Tool residual life prediction method based on weighted trend coefficient and self-encoding

Abstract

The tool residual life prediction method based on the weighted trend coefficient and the self-encoding includes the following steps: step one, collecting tool full life process data; step two, screening the features extracted in step one; step three, encoding the screened feature matrix obtained in step two to obtain a health index sequence; step four, constructing a Bi-LSTM model; and step five, collecting new tool monitoring signals to output the residual useful life corresponding to the current working step. The application effectively reduces feature information redundancy, uses a self-encoder to reduce and fuse the features to obtain the most essential representation of tool degradation, retains the interpretability while reducing the dimension, uses a data-driven method to predict the tool residual useful life, and fully plays the advantages of the Bi-LSTM network modeling based on the screened and fused features, improves the model precision and robustness, and provides technical support for predictive maintenance of the tool. In conclusion, the application has good application prospect.

Description

Technical Field

[0001] This invention relates to the field of judgment method technology, and in particular to a tool remaining life prediction method based on weighted trend coefficient and self-encoding. Background Technology

[0002] Cutting tools are a crucial component of machine tools, subjected to various forces, heat, and frictional loads during machining. The ongoing wear, repair, and replacement costs account for 25-35% of production costs. Therefore, accurately predicting tool wear and calculating remaining tool life is essential for ensuring component quality, minimizing machining losses, and reducing production costs. Existing methods for predicting remaining tool life mainly fall into four categories: 1) empirical methods; 2) physical model methods; 3) data-driven methods and data-driven methods combined with physical model methods. Empirical and physical model methods, due to the need for human intervention, suffer from relatively inaccurate prediction efficiency.

[0003] In recent years, due to the continuous development of sensor functionality and accuracy, computer systems and computing speed, and related algorithms, data-driven methods have developed rapidly. Data-driven methods for predicting tool life mainly include the following modes: First, after wavelet feature decomposition and correlation analysis of signals during tool machining using relevant software units, a feature selection method is used to obtain the final feature set, and then Logistic Regression (LSTM) is used as the tool wear prediction algorithm model. Second, combining the force and vibration data of the tool in three directions during machining, and then using LSTM to predict the actual tool wear. Third, after wavelet decomposition of the vibration monitoring information during the milling process to obtain sub-signals, these are input into a convolutional neural network to implement a tool wear monitoring scheme. Although the above research methods have achieved certain results in the field of remaining tool life, the redundancy of features in the preprocessing process of the above methods still needs to be reduced, and the interpretability of feature selection needs to be improved. In actual calculations, on the one hand, too many features will bring a large amount of redundant information, burdening subsequent model construction and even potentially causing the curse of dimensionality; on the other hand, directly reducing the dimensionality of features will lead to poor interpretability of the subsequently constructed model and an inability to handle noisy data well. Therefore, after acquiring signals and extracting features during actual machining, how to evaluate the performance of these features and how to retain their interpretability while reducing feature redundancy are urgent problems that need to be solved by the academic and industrial communities using data-driven methods to predict tool life. Summary of the Invention

[0004] To overcome the shortcomings of existing data-driven methods for tool life prediction, as described in the background, due to technological limitations, this invention provides a tool life prediction method based on weighted trend coefficients and autoencoders. This method, through the combined action of related steps, utilizes a feature performance evaluation method—weighted trend coefficients—to remove redundant features from the calculated total feature set. This fully leverages the information contained in the original signals from the workpiece being machined, reducing the time and computational load required for subsequent remaining tool life model construction. Furthermore, based on autoencoders and bidirectional long short-term memory networks, a subset of well-performing feature sets is fused as a health indicator, effectively constructing a tool life prediction model.

[0005] The technical solution adopted by this invention to solve its technical problem is:

[0006] A tool remaining life prediction method based on weighted trend coefficients and autoencoders is characterized by the following steps: Step 1, synchronously collecting tool life data from the start of machining with a new tool until it no longer meets working requirements using vibration and power sensors mounted on the machine tool spindle, wherein vibration signals and power signals {X, Y, Z, P} in three directions are collected; Step 2, using a data-driven approach and weighted trend coefficients, the features extracted in Step 1 are comprehensively screened from both monotonicity and predictability dimensions to select features that can effectively reflect the tool degradation trend, thereby improving the accuracy of the prediction model; Step 3, training an autoencoder to encode the selected feature matrix obtained in Step 2 to obtain a health index sequence I. m Step 4: Construct a Bi-LSTM model, normalize the health index sequence and remaining service life, and input the model for parameter learning; Step 5: Collect new tool monitoring signals, extract the features from the feature matrix selected in Step 2, encode them using the autoencoder trained in Step 3, and input the obtained health index sequence into the Bi-LSTM model obtained in Step 4 for remaining service life prediction. The model output is the remaining service life corresponding to the current step.

[0007] Furthermore, in step one, the collected vibration and power signal data are specifically processed by extracting time-domain and frequency-domain statistical features for each channel signal on a per-step basis. These features include mean, standard deviation, skewness, root mean square value, kurtosis, and centroid frequency, resulting in a feature matrix A. m×n , where m is the number of work steps in chronological order, and n is the number of features extracted for the corresponding work step.

[0008] Further, in step two, the weighted trend coefficient of the features extracted in step one is calculated using the formula fit=ω1*mon+ω2*prog, and the formula mon=mean is used. i|corr(rank(x i ),rank(t i Calculate the monotonic coefficients mon for i = 1, ..., M using the formula When calculating the prediction coefficient prog for i = 1, ..., M, a threshold is set. If the weighted trend coefficient of a feature is lower than the threshold, the feature is not used. The selected features are then reconstructed into the selection feature matrix B.

[0009] Furthermore, in step three, the autoencoder uses the formulas H = f(WX + b) and X′ = f(W′H + b′) for encoding and decoding; during training, the loss function is used to minimize the error between the output and input, and the formula is... Specifically, input the filtering feature matrix B obtained in step two, train the autoencoder to learn parameters, set the number of neurons in the hidden layer to 1, extract the output from the hidden layer, and obtain the health indicator sequence I.

[0010] Furthermore, in step three, the feature dimensionality reduction is completed by using an autoencoder encoding process, compressing the features in the selected feature matrix into 1 dimension to obtain the health index sequence that best characterizes the tool degradation process.

[0011] Furthermore, in step four, the health indicator sequence and remaining service life are normalized and then input into the model for parameter learning. Specifically, the remaining service life adopts a segmented life, that is, in the early stage, as the tool steps increase, the tool decay is slower and the remaining tool life is at its maximum value. In the later stage, the tool decay is faster and the remaining service life decreases linearly with the increase of tool steps.

[0012] Furthermore, in step four, the Bi-LSTM model utilizes the formula... h t =g o ∶tanhM t To update the time state, the Bi-LSTM model consists of three Bi-LSTM model layers, each with 256 nodes. To prevent overfitting, each Bi-LSTM model layer is followed by a dropout layer. After passing through three fully connected layers, the remaining lifetime is output by a regression layer. The first two fully connected layers have 128 and 56 nodes, respectively, and the activation functions are ReLU and tanh, respectively.

[0013] The beneficial effects of this invention are: (1) This invention comprehensively filters the extracted features through two dimensions: monotonicity and predictability, effectively reducing feature information redundancy and leaving indicators with low dimensionality and high sensitivity for the prediction of tool remaining life, thereby improving the efficiency of tool remaining life model construction and the accuracy of tool remaining life prediction; (2) It uses an autoencoder to perform dimensionality reduction and fusion of features to obtain the most essential representation of tool degradation, while retaining its interpretability; (3) It adopts a data-driven method to predict tool remaining life. Based on the screening and fusion of features, it uses a Bi-LSTM network for modeling, giving full play to its advantages, improving model accuracy and robustness, and providing a new and effective tool for predictive tool maintenance. In summary, this invention has good application prospects. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Figure 1 This is a schematic flowchart of the present invention.

[0016] Figure 2 This is the original signal diagram of one step in the present invention.

[0017] Figure 3 This is the weighted trend coefficient chart of the present invention.

[0018] Figure 4 This invention constructs a health indicator chart.

[0019] Figure 5 This is a graph showing the prediction results of this invention. Detailed Implementation

[0020] Figure 1 As shown, a method for predicting the remaining tool life based on weighted trend coefficients and autoencoders is presented. The specific steps are as follows: Step 1: Simultaneously collect data on the entire tool life from the start of machining with a new tool until it no longer meets working requirements using vibration and power sensors mounted on the machine tool spindle. Collect vibration signals and power signals {X, Y, Z, P} in the axial, radial, and tangential directions. Specifically, extract time-domain and frequency-domain statistical features from the signal of each channel for each complete machining process (i.e., each machining step), such as mean, standard deviation, skewness, root mean square value, kurtosis, and centroid frequency, to obtain the feature matrix A. m×n Where m is the number of work steps in chronological order, and n is the number of features extracted for the corresponding work step. The vibration signal and power signal collected in one work step are as follows: Figure 2 As shown. Step one involves collecting raw signals and calculating the feature matrix to provide data support for subsequent feature selection, health indicator reconstruction, and training of the tool life prediction model.

[0021] Figure 1 As shown, in step two, based on data-driven methods, the features extracted in step one are comprehensively screened from two dimensions: monotonicity and predictability using a weighted trend coefficient. Specifically, the weighted trend coefficient of the features extracted in step one is calculated using formula (1):

[0022] fit=ω1*mon+ω2*prog (1)

[0023] In the formula, fit is the weighted trend coefficient, mon is the monotonic coefficient, prog is the prediction coefficient, and ω1 and ω2 are the weights.

[0024] Calculate the monotonic coefficient mon using formula (2):

[0025] mon = mean i |corr(rank(x i ),rank(t i ))|、i=1,...,M (2)

[0026] In the formula, x is the feature vector, t is the process sequence, M is the number of different tools, rank is the ranking level of the calculated sequence, corr is the correlation between the two sequences, and the monotonic coefficient reflects the correlation between the feature and the tool degradation process.

[0027] The prediction coefficient prog is calculated using formula (3):

[0028]

[0029] In the formula, x is the feature vector, x(end) is the last point of the feature vector, x(start) is the first point of the feature vector, M is the number of different tools, std is the calculated standard deviation, and mean is the calculated mean. After calculating the weighted trend coefficient of each feature, a threshold is set. If the weighted trend coefficient of a feature is lower than the threshold, that feature is not used. The selected features are reconstructed into a selection feature matrix B. The weighted trend coefficient of the features is extracted as follows: Figure 3 As shown, the threshold is set to 0.9. Step two filters out features that can effectively reflect the trend of tool degradation, which helps to improve the accuracy of the prediction model.

[0030] Figure 1 As shown, in step three, the autoencoder is trained to encode the selected feature matrix obtained in step two. An autoencoder is an unsupervised machine learning model consisting of an input layer, a hidden layer, and an output layer, divided into an encoder and a decoder. The autoencoder encodes the input, performs a nonlinear transformation to map the input to a low-dimensional space, and then reconstructs it using the decoder to minimize the error between the output and the input. The autoencoder uses formulas (4) and (5) for the encoding and decoding process:

[0031] H = f(WX + b) (4)

[0032] X′=f(W′H+b′) (5)

[0033] In the formula, f is the nonlinear activation function, W and W′ are the hidden weights, b and b′ are the biases, X′ is the output, and X is the input. The formula for minimizing the error between the output and input using the loss function during training is as follows.

[0034]

[0035] Then, input the selected feature matrix B obtained in step two, train the autoencoder for parameter learning, set the number of neurons in the hidden layer to 1, extract the output from the hidden layer, and obtain the health indicator sequence I. Specifically, the feature fusion process using the autoencoder is as follows: Figure 4 As shown. Step 3 utilizes the autoencoder encoding process to complete feature dimensionality reduction, compressing the features in the selected feature matrix into 1 dimension, thus obtaining the health index sequence that best characterizes the tool degradation process.

[0036] Figure 1 As shown, in step four, a Bi-LSTM model is constructed. The health indicator sequence and remaining lifespan are normalized and then input into the model for parameter learning. The Bi-LSTM unit updates the time-state using formulas (7), (8), and (9):

[0037]

[0038]

[0039] h t =g o ∶tanhM t (9)

[0040] in, For temporary state information, M t In the unit cell state, h t The output of a unit at time t is represented by x. The output of a Bi-LSTM unit consists of the outputs of two LSTM units in opposite directions, allowing the use of past and subsequent information. t Let W be the input vector at time t. c and b c For weights and biases, g o For output gate, g i For the input gate, g fThe forget gate is used, and the three gates work together to update the unit state. tanh is the activation function. The Bi-LSTM model consists of three Bi-LSTM layers, each with 256 nodes. To prevent overfitting, a dropout layer is connected after each Bi-LSTM layer. After three fully connected layers, a regression layer outputs the remaining lifetime. The first two fully connected layers have 128 and 56 nodes respectively, with ReLU and tanh activation functions. The remaining lifetime is calculated using a piecewise lifetime approach: initially, tool decay is slow with increasing tool steps, resulting in the maximum remaining lifetime; later, tool decay is rapid, and the remaining lifetime decreases linearly with increasing tool steps.

[0041] Figure 1 As shown, in step five, new tool monitoring signals are acquired, features are extracted from the feature matrix selected in step two, and encoded using the autoencoder trained in step three. The resulting health indicator sequence is then input into the Bi-LSTM model obtained in step four for remaining service life prediction. The model output is the remaining service life corresponding to the current step. The model prediction results are as follows: Figure 5 As shown, the mean error is 0.0008, which can predict the remaining tool life with high accuracy.

[0042] Through the above, this invention comprehensively filters the extracted features by considering both monotonicity and predictability, effectively reducing feature redundancy and leaving indicators with low dimensionality and high sensitivity for tool life prediction. This improves the efficiency of tool life model construction and the accuracy of tool life prediction. An autoencoder is used to reduce and fuse features to obtain the most essential representation of tool degradation, preserving interpretability while reducing dimensionality. A data-driven method is employed to predict tool life. Based on feature filtering and fusion, a Bi-LSTM network is used for modeling, fully leveraging its advantages to improve model accuracy and robustness, providing effective technical support for predictive tool maintenance.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0044] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting remaining tool life based on weighted trend coefficients and auto-encoding, characterized in that, Includes the following steps: Step one involves synchronously collecting data on the entire tool lifecycle, from the start of machining with a new tool until it no longer meets working requirements, using vibration and power sensors mounted on the machine tool spindle. This includes collecting vibration signals in three directions and power signals. Step two: Based on data-driven analysis, using weighted trend coefficients, the features extracted in step one are comprehensively screened from both monotonicity and predictability dimensions to identify features that effectively reflect tool degradation trends, thereby improving the accuracy of the prediction model. Step three: An autoencoder is trained to encode the selected feature matrix B obtained in step two, resulting in a health indicator sequence. And obtain an effective description of the tool degradation process; Step 4, construct a Bi-LSTM model to convert the health indicator sequence After normalizing with the remaining service life, the data is input into the model for parameter learning; Step 5: Collect new tool monitoring signals, extract features from the feature matrix B selected in Step 2, encode them using the autoencoder trained in Step 3, and then generate the resulting health indicator sequence. Input the Bi-LSTM model obtained in step four to predict the remaining useful life; the model output is the remaining useful life corresponding to the current work step. In step two, the formula is used... Calculate the weighted trend coefficient of the features extracted in step one using the formula. Calculate the monotonic coefficient Using the formula Calculate the prediction coefficient In calculating the weighted trend coefficient of each feature, a threshold is set. If the weighted trend coefficient of a feature is lower than the threshold, that feature is not used. The selected features are then reconstructed into a selection feature matrix. In step three, the autoencoder uses the formula Perform the encoding and decoding process; Using loss function during training Minimize the error between the output and the input, the formula is: Specifically, input the filtering feature matrix obtained in step two. The autoencoder is trained to learn parameters. The number of neurons in the hidden layer is set to 1. The output of the hidden layer is extracted to obtain the health indicator sequence. In step three, the feature dimensionality reduction is completed using the autoencoder encoding process, and the selected feature matrix is ​​then processed. By compressing the features in the data into one dimension, a health index sequence that best characterizes the tool degradation process is obtained. .

2. The tool remaining life prediction method based on weighted trend coefficient and auto-encoding according to claim 1, characterized in that, In step one, the collected vibration and power signal data are processed on a per-step basis. Time-domain and frequency-domain statistical features are extracted from each channel's signal, including mean, standard deviation, skewness, root mean square value, kurtosis, and centroid frequency, resulting in a feature matrix. , where m is the number of work steps in chronological order, and n is the number of features extracted for the corresponding work step.

3. The tool remaining life prediction method based on weighted trend coefficient and auto-encoding according to claim 1, characterized in that, In step four, the health indicator sequence is... After being normalized with the remaining service life, the remaining service life is input into the model for parameter learning. Specifically, the remaining service life adopts a segmented life, that is, in the early stage, as the tool steps increase, the tool decay is slower and the remaining tool life is at its maximum value. In the later stage, the tool decay is faster and the remaining service life decreases linearly with the increase of the tool steps.

4. The tool remaining life prediction method based on weighted trend coefficient and auto-encoding according to claim 1, characterized in that, In step four, the Bi-LSTM model uses the formula To update the time state, the Bi-LSTM model consists of three Bi-LSTM model layers, each with 256 nodes. To prevent overfitting, each Bi-LSTM model layer is connected to a dropout layer. After passing through three fully connected layers, the remaining lifetime is output by a regression layer. The first two fully connected layers have 128 and 56 nodes, respectively, and the activation functions are ReLU and tanh, respectively.

Images