Mobile Net-BiLSTM-Attention tool residual life prediction method based on improved sparrow algorithm

Abstract

The invention discloses a MobileNet-BiLSTM-Attention tool remaining life prediction method based on an improved sparrow algorithm, and aims to solve the problems of accuracy and uncertainty of tool remaining life prediction in an actual machining environment. The method comprises the following steps that a milling experiment platform is built, and multi-channel signals of a cutter under different feeding times are obtained; intercepting a vibration signal stable interval and carrying out noise reduction processing by adopting a moving average filtering method; constructing a MobileNet-BiLSTM-Attention model, inputting the preprocessed signal into the model, and determining key parameters in the model by adopting a sparrow search algorithm (SCSSA) fused with sine and cosine and Cauchy variation so as to give full play to the advantages of the model to realize accurate and efficient prediction of the residual life of the cutter; a confidence interval of a prediction value is calculated by using an adaptive bandwidth kernel density estimation (ABKDE) method, the uncertainty degree of a prediction result is reflected, and a more stable and reliable prediction result is provided.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical processing state monitoring, and specifically relates to a MobileNet-BiLSTM-Attention tool remaining life prediction method based on an improved sparrow algorithm. Background Art

[0002] As the "teeth" of the machine tool, the tool is inevitably affected by thermal-mechanical coupling and wears during the cutting process. The wear condition is a key factor affecting the tool's service life.

[0003] Statistics show that in industrial production, only 38% of the tool life is fully utilized. The irrational use of tools leads to an additional $10 billion in manufacturing costs each year. Effective prediction of the remaining tool life can reduce production costs by 10% to 30% and increase production efficiency by 10% to 40%.

[0004] Therefore, achieving accurate and reliable prediction of remaining tool life is crucial to formulating reasonable tool change strategies, reducing production costs, and ensuring the stability of the machining process.

[0005] In recent years, deep learning has been widely used in the field of tool wear status monitoring due to its powerful data processing capabilities. Although the tool wear monitoring method based on deep learning has achieved great success, it still faces two core challenges.

[0006] On the one hand, how to further optimize the deep learning model to achieve more accurate prediction of tool remaining life.

[0007] On the other hand, there are many factors that affect the remaining life of the tool during the cutting process, and the simple deep learning model prediction ignores the potential uncertainty in the prediction results, which to a certain extent limits the formulation of economical and efficient tool operation and maintenance strategies, and also affects the adaptability of the model in the actual complex and changeable processing environment. Summary of the invention

[0008] The purpose of the present invention is to propose a method for predicting the remaining life of a tool in view of the deficiencies in the prior art, mainly to solve the problems of low prediction accuracy of the model and inability to effectively quantify the uncertainty of the prediction results.

[0009] To achieve the above purpose, the technical solution of the present invention is as follows: The present invention proposes a MobileNet-BiLSTM-Attention tool remaining life prediction method based on an improved sparrow algorithm, which adopts the following steps:

[0010] S1: Obtain the processing signal of the entire life cycle of the tool during the milling process.

[0011] S2: intercept the stable interval of the original signal and perform noise reduction and normalization processing on it.

[0012] S3: The preprocessed data is labeled according to the remaining number of cuts to form a data set, and is divided into a training set and a test set.

[0013] S4: Construct a MobileNet-BiLSTM-Attention tool remaining life prediction model optimized by the improved Sparrow algorithm;

[0014] S4.1: The model mainly consists of MobileNetV2 module, deep BiLSTM module and attention module;

[0015] S4.2: MobileNetV2 is a lightweight convolutional neural network. Compared with traditional convolutional networks, it can significantly reduce model parameters and computation while maintaining accuracy. The network mainly introduces an inverted residual module, which first increases the dimension through 1×1 convolution to capture more features, and then reduces the number of channels back to the number of output channels through 1×1 convolution, reducing the number of model parameters and complexity.

[0016] S4.3: Each inverted residual module in MobileNetV2 has a linear bottleneck layer. Unlike the traditional ReLU activation function, the linear bottleneck layer uses a linear activation function to retain more feature information;

[0017] S4.4: Deep BiLSTM module. BiLSTM can capture the dependencies in the sequence from both the front and back directions, overcoming the problems of gradient vanishing and gradient exploding in traditional RNN when processing long sequences. By stacking multiple BiLSTM layers, the model’s expressiveness and ability to capture complex features are improved. Since the number of neurons is the key factor affecting the performance of the BiLSTM layer, the optimal number of neurons is found by comprehensively considering the model complexity and performance through the improved sparrow algorithm.

[0018] S4.5: Attention module, Squeeze-and-Excitation (SE) attention mechanism can adaptively adjust the importance of different channels to improve network performance. It compresses features in the spatial dimension through global average pooling operation so that the features have global information. Then, two fully connected layers adaptively adjust the importance of different channels and assign different weights to feature maps so that the neural network can focus on key information among a lot of input information.

[0019] S5: The setting of hyperparameters plays a vital role in the model training process, directly affecting the model's performance, training efficiency, and stability.

[0020] S6: Therefore, this method introduces the sparrow search algorithm (SCSSA) that combines sine and cosine and Cauchy mutations to optimize and determine hyperparameters. The SCSSA algorithm is a new improved algorithm based on the sparrow search algorithm (SSA), which significantly improves the convergence speed and optimization ability. The specific steps are as follows:

[0021] S6.1: Initialize the sparrow population and population parameters, and calculate the current fitness value of each sparrow;

[0022] S6.2: Find the sparrow individuals corresponding to the best and worst fitness values ​​respectively, and record their locations;

[0023] S6.3: Introduce the sine and cosine search strategy in the seeker position of the sparrow search algorithm and update the seeker position to enhance the global search capability;

[0024] S6.4: Introduce Cauchy mutation into the follower position update formula to update the follower position to increase the search diversity and the ability to escape from the local optimum;

[0025] S6.5: Update the location of sentinels in the sparrow population to ensure population diversity and continuity of search;

[0026] S6.6: Determine whether the preset number of iterations is reached or other termination conditions are met. If so, output the optimal fitness value and the optimal position. Otherwise, continue iterating.

[0027] S7: Use optimization algorithms to determine the optimal initial learning rate to speed up the convergence of the model and avoid falling into the local optimal solution, as well as the optimal L2 regularization coefficient to prevent model overfitting and improve the generalization ability of the model.

[0028] S8: The Adam optimizer is used to update the network parameters to minimize the loss function so that the model can converge faster and achieve better performance during the training process, thus realizing the prediction of the remaining life of the tool.

[0029] S9: In order to verify the effectiveness of the model proposed in this invention, three performance indicators are used to evaluate the prediction effect of the model, including mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R 2 ); The calculation formulas for these three performance indicators are as follows:

[0030]

[0031]

[0032]

[0033] Where: N is the number of test data, X(t) is the actual value, is the predicted value; the smaller the value of MAE and RMSE, the higher the prediction accuracy of the model. 2 The closer it is to 1, the better the model fit.

[0034] S10: The prediction results output by the model are used as the input of the adaptive bandwidth kernel density estimation (ABKDE) to estimate the probability density function of the predicted value, and then calculate the prediction interval according to the confidence level.

[0035] S11: In order to verify the reliability of the interval prediction result of the present invention, the performance evaluation indicators interval coverage ratio (PICP), normalized average width (PINAW), and coverage width criterion (CWC) are used for evaluation, and the calculation formula is as follows:

[0036]

[0037]

[0038]

[0039] Where: Counti is used to determine whether the i-th true value is within the prediction interval.

[0040] The beneficial effects of the present invention are as follows:

[0041] The present invention builds a milling cutter wear monitoring test data acquisition platform to predict the remaining life of the tool, collects raw signal data and performs denoising and normalization processing on the vibration signal to improve the accuracy of the prediction result.

[0042] The present invention constructs a MobileNet-BiLSTM-Attention model to adaptively extract features without relying on subjective experience, so that the model has higher flexibility and accuracy when processing signals of different types and complexities.

[0043] The MobileNetV2 module in the model of the present invention can achieve more efficient calculation while maintaining high model performance. The deep BiLSTM module enables the model to fully understand the temporal dependency of the data and improve the prediction accuracy of the model. The attention SE module adjusts the importance of information in the channel dimension to further enhance the feature expression ability of the model.

[0044] The present invention uses an optimization algorithm to intelligently iteratively optimize three important parameters of the BiLSTM layer neurons, learning rate and L2 regularization coefficient of the deep BiLSTM module in the model, and obtains the optimal parameters to give full play to the advantages of the model and improve the accuracy of model prediction.

[0045] The present invention adopts adaptive bandwidth kernel density estimation to estimate the uncertainty of the predicted value. Since the traditional kernel density estimation method has the limitation of fixed bandwidth parameters, which affects the degree of smoothness and the accuracy of the estimation, the adaptive bandwidth kernel density estimation can adaptively adjust the bandwidth parameters according to the local density of the data, which can better adapt to the local characteristics and density changes of the data to improve the estimation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the MobileNet-BiLSTM-Attention tool remaining life prediction method based on the improved sparrow algorithm of the present invention;

[0047] Figure 2 This is a schematic diagram of the principle of the sparrow search algorithm that combines sine and cosine and Cauchy mutations;

[0048] Figure 3 Schematic diagram of the MobileNet-BiLSTM-Attention model;

[0049] Figure 4 This is the structure diagram of the MobileNetV2 module;

[0050] Figure 5 This is the BiLSTM module structure diagram;

[0051] Figure 6 This is the structure diagram of the SE attention module;

[0052] Figure 7 This is the tool remaining life prediction diagram based on the CNN-LSTM model;

[0053] Figure 8 This is the remaining tool life interval prediction diagram of MobileNet-BiLSTM-Attention based on the improved sparrow algorithm. DETAILED DESCRIPTION

[0054] In order to more clearly describe the technical content and effects of the present invention in detail, further description and illustration are given below in conjunction with the drawings and examples. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0055] like Figure 1 In the present invention, a MobileNet-BiLSTM-Attention tool remaining life prediction method based on an improved sparrow algorithm is provided, and the specific steps are as follows:

[0056] S1: By installing piezoelectric dynamometers, vibration sensors and acoustic emission sensors, the X, Y and Z cutting force signals, vibration signals and acoustic emission signals are collected during the entire life cycle of tool machining workpieces to provide original data sets for model training and testing.

[0057] S2: Preprocess the vibration signal in the original data, including signal noise reduction and normalization processing in the data, as follows:

[0058] S2.1: Considering that the tool cutting in and out will generate some invalid signal data, the sliding average filtering method is used to reduce noise in the signal stability range, and the number of windows is set to 1000 to obtain smoother signal data with less noise interference;

[0059] S2.2: In order to reduce the impact caused by differences in data scale or dimension, so as to better train the model and find the optimal solution faster, the denoised data is normalized. The calculation formula is as follows:

[0060] y=(x-xmin) / (xmax-xmin)

[0061] In the formula: xmax is the maximum value in the data, xmin is the minimum value in the feature data, and x is the input data.

[0062] S3: Build a MobileNet-BiLSTM-Attention model optimized by the improved sparrow algorithm, such as Figure 3 As shown;

[0063] S3.1: This model mainly consists of three modules: MobileNetV2 module, deep BiLSTM module and attention module;

[0064] S3.2: Reference Figure 4 The MobileNetV2 module introduces a reverse residual structure with a linear bottleneck. The structure first uses a 1×1 convolution kernel to increase the dimension, then uses deep convolution to process the increased-dimensional features, and finally restores the dimension of the input features through a 1×1 convolution kernel. The linear bottleneck layer then maps the features to a low-dimensional compressed expression. At the same time, in order to avoid the loss or destruction of information caused by the use of the ReLU activation function, the last convolution layer of the bottleneck structure no longer uses the ReLU activation function but a linear activation function.

[0065] S3.3: Reference Figure 5 , construct a deep BiLSTM module to extract features of different time scales. The deep BiLSTM module includes two layers of BiLSTM. The optimization algorithm is used to determine that the number of neurons in the first layer is 63 and the number of neurons in the second layer is 32. The BiLSTM structure is as follows Figure 5As shown; its implementation is as follows:

[0066]

[0067]

[0068] ht′=atht+bthi+ct

[0069] Where: is the forward hidden layer state, is the backward hidden layer state, at and bt are weights, and ct is the bias;

[0070] S3.4: Reference Figure 6 ,The SE module is a lightweight channel attention mechanism, which mainly consists of three parts: Squeeze, Excitation and Scale;

[0071] The compression operation mainly compresses the input feature map to make it have global spatial information. The calculation formula is as follows:

[0072]

[0073] Where: uc is the input feature, W and H represent the width and height of the feature map respectively;

[0074] The excitation operation weights the features of different channels by adjusting their importance. The calculation formula is as follows:

[0075] s=Fex(z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0076] Where: W1 and W2 represent the weights of the two fully connected layers for dimensionality reduction and dimensionality increase, respectively, and σ is the activation function;

[0077] The Scale operation mainly multiplies the weight value obtained from the Excitation part by the corresponding channel of the original feature map, so that the network can pay more attention to important channel features.

[0078] S4: Construct a fully connected layer after the MobileNetV2 module, the deep BiLSTM module and the attention module, and construct a regression layer based on the output results of the fully connected layer.

[0079] S5: The sparrow search algorithm that integrates sine and cosine and Cauchy mutations is improved on the basis of the sparrow search algorithm, effectively solving the problems of insufficient search ability and difficulty in escaping local extreme values ​​in the later stage of the search. It mainly introduces a refraction reverse learning strategy to initialize the sparrow population to enhance species diversity, thereby reducing the stagnation problem of the sparrow search algorithm in the later stage of the search due to insufficient species diversity; secondly, the sine and cosine strategy is introduced in the position update of the discoverer of the sparrow search algorithm, and the sine and cosine algorithm is improved through nonlinear decreasing search factors and weight factors, thereby achieving the purpose of balancing local development and global exploration; finally, the Cauchy mutation mechanism is used to disturb the optimal solution in the follower position update, expand the sparrow search range, and increase the probability of the algorithm escaping local extreme values;

[0080] S5.1: The present invention sets the population size in the algorithm to 20, the number of iterations to 10, and optimizes the L2 regularization parameter, the initial learning rate, and the number of neurons in the BiLSTM layer in the model;

[0081] S5.2: The optimal L2 regularization coefficient is determined to be 0.0070, the optimal initial learning rate is 0.0051, the optimal number of neurons in the first BiLSTM layer is 63, and the optimal number of neurons in the second BiLSTM layer is 32.

[0082] S6: The Adam algorithm is used to automatically update the network model weights according to data iteration, so as to better establish the mapping relationship between signal characteristics and the remaining life of the tool and realize the prediction of the remaining life of the tool.

[0083] S7: In order to verify the effectiveness of the prediction results of the model proposed in this invention, three performance indicators are used to evaluate the model, including mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R 2 ); The calculation formulas of the three performance indicators are as follows:

[0084]

[0085]

[0086]

[0087] Where: N is the number of test data, X(t) is the actual value, is the predicted value.

[0088] S8: Both MAE and RMSE can reflect the difference between the predicted value and the true value. The smaller the index value, the higher the prediction accuracy of the model. 2 Indicates how well the model fits the original data. 2 The closer it is to 1, the better the goodness of fit.

[0089] S9: In order to effectively verify the reliability of the interval prediction results of the present invention, the performance evaluation indicators interval coverage ratio (PICP), normalized average width (PINAW), and coverage width criterion (CWC) are used for evaluation, and the calculation formula is as follows:

[0090]

[0091]

[0092]

[0093] Where: Counti is used to determine whether the i-th true value is within the prediction interval.

[0094] S10: The interval coverage rate in the evaluation index indicates the probability that the actual observation value falls within the prediction interval. The closer the PICP value is to 1, the higher the coverage rate of the prediction interval. The normalized average width indicates the average width of the prediction interval, and the coverage width criterion is a comprehensive indicator for measuring the width and coverage rate of the prediction interval. The three indicators can comprehensively and effectively measure the performance of the model.

[0095] Embodiment: Predict the remaining life of a milling cutter, and verify the effectiveness and feasibility of the present invention by predicting the uncertainty of the remaining life of the cutter.

[0096] 1. Build a data acquisition platform for milling 508Ⅲ steel tool wear test and construct a data set. The data set has 3 groups of experimental data, numbered Cl~C3, and each group of experimental data contains 68 milling records. The experiment uses piezoelectric dynamometer, vibration sensor and acoustic emission sensor to collect signal data under tool milling operation, and uses a small digital microscope industrial camera to observe and measure the tool flank wear morphology and wear amount (VB) offline after each pass. The main equipment and processing parameters used in the experiment are shown in Table 1:

[0097] Table 1 Main experimental equipment and processing parameters

[0098]

[0099] 2. Noise reduction processing: Cut off the signal when the tool cuts in and out, and use the sliding average filtering method in the signal stability range to reduce the impact of random noise on the overall data, so as to achieve simple and efficient noise reduction processing.

[0100] 3. Since each set of experimental data contains 68 milling records, one set of data is selected as the test set, and the other set is selected as the training set. The remaining cutting times of the tool are used as the label. The normalized data is input into the MobileNet-BiLSTM-Attention model optimized by the improved sparrow algorithm, and the CNN-LSTM model is used to compare it. The prediction results of the CNN-LSTM model are as follows: Figure 7 As shown in the figure, the prediction results of the MobileNet-BiLSTM-Attention model optimized by the improved sparrow algorithm are as follows Figure 8 The calculation results of the three indicators are shown in Table 2:

[0101] Table 2 Comparison of tool remaining life prediction results

[0102] Fourth, the sliding average filtering method is used to process the signal. This method is simple to operate and can effectively reduce redundant signals and noise interference in the signal. The MobileNet-BiLSTM-Attention model optimized based on the improved sparrow algorithm has achieved good results in predicting the remaining life of the tool. Compared with the CNN-LSTM model, its prediction error is smaller and has higher accuracy.

[0103] 5. The prediction error distribution of the model proposed in the present invention is fitted by the adaptive bandwidth kernel density estimation method, and the uncertainty range of the tool remaining life prediction result is provided at a confidence level of 95%. The interval prediction result of the model is shown in Figure 8, and the calculation results of the three indicators are shown in Table 3:

[0104] Table 3 Evaluation of model interval prediction results

[0105] 6. The prediction results are evaluated from the three perspectives of reliability, effectiveness and computational efficiency through three evaluation indicators: interval coverage, normalized average width and coverage width criterion. The results show that it can not only achieve accurate prediction of remaining life, but also effectively quantify the uncertainty range of the predicted value.

[0106] The present invention predicts the remaining life of the tool and determines the prediction result range, thereby effectively avoiding the problems of tool waste and damage to the workpiece, and reduced processing quality caused by replacing the tool too early or too late, and provides a certain theoretical basis and technical support for the reasonable use and replacement of tools in actual production.

[0107] Finally, it should be pointed out that the above embodiments are only used as examples to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail according to the preferred embodiments, it should be clear to those skilled in the art that appropriate adjustments or equivalent alternatives can be made without violating the core principles and scope of the technical solutions of the present invention.

Claims

1. A MobileNet-BiLSTM-Attention tool remaining life prediction method based on an improved sparrow algorithm, the method comprising the following steps: S1: Obtain multi-channel raw signal data of the milling process during the entire life cycle of the tool through sensors, and construct raw data set samples; S2: Data preprocessing is performed on the original vibration signal to intercept the signal stability interval in order to reduce the data fluctuation of the tool cutting in and out during the milling process, and the sliding average filter algorithm is used for noise reduction; S3: Normalize the denoised data. The normalization formula is as follows: y=(x-x min ) / (x max -x min ) Where: x max is the maximum value in the data, x min is the minimum value in the data, and x is the input data; S4: The normalized vibration data is input into the MobileNet-BiLSTM-Attention model optimized by the improved sparrow algorithm to predict the remaining tool life; S5: The mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R 2 ) verify the validity of the prediction results of the model disclosed in the present invention; S6: Combine the model prediction results with the adaptive bandwidth kernel density estimate to calculate the prediction interval and quantify the uncertainty of the prediction results; S7: The uncertainty of the model prediction results was verified using the percentage of interval coverage (PICP), the normalized average width (PINAW), and the coverage width criterion (CWC).

2. According to the MobileNet-BiLSTM-Attention tool remaining life prediction method of the improved sparrow algorithm of claim 1, it is characterized by: The specific operation method of S1 is as follows: a three-dimensional coordinate system is established based on the tool to collect the cutting force signal, vibration signal and acoustic emission signal data in the X, Y and Z directions of the real-time CNC machine tool.

3. The MobileNet-BiLSTM-Attention tool remaining life prediction method based on the improved sparrow algorithm according to claim 1 is characterized by: The specific operation method of S2 is as follows: set the sliding window size and initialize the sliding window, average the data in the window to obtain a new output value, then store the new output value in the output sequence, and move the window to repeat the above process until the entire data sequence is traversed.

4. The MobileNet-BiLSTM-Attention tool remaining life prediction method based on the improved sparrow algorithm according to claim 1 is characterized by: The MobileNetV2 module is used to efficiently extract different spatial features, and then the sample time series features are deeply extracted through two BiLSTM layers. The fully connected layer and regression layer are constructed to map the features to the sample label space to predict the remaining life of the tool.

5. The MobileNet-BiLSTM-Attention tool remaining life prediction method based on the improved sparrow algorithm according to claim 4 is characterized by: Each MobileNetV2 module in this model is connected to an attention module. SE attention can adaptively learn the importance of each channel to enhance key channels and compress useless channels.

6. The MobileNet-BiLSTM-Attention tool remaining life prediction method based on the improved sparrow algorithm according to claim 4 is characterized by: The sparrow search optimization algorithm (SCSSA) that combines sine-cosine and Cauchy mutations is used to optimize and determine the optimal configuration of key parameters in the model, such as learning rate, L2 regularization coefficient and the number of neurons in the BiLSTM layer.

7. The MobileNet-BiLSTM-Attention tool remaining life prediction method based on the improved sparrow algorithm according to claim 1 is characterized by: The specific operation method of S6 is as follows: The adaptive bandwidth kernel density estimation method (ABKDE) is a non-parametric kernel density estimation method, which can automatically select the bandwidth of the kernel function and use the adaptive bandwidth kernel density estimation method to estimate the residual distribution of the model and construct the prediction interval.

Images