more representative fusion features to construct a prediction model that can comprehensively reflect the tool wear status, thereby improving the prediction accuracy and generalization ability.

[0020] The gated recurrent unit GRU enhanced by the attention mechanism can better assign attention weight parameters to each historical moment; the gated recurrent unit GRU optimized by WOA for feature extraction and prediction can better extract key information from the features after multi-source fusion, and improve the prediction performance and robustness of the tool wear state prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flow chart of a tool wear state prediction method based on multi-source feature fusion provided by an embodiment of the present invention; Figure 2 Another flow chart of a tool wear state prediction method based on multi-source feature fusion provided by an embodiment of the present invention; Figure 3 It is the structural schematic diagram of the tool life test bench; Figure 4 This is the architecture diagram of the multimodal fusion autoencoder AE; Figure 5 This is the architecture diagram of the multi-source gating prediction model WGX; Figure 6 for Figure 1 Flow chart of step S30; Figure 7 This is the structural diagram of the GRU-A model; Figure 8 A schematic diagram of the structure of a tool wear state prediction device based on multi-source feature fusion provided by another embodiment of the present invention; Fig. 9 for Figure 8 Schematic diagram of the module structure formed by fusion feature data set; Fig.10 for Figure 8 Schematic diagram of the prediction model building module structure; Fig.11 for Fig.10 A schematic diagram of the structure of the first prediction result generation module; Fig.12 for Fig.10 The third prediction result generation module structure diagram; Wherein, the reference numerals are: 1-Tool wear state prediction method based on multi-source feature fusion; 10-Tool life test bench; 11-Vertical machining center; 110-vibration sensor; 111-end milling cutter; 112-workpiece; 12-Universal tool microscope; 13-QMS3D measurement software; 2-Tool wear state prediction device based on multi-source feature fusion; 20-Multi-source feature dataset construction module; 21- Fusion feature data set formation module; 210- coding unit; 211-decoding unit; 22-Prediction model building module; 220-first prediction result generating module; 221-second prediction result generating module; 222-third prediction result generating module; 223-final prediction result generation module; 2200-attention score calculation unit; 2201-calculate attention weight coefficient unit; 2202-calculate the output unit of the attention layer; 2220-Splicing unit; 2221-Generation unit; 23- prediction module; S10~S40, S300~S303-steps. DETAILED DESCRIPTION

[0022] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments to further understand the purpose, solution and effect of the present invention, but it is not intended to limit the scope of protection of the claims attached to the present invention.

[0023] References to "an embodiment", "another embodiment", "this embodiment", etc. in the specification refer to the embodiment being described which may include specific features, structures or characteristics, but not every embodiment must include these specific features, structures or characteristics. In addition, such expressions do not refer to the same embodiment. Furthermore, when describing specific features, structures or characteristics in conjunction with an embodiment, whether or not there is an explicit description, it is indicated that incorporating such features, structures or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0024] Certain words are used in the specification and subsequent claims to refer to specific components or parts. People of ordinary skill in the art should understand that technical users or manufacturers can refer to the same component or part with different nouns or terms. This specification and claims do not use differences in names as a way to distinguish components or parts, but use differences in the functions of components or parts as the criteria for distinction. "Including" and "including" mentioned throughout the specification and claims are open-ended terms and should be interpreted as "including but not limited to". In addition, the word "connected" here includes any direct and indirect electrical connection means. Indirect electrical connection means include connection through other devices.

[0025] When predicting tool wear status, online prediction methods have the following problems. First, the cutting process is a complex physical process involving the interaction of multiple parameters and variables, and a single sensor signal often cannot fully reflect the tool wear status. Second, different processing materials and cutting conditions have different effects on tool wear, and multiple factors need to be considered comprehensively to accurately predict the tool wear status.

[0026] In order to solve the above problems, this paper proposes a tool wear state prediction method based on multi-source feature fusion. This method collects multiple sensor signals in the cutting process, including three vibration signals of cutting force, cutting temperature, sound signals, etc., and uses signal processing technology and feature extraction algorithms to extract key features reflecting the tool wear state. Then, by fusing these multi-source features, a prediction model that can comprehensively reflect the tool wear state is constructed. This method can make full use of multi-source information in the cutting process, improve prediction accuracy and generalization ability, and provide a reliable basis for real-time monitoring and replacement of cutting tools.

[0027] The present invention aims to provide a tool wear state prediction method based on multi-source feature fusion, which can realize accurate prediction and real-time monitoring of tool wear state by collecting multiple sensor signals in the cutting process and using advanced signal processing and feature extraction algorithms. This method can not only improve the precision and efficiency of cutting processing and reduce production costs, but also provide strong technical support for the intelligence and automation of the manufacturing industry.

[0028] Figure 1 or Figure 2 A flowchart of a tool wear state prediction method based on multi-source feature fusion disclosed in an embodiment of the present invention includes the following steps: S10: Acquire multiple tool wear state data to construct a multi-source feature data set, and pre-process the multi-source feature data set; S20: Use the multimodal fusion autoencoder AE to perform deep fusion feature learning on the preprocessed multi-source feature dataset to generate a fusion feature dataset; S30: Use a neural network to extract features from the fused feature data set, generate a feature sequence, and input the feature sequence into the multi-source gating prediction model WGX for training; S40: Use the trained multi-source gating prediction model WGX to predict the wear state of the tool to be detected.

[0029] In step S10: The tool life test bench 10 is used to collect tool wear status data. A CNC tool is run from new to scrapped. During each milling process, vibration sensor signals, acoustic signals, and temperature signals are collected, and a high-power microscope is used to measure the tool wear value. The collection device is the tool life test bench 10, such as Figure 3 As shown, the tool life test bench 10 includes a vertical machining center 11, a universal tool microscope 12, and QMS3D measurement software 13. The vertical machining center 11 includes a vibration sensor 110, an end mill 111, and a workpiece 112. The universal tool microscope 12 is connected to the QMS3D measurement software 13 by signal. A brand new tool is placed on the workpiece 112, and then the end mill 111 is put down to run the tool from brand new to scrapped. During this process, the vibration signal generated is transmitted by the vibration sensor 110 to the QMS3D measurement software 13, and the universal tool microscope 12 is used to measure the wear area on the surface of the worn tool.

[0030] Preprocessing the multi-source feature data set specifically includes: cleaning the collected original data set and unifying the format, and then using time windowing to preprocess the data. The present invention sets the feature data of the first three time points to predict the trajectory information of the current time each time, that is, time_step=8, so that the vibration signal, acoustic signal and temperature signal can be cut into multiple windows of fixed length.

[0031] In step S20: like Figure 4 As shown in the figure, it is the architecture diagram of the multimodal fusion autoencoder AE. The multimodal fusion autoencoder AE is a multi-input autoencoder. It uses the data itself as supervision based on the back propagation algorithm and optimization method to guide the neural network to learn the mapping relationship, thereby obtaining a reconstructed output result. Figure 4 As shown in the figure, the algorithm model of AE consists of two main parts: encoder and decoder. The encoder maps the three high-dimensional signal data of vibration signal, acoustic signal and temperature signal from the input layer to a low-dimensional feature space of the hidden layer, and encodes them into low-dimensional latent variables, so that the neural network can learn the most informative features; the decoder restores the latent variables of the hidden layer to the initial dimension, so as to obtain the feature fusion data set, which is output from the output layer. The process of multi-source data encoding is as follows: The encoding process of the preprocessed multi-source data X (vibration signal), Y (acoustic signal), and Z (temperature signal) from the input layer to the hidden layer is shown in formula (1): (1); The decoding process from the hidden layer to the output layer is shown in formula (2): (2); in, represents the output of the encoding process, represents the output of the decoding process, where represents the output of the encoding process, represents the output of the decoding process, represents the Sigma activation function;Sigma activation function; , , , represents the weight matrix of the fully connected layer, , represents the bias term, X represents the vibration signal data, Y represents the sound signal data, and Z represents the temperature signal data.

[0032] In step S30: Use neural network to extract features from feature fusion datasets and generate feature sequences.

[0033] First, the convolutional neural network (CNN) performs a convolution operation on the input image through a sliding convolution kernel (filter) to extract local features. In multimodal signal input, the convolution operation captures the local features of each signal type. Secondly, the pooling layer selects the maximum value or average value in the local area to reduce the amount of calculation and enhance the robustness of the feature. Finally, the extracted features are integrated through the fully connected layer to generate the final feature sequence.

[0034] The feature sequence is input into the multi-source gating prediction model WGX for training. The multi-source gating prediction model WGX is a fusion model of the whale optimization algorithm WOA, the gated recurrent unit GRU, the connection fusion layer MLP and XGBoost (extreme steep boost). The architecture diagram of the multi-source gating prediction model WGX of the present invention is as follows: Figure 5 shown.

[0035] like Figure 6 As shown, step S30 also includes: S300: Modeling and predicting the feature sequence using the WOA-GRUA model to obtain a first prediction result; S301: Use XGBoost to model and predict the feature sequence to obtain a second prediction result; S302: The first prediction result and the second prediction result are fused through a connection fusion layer MLP to generate a third prediction result; S303: Binary classify the third prediction result through the second activation function to generate a final prediction result.

[0036] Specifically, in step S300, a branch of the multi-source gating prediction model WGX uses the WOA-GRUA model to model and predict the feature sequence to obtain a prediction result A.

[0037] The WOA-GRUA model is a model that combines the attention mechanism-enhanced gated recurrent unit GRU and the whale optimization algorithm WOA.

[0038] Specifically, the WOA-GRUA model uses the two gating devices of the gated recurrent unit GRU enhanced by the attention mechanism to obtain the prediction result A (i.e., the first prediction result), as shown in Figure 7As shown in the figure, the gated recurrent unit GRU includes a reset gate, an update gate, and a candidate hidden state, where the reset gate determines the extent to which the new input is combined with the previous hidden state, and this process is controlled by a sigmoid function and a point-by-point multiplication operation; the update gate determines the extent to which the old hidden state information is passed to the new state, which is generated by a sigmoid function; the candidate hidden state is a new hidden state candidate value, which combines the current input and the previous hidden state, and is generated by a hyperbolic tangent function tanh.

[0039] The mathematical expression of the final output of the hidden state of the gated recurrent unit GRU is as follows (3): (3), Among them, H t is at the time step t The hidden state output of z t is the update gate at time step t The output of , which determines the extent to which the previous hidden state is retained H t−1 . h t is the candidate hidden state at time step t The output of is calculated based on the current input and the previous hidden state.

[0040] In this process, the whale optimization algorithm WOA is used to tune the weights of the GRU-A model (i.e., the attention mechanism enhanced gated recurrent unit GRU) and bias the parameters of the GRU-A model. The structural diagram of the GRU-A model is shown in the figure. Figure 7 As shown in the figure, the GRU-A model includes the Attention layer and the Gated Recurrent Unit (GRU). The GRU-A model processing process has three parts: Calculate the attention score: For the output of the hidden state of the GRU-A model at each time step, calculate its value score through the fully connected layer, and use the hyperbolic tangent function tanh as the first activation function to ensure that the score is between -1 and 1. The calculation formula is shown in formulas (4) and (5): (4), (5), in Is a hidden state The score represents the value score of each input entering the Attention layer (i.e., the attention layer). It represents the weight matrix of the fully connected layer. represents the bias vector of the fully connected layer, and tanh is the hyperbolic tangent function, which is used to compress the score into the range of [-1, 1]; yes The attention weight coefficient indicates the importance of the corresponding historical moment.

[0041] Calculate the attention weight coefficient: Use the second activation function (such as the Softmax function) to convert the score into a probability distribution, ensuring that the sum of all weight coefficients is 1, and each weight coefficient represents the relative importance of the hidden state output of the GRU-A model at the corresponding time step in the sequence; Calculate the output of the attention layer: Use the weight coefficient and the corresponding GRU-A model hidden state output to realize the output of the attention layer. The calculation formula is shown in formula (6): (6), in, is the output of the Attention layer (i.e., the attention layer), y is the final output, is the activation function, usually using sigmoid function or ReLU function, is the weight matrix of the fully connected layer. is the bias vector of the fully connected layer.

[0042] When training the GRU-A model, the L1 regularization term and the Adam optimizer are used for model training. The mathematical expression of the L1 regularization term is: (7), in, is the regularization parameter, which is used to control the regularization strength; is the weight of the GRU-A model; n is the number of parameters; is the true label vector; is the input feature matrix.

[0043] In the Adam optimizer, the number of whales is set to 100, the initialization of the objective function is set to 1e10, and the maximum number of iterations is set to 100. The Whale Swarm Algorithm (WOA) optimizes the connection weights of the neural network, and the position vectors of individuals in the whale group correspond to the weights and thresholds during the network iteration process. Iterative training continues until the expected error or the number of iterations reaches the standard.

[0044] In order to ensure the model’s predictive power for samples outside the training set and maintain high accuracy on the test sets of different depths in the deep sea layer, the L1 regularization term retains all variables but reduces the magnitude of the feature variables, and the regularization parameter balances the goal of fitting training and keeping the parameter value with smaller feature contribution small.

[0045] The Adam optimizer is used to achieve simple and efficient calculation of the model. The weight and bias parameters are tuned through model training, and the prediction result sequence is obtained through the test set.

[0046] In step S301 , another branch of the multi-source gating prediction model WGX uses XGBoost (extreme steep boosting) to model and predict the feature sequence to obtain a prediction result B (ie, a second prediction result).

[0047] XGBoost modeling is implemented through the gradient boosting algorithm, which starts with an initial simple tree. In each round, the newly added tree will fit the residual according to the error of the previous round of model, and gradually optimize the error. Each tree will adjust the prediction results based on the current model. Through multiple rounds of iteration, a set of decision trees is generated, and the objective function is gradually optimized using the loss function and regularization term.

[0048] In step S302, prediction result A and prediction result B are combined through a fully connected fusion layer (eg, gradient perceptron, MLP), and the MLP projects the two prediction results into a new space and learns how to weight the outputs of the two.

[0049] During fusion, in order to ensure that the final result contains all the information of the two models, the present invention uses a concatenation function to directly concatenate the output vectors of the two models together to form a new feature vector F. (8), where F is the concatenated feature vector, concat is the concatenation function, A is the prediction result A of the WOA-GRUA model, and B is the prediction result B of XGBoost.

[0050] The fused feature vector F is input into an MLP, and the processing steps of MLP are as follows: First layer: Receive the fused feature vector F. (9), The second layer: multiple fully connected layers, each layer uses the nonlinear activation function ReLU. (10), The third layer: outputs the final prediction result y. (11), here, , and is the weight matrix of the fully connected layer, , , is the bias term.

[0051] In step S303: the output result of the MLP is classified into two categories through a second activation function (such as a Softmax activation function) to obtain a final prediction result.

[0052] After the MLP output, the Softmax activation function is used to normalize the MLP output. The output of the Softmax activation function is converted into a probability distribution P: (12), in, yes categories (i=1, 2, 3 in the present invention, representing normal tool, light wear and heavy wear respectively), The corresponding category The output result is, is the sum of all class outputs.

[0053] During the entire training phase, the entire multi-source gating prediction model WGX will be jointly optimized. The WOA-GRUA part will use the whale optimization algorithm WOA to optimize the hyperparameters of the GRU-A model, further improving the multi-source gating prediction model WGX's ability to fit time series data. The XGBoost part uses the gradient boosting method to process the data and captures complex nonlinear relationships by integrating decision trees. During training, the weights in the MLP will be adjusted to optimize the learning ability of the fused features, so as to find the best nonlinear mapping on the output of WOA-GRUA and XGBoost.

[0054] In step S40: Using accuracy and ROC curve as evaluation criteria, the tool wear status is analyzed based on the prediction results, and the wear trend and remaining life prediction are provided. Combined with practical applications, early warnings or maintenance suggestions can be provided to operators to optimize tool replacement or maintenance strategies.

[0055] It should also be noted that the data of the present invention will also use the wear data of different lathes and milling cutters, and then combine them to obtain a new data set, thereby improving the generalization ability of the training model.

[0056] The following is a device embodiment corresponding to the above method embodiment. This embodiment can be implemented in conjunction with the above embodiment. The relevant technical details mentioned in the above embodiment are still valid in this embodiment. In order to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied in the above embodiment.

[0057] In another embodiment, if Figure 8 As shown, the present invention also provides a tool wear state prediction device 2 based on multi-source feature fusion, and the tool wear state prediction device 2 based on multi-source feature fusion includes: A multi-source feature data set construction module 20 is used to obtain a variety of tool wear state data to construct a multi-source feature data set and pre-process the multi-source feature data set; A fused feature data set forming module 21 is used to perform deep fusion feature learning on the preprocessed multi-source feature data set using a multimodal fusion autoencoder AE to generate a fused feature data set; The prediction model building module 22 is used to extract features from the fused feature data set using a neural network, generate a feature sequence, and input the feature sequence into the multi-source gating prediction model WGX for training; The prediction module 23 is used to predict the wear state of the tool to be detected by using the trained multi-source gating prediction model WGX.

[0058] In this embodiment, if Fig. 9 As shown, the fusion feature data set forming module 21 also includes: an encoding unit 210, which is used for the encoding process from the input layer to the hidden layer, and the calculation formula is formula (1); a decoding unit 211, which is used for the decoding process from the hidden layer to the output layer, and the calculation formula is formula (2).

[0059] In this embodiment, if Fig.10 As shown, the prediction model building module 22 also includes: A first prediction result generating module 220 is used to model and predict the feature sequence using the WOA-GRUA model to obtain a first prediction result, wherein the WOA-GRUA model is a model that combines the attention mechanism enhanced gated recurrent unit GRU and the whale optimization algorithm WOA; The second prediction result generating module 221 is used to use XGBoost to model and predict the feature sequence to obtain a second prediction result; A third prediction result generating module 222, configured to fuse the first prediction result and the second prediction result through a connection fusion layer MLP to generate a third prediction result; The final prediction result generating module 223 is used to perform binary classification on the third prediction result through the Softmax activation function to generate the final prediction result.

[0060] In this embodiment, the WOA-GRUA model uses two gating devices of the gated recurrent unit GRU enhanced by the attention mechanism to obtain the first prediction result, such as Fig.11 As shown, the first prediction result generating module 220 further includes: The attention score calculation unit 2200 is used to calculate the value score of the output of the hidden state of the gated recurrent unit GRU at each time step through the fully connected layer, and use the first activation function to ensure that the score is between -1 and 1; An attention weight coefficient calculation unit 2201 is used to convert the score into a probability distribution using a second activation function to ensure that the sum of all weight coefficients is 1, and each weight coefficient represents the relative importance of the hidden state output of the gated recurrent unit GRU at the corresponding time step in the sequence; The output unit 2202 of the attention layer is calculated, and is used to realize the output of the attention layer by using the weight coefficient and the corresponding gated recurrent unit GRU hidden state output.

[0061] In this embodiment, if Fig.12 As shown, the third prediction result generating module 222 also includes: a splicing unit 2220, which is used to splice the first prediction result and the second prediction result using a splicing function to generate a fused feature vector; and a generating unit 2221, which is used to input the fused feature vector into the connection fusion layer MLP to generate a third prediction result.

[0062] In addition, technicians in the relevant field can clearly understand that for the convenience and simplicity of description, the specific working process of the tool wear state prediction device 2 based on multi-source feature fusion can refer to the corresponding process in the aforementioned embodiment of the tool wear state prediction method 1 based on multi-source feature fusion, and will not be repeated here.

[0063] In summary, the present invention uses an autoencoder (AE) model to fuse the multi-source features of the tool wear state, effectively extract key information from the data, and generate a set of more representative fusion features, so as to construct a prediction model that can fully reflect the tool wear state, and improve the prediction accuracy and generalization ability. The present invention uses the gated recurrent unit GRU optimized by the whale optimization algorithm WOA for feature extraction and prediction, which can better extract key information from the features after multi-source fusion. In order to ensure the prediction ability of the model for samples outside the training set, the deep-sea layer maintains a high accuracy on different depth test sets, the L1 regularization term retains all variables, but reduces the order of magnitude of the feature variables, and the regularization parameter balances the goal of fitting training and keeps the parameter value of the feature contribution small. The Adam optimizer is used to realize simple and efficient calculation of the model, and the weight and bias parameters are tuned through model training, and the prediction result sequence is obtained through the test set. The GRU model is enhanced by the attention mechanism, which can better give attention weight parameters to each historical moment.

[0064] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which belong to the protection scope of the present invention.

Claims

1. A tool wear state prediction method based on multi-source feature fusion, characterized in that: The following steps are involved: Acquire multiple tool wear state data to construct a multi-source feature data set, and pre-process the multi-source feature data set; Using a multimodal fusion autoencoder AE to perform deep fusion feature learning on the preprocessed multi-source feature dataset to generate a fused feature dataset; Using a neural network to extract features from the fused feature data set, generating a feature sequence, and inputting the feature sequence into a multi-source gating prediction model WGX for training; The trained multi-source gating prediction model WGX is used to predict the wear state of the tool to be detected.

2. The tool wear state prediction method based on multi-source feature fusion according to claim 1 is characterized in that: The multi-source gating prediction model WGX is a fusion model of the whale optimization algorithm WOA, the gated recurrent unit GRU, the connection fusion layer MLP and XGBoost.

3. The tool wear state prediction method based on multi-source feature fusion according to claim 1 or 2, characterized in that: The plurality of tool wear state data include vibration signal data, sound signal data and temperature signal data.

4. The tool wear state prediction method based on multi-source feature fusion according to claim 3 is characterized in that: The method of using the multimodal fusion autoencoder AE to perform deep fusion feature learning on the preprocessed multi-source feature dataset to generate a fusion feature dataset also includes the following steps: The encoding process from the input layer to the hidden layer is calculated as follows (1): (1); The decoding process from the hidden layer to the output layer is calculated as follows (2): (2); in, represents the output of the encoding process, represents the output of the decoding process, represents the Sigma activation function;Sigma activation function; , , , represents the weight matrix of the fully connected layer, , represents the bias term, X represents the vibration signal data, Y represents the sound signal data, and Z represents the temperature signal data.

5. The tool wear state prediction method based on multi-source feature fusion according to claim 2 is characterized in that: The step of inputting the feature sequence into the multi-source gating prediction model WGX for training also includes the following steps: Modeling and predicting the feature sequence using the WOA-GRUA model to obtain a first prediction result, wherein the WOA-GRUA model is a model that combines the attention mechanism-enhanced gated recurrent unit GRU and the whale optimization algorithm WOA; Using the XGBoost to model and predict the feature sequence to obtain a second prediction result; The first prediction result and the second prediction result are fused through the connection fusion layer MLP to generate a third prediction result; The third prediction result is classified into two categories through a second activation function to generate a final prediction result.

6. The tool wear state prediction method based on multi-source feature fusion according to claim 5 is characterized in that: The method of using the WOA-GRUA model to model and predict the feature sequence to obtain a first prediction result also includes the following steps: Calculate the attention score, for each time step of the output of the gated recurrent unit GRU hidden state, calculate the value score through the fully connected layer, and use the first activation function to ensure that the score is between -1 and 1; Calculate the attention weight coefficient, and use the second activation function to convert the score into a probability distribution, ensuring that the sum of all weight coefficients is 1; The output of the attention layer is calculated, and the output of the attention layer is realized by using the weight coefficient and the output of the corresponding gated recurrent unit GRU hidden state.

7. The tool wear state prediction method based on multi-source feature fusion according to claim 5 is characterized in that: The step of fusing the first prediction result and the second prediction result through the connection fusion layer MLP to generate a third prediction result further includes the following steps: Using a concatenation function to concatenate the first prediction result and the second prediction result to generate a fused feature vector; The fused feature vector is input into the connection fusion layer MLP to generate the third prediction result.

8. The tool wear state prediction method based on multi-source feature fusion according to claim 7 is characterized in that: The calculation formula of the fused feature vector is as follows (8): (8), Among them, F is the fused feature vector, A is the first prediction result, B is the second prediction result, and concat is the concatenation function.

9. The tool wear state prediction method based on multi-source feature fusion according to claim 8 is characterized in that: The connection fusion layer MLP includes: The first layer receives the fused feature vector F, and the calculation formula is as follows (9): (9), in, is the output of the first layer, is the weight matrix of the first layer, is the bias term of the first layer, is the activation function; The second layer, the calculation formula is as follows (10): (10), in, is the output of the second layer, is the weight matrix of the second layer, is the bias term of the second layer; The third layer outputs the third prediction result, and the calculation formula is as follows (11): (11), Wherein, y is the third prediction result, W out is the weight matrix of the third layer, b out is the bias term of the third layer, represents the second activation function.

10. A tool wear state prediction device based on multi-source feature fusion, characterized in that: The tool wear state prediction method based on multi-source feature fusion according to any one of claims 1 to 9 at least comprises: A multi-source feature data set construction module, used for acquiring a variety of tool wear state data to construct a multi-source feature data set, and preprocessing the multi-source feature data set; A fused feature data set forming module is used to perform deep fusion feature learning on the preprocessed multi-source feature data set using a multimodal fusion autoencoder AE to generate a fused feature data set; A prediction model building module, used to extract features from the fused feature data set using a neural network, generate a feature sequence, and input the feature sequence into a multi-source gating prediction model WGX for training; The prediction module is used to predict the wear state of the tool to be detected by using the trained multi-source gating prediction model WGX.

Images