A tool wear prediction method based on an improved subtraction optimizer combined with an improved bidirectional long short-term memory neural network

Abstract

The application provides a tool wear prediction method based on an improved subtraction optimizer combined with an improved bidirectional long short-term memory neural network. The method comprises the following steps: collecting machine acceleration vibration and cutting force signals, processing the collected signals, decomposing and denoising the down-sampled signals by using ICEEMDAN-improved wavelet mother function, extracting features from the denoised signals, screening the features as inputs of a tool prediction model, constructing a tool wear prediction model, and predicting the tool wear. The tool wear prediction model of the method has high prediction accuracy, and provides guidance for the selection of tool wear prediction methods.

Description

Technical Field

[0001] This invention relates to the field of tool wear assessment and prediction technology, and in particular to a tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network. Background Technology

[0002] The cutting tool is the end effector of a CNC machine tool. During machining, mechanical friction occurs between the cutting tool, the chips, and the workpiece. Simultaneously, affected by cutting forces and cutting heat, the cutting tool will continuously wear down, and this wear will gradually increase over time, eventually leading to chipping. This severely affects machining efficiency, workpiece surface quality, and machining costs.

[0003] Predicting tool wear during CNC machine tool machining is crucial for improving machining efficiency, enhancing workpiece surface quality, and reducing machining costs. However, existing tool wear prediction models lack optimization algorithms and have imperfect network structures, resulting in low accuracy and preventing their widespread adoption in practical applications.

[0004] Therefore, there is an urgent need to develop a tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network. Summary of the Invention

[0005] The purpose of this invention is to provide a tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network, so as to solve the problems existing in the prior art.

[0006] The technical solution adopted to achieve the purpose of this invention is as follows: a tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network, comprising the following steps:

[0007] 1) Simultaneously collect vibration signals and cutting force signals of the worktable and cutting tool during tool operation. 2) Preprocess the signal data collected in step 1).

[0008] 2.1) Truncate the input and output portions of the acquired signal data.

[0009] 2.2) Use the Hampel filtering method to remove outliers from the truncated signal data.

[0010] 2.3) Downsampling is performed on the signal data after removing outliers.

[0011] 2.4) The ICEEMDAN-modified wavelet mother function is used to decompose and denoise the downsampled signal data.

[0012] 3) Extract the time domain, frequency domain, and time-frequency domain feature values ​​of the preprocessed signal data to obtain the original feature dataset.

[0013] 4) Features are selected based on both Spearman correlation coefficient and maximum mutual information correlation coefficient. The selected feature values ​​are used as input to the prediction model to obtain a new feature dataset. After normalizing the new feature dataset with the actual tool wear dataset, the model construction dataset is obtained.

[0014] 5) Constructing the initial tool wear prediction ISABO-IBiLSTM model. The initial tool wear prediction ISABO-IBiLSTM model uses the IBiLSTM network as the base network and is improved using the ISABO algorithm. The IBiLSTM model includes a three-layer bidirectional long short-term memory neural network, an activation layer, a dropout layer, two fully connected layers, and a regression layer. The ISABO algorithm is used to search for the optimal hyperparameters of the IBiLSTM model. The Adam solver is used for optimization.

[0015] 6) The IBiLSTM model is trained based on the model-built dataset and the optimized hyperparameters to obtain the trained tool wear prediction model. During the training process, the root mean square error, mean absolute percentage error, and mean absolute error are calculated based on the predicted wear amount and the actual wear amount to evaluate the prediction accuracy of tool wear.

[0016] 7) Monitor the vibration signals of the worktable and the cutting force signals of the tool under test during the operation of the tool under test, and input the processed real-time signal feature set into the tool wear prediction model to obtain the wear state of the tool under test.

[0017] Further, in step 1), the vibration signal measurement points are distributed at the tool head and the X-table. The vibration signals are collected using a PCB-356A15 accelerometer. Iron sheets are glued to the surfaces of the tool head and the X-table, and the accelerometer is fixed to the iron sheet on the workpiece surface using magnetic force. The cutting force signal is collected using a Kistler force sensor.

[0018] Furthermore, in step 2.3), the sampling ratio is 1 to 10.

[0019] Further, in step 3), 11 time-domain features are extracted, including absolute mean, peak value, root mean square value, root mean square amplitude, skewness value, kurtosis value, waveform factor, impulse factor, skewness factor, kurtosis factor, and margin factor; 4 frequency-domain features, including centroid frequency, mean square frequency, root mean square frequency, and frequency variance; and 24 energy features obtained after three-level wavelet decomposition.

[0020] Furthermore, in step 4), the Spearman correlation coefficient and the maximum mutual information correlation coefficient are used as feature selection indicators, with the following mathematical formulas:

[0021]

[0022] In the formula, The Spearman correlation coefficient. The difference in grade between the eigenvalue and the corresponding actual wear value. This represents the number of samples.

[0023]

[0024] In the formula, For mutual information, For joint probability density, These represent the eigenvalue and the corresponding actual wear value, respectively. The maximum mutual information correlation coefficient is obtained from the mutual information.

[0025] The thresholds for Spearman correlation coefficient and maximum mutual information correlation coefficient were set to 0.95, and 24 feature values ​​were finally selected as inputs to the prediction model.

[0026] Furthermore, in step 4), 80% of the model building dataset is randomly divided into training dataset and 20% into validation dataset.

[0027] Furthermore, the ISABO algorithm sets the population size to 25. The maximum number of iterations is 50. The upper limit of the parameter optimization range is [0.1, 0.1, 300]. The lower limit of the parameter optimization range is [1×10]. -8 1×10 -4

[50] . The maximum number of training iterations is 1000. The retention ratio of the Dropout layer is 0.4.

[0028] The present invention also discloses a computer-readable storage medium storing a computer program for implementing any of the above-described tool wear prediction methods based on an improved bidirectional long short-term memory neural network with an improved subtraction optimizer.

[0029] This invention also discloses a tool wear prediction system implementing any of the above methods, comprising a data acquisition unit, a processor, an input unit, an output unit, and a storage unit. The data acquisition unit monitors the vibration signals of the worktable and the cutting force signals of the tool under test during operation and processes them into a real-time signal feature set. The input unit receives the real-time signal feature set. The output unit displays the dynamic identification results of the tool wear state in real time. The storage unit stores data and a sequence of instructions executed by the processor to implement the tool wear prediction model.

[0030] The technical effects of this invention are beyond doubt:

[0031] A. The improved subtraction optimizer algorithm eliminates the shortcomings of the original subtraction optimizer algorithm, which has fast convergence but is prone to getting trapped in local optima. Theoretically, it can search for the optimal hyperparameters, thereby further improving the prediction accuracy of the model.

[0032] B. By improving the structure of the BiLSTM model, the complexity, robustness, and representation ability of the BiLSTM model are increased, thereby improving the stability and prediction accuracy of the model. Attached Figure Description

[0033] Figure 1 Flowchart of tool wear prediction method;

[0034] Figure 2 To eliminate the comparison before and after the cut-in signal data;

[0035] Figure 3 To eliminate the comparison before and after the cut-out signal data;

[0036] Figure 4 Before and after comparison to remove invalid data (partial) in the middle;

[0037] Figure 5 Before and after comparison of signal (partial) decomposition and noise reduction;

[0038] Figure 6 Spearman's correlation coefficient and maximum mutual information correlation coefficient plots for partial features;

[0039] Figure 7 The image shows the prediction results of the ISABO-IBiLSTM model on the dataset.

[0040] Figure 8 This is a schematic diagram of tool condition monitoring during high-speed milling.

[0041] In the diagram: 1. Tool; 2. Worktable; 3. Workpiece; 4. Triaxial accelerometer; 5. Triaxial force gauge. Detailed Implementation

[0042] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0043] Example 1:

[0044] To address the issue of low prediction accuracy in existing tool wear prediction methods due to the lack of optimization algorithms and imperfect network structures, this embodiment provides a tool wear prediction method based on an improved subtraction optimizer combined with an improved bidirectional long short-term memory neural network, comprising the following steps:

[0045] 1) Vibration signals and cutting force signals of the worktable and cutting tool are simultaneously acquired during tool operation. The workpiece is arranged in a spatial rectangular coordinate system O-XYZ. The origin of the spatial rectangular coordinate system O-XYZ is located at the center of the bottom surface of the workpiece. The X-axis of the spatial rectangular coordinate system O-XYZ is parallel to the radial direction of the tool, and the Y-axis is parallel to the feed direction of the tool. Force sensors and vibration sensors can acquire force signals and acceleration vibration signals in the X, Y, and Z directions, respectively.

[0046] 2) Preprocess the signal data acquired in step 1).

[0047] 2.1) Truncate the input and output portions of the acquired signal data.

[0048] The truncation method uses the upper quartile value of the original data as the boundary point for invalid data. That is, starting from the beginning of the original data and moving backwards and forwards, all data smaller than this boundary value are removed.

[0049] 2.2) Use the Hampel filtering method to remove outliers from the truncated signal data.

[0050] The Hampel filtering method employs a sliding window mechanism and the principle of median filtering, which replaces data points in the signal that exceed three times the standard deviation with the average of 10,000 adjacent points.

[0051] 2.3) The signal data after removing outliers is downsampled at a ratio of 1 to 10.

[0052] 2.4) The ICEEMDAN-modified wavelet mother function is used to decompose and denoise the downsampled signal data.

[0053] The ICEEMDAN algorithm (Improved Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise) is used to decompose the signal, which can solve the problems of residual noise and pseudo-mode components, obtain better IMF components, and thus achieve better noise reduction. The wavelet thresholding function is the sym18 function, where the threshold function is an improved function using a polynomial method.

[0054] 3) Extract the time domain, frequency domain, and time-frequency domain feature values ​​of the preprocessed signal data to obtain the original feature dataset.

[0055] 4) Features are selected based on both Spearman's correlation coefficient and the maximum mutual information correlation coefficient. The selected feature values ​​are used as input to the prediction model to obtain a new feature dataset. After normalizing the new feature dataset with the actual tool wear dataset, the model construction dataset is obtained. The Spearman's correlation coefficient and the maximum mutual information correlation coefficient are used as feature selection metrics, and the mathematical formulas are as follows:

[0056]

[0057] In the formula, The Spearman correlation coefficient. The difference in grade between the eigenvalue and the corresponding actual wear value. This represents the number of samples.

[0058]

[0059] In the formula, For mutual information, For joint probability density, These represent the eigenvalue and the corresponding actual wear value, respectively. The maximum mutual information correlation coefficient is obtained from the mutual information.

[0060] The thresholds for Spearman correlation coefficient and maximum mutual information correlation coefficient were set to 0.95, and 24 feature values ​​were finally selected as inputs to the prediction model.

[0061] 5) Constructing the initial tool wear prediction ISABO-IBiLSTM model. The initial tool wear prediction ISABO-IBiLSTM model uses an IBiLSTM network as the base network and improves upon it using the ISABO algorithm. The IBiLSTM (Improved Bidirectional Long Short-Term Memory) model includes three layers of bidirectional long short-term memory neural network, one activation layer, one dropout layer, two fully connected layers, and one regression layer. Three layers of BiLSTM replace the single-layer BiLSTM layer of the base network to achieve deeper temporal feature extraction. A modified linear unit (ReLU) layer is added after the three BiLSTM layers to prevent gradient vanishing and overfitting. Furthermore, a dropout layer is added after the ReLU layer to shield some neurons in the neural network, improving the model's robustness. Finally, two fully connected layers are used to increase the model's representational power and match the output scale of the object detection network. The ISABO algorithm (Improved Subtraction-Average-Based Optimizer) is used to search for the optimal hyperparameters of the IBiLSTM model. The Adam solver is employed. The improved ISABO algorithm overcomes the shortcomings of the original algorithm, such as fast convergence but susceptibility to local optima, and can search for optimal hyperparameters, thereby further improving the model's prediction accuracy.

[0062] 6) The IBiLSTM model is trained based on the model-built dataset and the optimized hyperparameters to obtain the trained tool wear prediction model. During the training process, the root mean square error, mean absolute percentage error, and mean absolute error are calculated based on the predicted wear amount and the actual wear amount to evaluate the prediction accuracy of tool wear.

[0063] The mathematical formulas for root mean square error, mean absolute percentage error, and mean absolute error are as follows:

[0064]

[0065]

[0066]

[0067] In the formula, It is the root mean square error. Mean absolute percentage error It is the mean absolute error. It is the sample size. It predicts tool wear values. This is the actual tool wear value.

[0068] The smaller the root mean square error, mean absolute percentage error, and mean absolute error, the higher the prediction accuracy.

[0069] 7) Monitor the vibration signals of the worktable and the cutting force signals of the tool under test during the operation of the tool under test, and input the processed real-time signal feature set into the tool wear prediction model to obtain the wear state of the tool under test.

[0070] 8) Display the dynamic identification results of the wear status of the tool to be detected in real time.

[0071] Example 2:

[0072] The main content of this embodiment is the same as that of Embodiment 1, wherein, please refer to Table 1 and Figure 8 In step 1), the vibration signal measurement points are distributed on the workpiece. The vibration signals are acquired using a Kistler triaxial accelerometer. The Kistler triaxial accelerometer is fixed to the workpiece surface. The cutting force signal is acquired using a Kistler triaxial force gauge. The Kistler triaxial force gauge is installed between the workpiece and the machining table.

[0073] Table 1

[0074]

[0075] Example 3:

[0076] The main content of this embodiment is the same as that of Embodiment 2, except that an acoustic emission sensor is also installed in actual production. The Kistler acoustic emission (AE) sensor is installed on the workpiece to monitor the high-frequency stress waves generated during the cutting process.

[0077] Example 4:

[0078] The main content of this embodiment is the same as that of Embodiment 1, wherein, see [link / reference]. Figure 1 This embodiment provides a tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network, specifically including the following steps:

[0079] 1) Force and vibration sensors are installed at the workpiece fixture of a high-speed milling machine to collect force signals and acceleration vibration signals in the X, Y, and Z directions during machining. The actual tool wear value is then measured using a wear measuring instrument. A total of [value missing] is obtained. signal data and Actual wear and tear value.

[0080] 2) Since the tool has an entry and exit phase during a single pass, during which the tool does not contact the workpiece, the tool wear process cannot be reflected. Therefore, this embodiment adopts the truncation method, that is, using the upper quartile value of the original signal data as the boundary point of the first invalid data, that is, from the beginning of the original data backwards and from the end forwards, all data smaller than this boundary value are removed. Figure 2 and Figure 3 The comparison between the X-axis force signal before and after removing the cutting-in and cutting-out signal data is shown.

[0081] 3) Outliers in the signal data are caused by chatter and other factors during the cutting process. If these outliers are not removed, the accuracy of the prediction model will decrease. The Hampel filtering method used in this embodiment utilizes the sliding window mechanism and the principle of median filtering to replace data points in the signal that exceed three standard deviations with the average of 10,000 adjacent points. Figure 4 The image shows a comparison before and after removing invalid data (partial) from the X-axis force signal.

[0082] 4) Due to the huge amount of signal data, directly denoising the signal data would take a lot of time. Therefore, the signal data with outliers removed is downsampled at a ratio of 1 to 10.

[0083] 5) To more clearly analyze and study the useful information of signals in practical engineering, signal denoising is crucial. This embodiment uses the ICEEMDAN-improved wavelet mother function to perform wavelet threshold denoising on the sensor signal. The ICEEMDAN algorithm is used to decompose the signal, which solves the problems of residual noise and pseudo-mode components, obtaining better IMF components and thus improving the denoising effect. The wavelet threshold denoising function used is the sym18 function, where the threshold function is an improved function using a polynomial method. A comparison of the x-axis force signal of the first tool pass in dataset c1 before and after denoising is shown below. Figure 5 As shown.

[0084] 6) Even after preprocessing, each milling operation generates tens of thousands of data points, which do not correspond one-to-one with wear values ​​and cannot be directly used as features. It is necessary to "reduce" the information in the data to extract effective features. This embodiment extracts 11 time-domain features: absolute mean, peak value, root mean square value, root square amplitude, skewness value, kurtosis value, waveform factor, impulse factor, skewness factor, kurtosis factor, and margin factor; 4 frequency-domain features: centroid frequency, mean square frequency, root mean square frequency, and frequency variance; and energy features obtained after three-level wavelet decomposition, totaling 24 features. Since features of force signals and acceleration vibration signals in the X, Y, and Z directions are extracted, the total number of features is 144.

[0085] 7) Not every extracted feature is effective. Therefore, this embodiment uses both the Spearman correlation coefficient and the maximum mutual information correlation coefficient for feature selection. Features are retained when both are greater than 0.95; otherwise, they are discarded. The values ​​of the Spearman correlation coefficient and the maximum mutual information correlation coefficient for some features in the dataset are shown below. Figure 6 As shown. The final dataset yielded 25 features, which were used as input to the model.

[0086] 8) In this embodiment, after normalizing the signal data after the above preprocessing operations and the actual wear dataset of each milling cutter, 80% of them are randomly used as the training dataset and 20% as the validation dataset.

[0087] 9) The initial parameter settings for the ISABO-IBiLSTM model are as follows: population size is 25; maximum number of iterations is 50; upper limit of parameter optimization range is [0.1, 0.1, 300]; lower limit of parameter optimization range is [1×10]. -8 1×10 -4

[50] ; Maximum training iterations are 1000; Dropout layer retention ratio is 0.4; Adam solver is used for optimization; gradient threshold is 1; learning rate descent period is 350; learning rate descent factor is 0.2. Because the number of hidden layer neurons in the BiLSTM layer significantly affects the model's complexity and performance; the learning rate determines whether the optimal solution can be obtained; and the L2 regularization coefficient affects the degree of overfitting, the ISABO algorithm is used to optimize these three hyperparameters. The optimized hyperparameters are then input into the IBiLSTM model along with the training set for training to obtain the final result. The tool wear prediction process is as follows: Figure 1 As shown, the prediction results are as follows Figure 7 As shown.

[0088] 10) To verify the superiority of the ISABO-IBiLSTM model in predicting tool wear conditions, this paper uses root mean square error (RMSE), mean absolute percentage error (MASE), and mean absolute error (MAE) as criteria for model accuracy. The smaller the RMSE, MASE, and MAE, the higher the model's prediction accuracy. The mathematical formulas are as follows:

[0089]

[0090]

[0091]

[0092] The optimized hyperparameters of the ISABO-IBiLSTM model are shown in Table 2. The evaluation metrics for the dataset prediction results are shown in Table 3.

[0093] Table 2

[0094]

[0095] Table 3

[0096]

[0097] The average MAE of the ISABO-IBiLSTM model was reduced by 60.52%, and the prediction accuracy reached 98.83%. This embodiment improves the BiLSTM model structure, increasing its complexity, robustness, and representational capabilities. Combining the ISABO algorithm with the model avoids the decrease in prediction accuracy and stability caused by improper parameter settings. Furthermore, the improved BiLSTM model effectively prevents underfitting and gradient vanishing. Finding the optimal hyperparameter values ​​further enhances the model's prediction accuracy and stability. The ISABO-IBiLSTM model proposed in this embodiment is more effective and feasible for tool wear prediction.

[0098] Example 5:

[0099] This embodiment provides a computer-readable storage medium storing a computer program for implementing any one of the tool wear prediction methods based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network in embodiments 1 to 3.

[0100] Example 6:

[0101] This embodiment provides a tool wear prediction system implementing any one of the methods in Embodiments 1 to 3, including a data acquisition unit, a processor, an input unit, an output unit, and a storage unit. The data acquisition unit monitors the vibration signals of the worktable and the cutting force signals of the tool under test during operation and processes them into a real-time signal feature set. The input unit receives the real-time signal feature set. The output unit displays the dynamic identification results of the tool wear state in real time. The storage unit stores data and a sequence of instructions executed by the processor to implement the tool wear prediction model.

Claims

1. A tool wear prediction method based on an improved subtraction optimizer combined with an improved bidirectional long short-term memory neural network, characterized in that, Includes the following steps: 1) During the operation of the tool (1), the vibration signal and cutting force signal of the worktable (2) and the tool (1) are collected simultaneously; 2) Preprocess the signal data acquired in step 1); 2.1) Truncate the input and output portions of the acquired signal data; 2.2) Use the Hampel filtering method to remove outliers from the truncated signal data; 2.3) Downsampling is performed on the signal data after removing outliers; 2.4) The ICEEMDAN-modified wavelet mother function is used to decompose and denoise the downsampled signal data; 3) Extract the feature values ​​in the time domain, frequency domain, and time-frequency domain of the preprocessed signal data to obtain the original feature dataset; 4) Features are selected based on both Spearman correlation coefficient and maximum mutual information correlation coefficient. The selected feature values ​​are used as input to the prediction model to obtain a new feature dataset. The new feature dataset is then normalized with the actual tool wear dataset to obtain the model construction dataset. 5) Construct an initial tool wear prediction ISABO-IBiLSTM model; the initial tool wear prediction ISABO-IBiLSTM model uses the IBiLSTM network as the base network and is improved using the ISABO algorithm; the IBiLSTM model includes a three-layer bidirectional long short-term memory neural network, an activation layer, a dropout layer, two fully connected layers, and a regression layer; the ISABO algorithm is used to search for the optimal hyperparameters of the IBiLSTM model; the optimization solver uses the Adam solver; 6) The IBiLSTM model is trained based on the model-built dataset and the optimized hyperparameters to obtain the trained tool wear prediction model; During the training process, the root mean square error, mean absolute percentage error, and mean absolute error are calculated based on the predicted wear amount and the actual wear amount to evaluate the prediction accuracy of tool wear. 7) Monitor the vibration signals of the worktable and the cutting force signals of the tool under test during the operation of the tool under test, and input the processed real-time signal feature set into the tool wear prediction model to obtain the wear state of the tool under test.

2. The tool wear prediction method based on the improved subtractive optimizer combined with the improved bidirectional long short-term memory neural network according to claim 1, characterized in that: Vibration signal measurement points are distributed on the workpiece (3); vibration signals are collected by a triaxial accelerometer (4); the triaxial accelerometer (4) is fixed on the surface of the workpiece; cutting force signals are collected by a triaxial force measuring instrument (5); the triaxial force measuring instrument (5) is installed between the workpiece (3) and the machining table (2).

3. The tool wear prediction method based on the improved subtractive optimizer combined with the improved bidirectional long short-term memory neural network according to claim 1, characterized in that: In step 2.3), the sampling ratio is 1 to 10.

4. The tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network according to claim 1, characterized in that: In step 3), 11 time-domain features are extracted, including absolute mean, peak value, root mean square value, root mean square amplitude, skewness value, kurtosis value, waveform factor, impulse factor, skewness factor, kurtosis factor, and margin factor; 4 frequency-domain features, including centroid frequency, mean square frequency, root mean square frequency, and frequency variance; and 24 energy features obtained after three-level wavelet decomposition.

5. The tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network according to claim 1, characterized in that: Step 4) Spearman correlation coefficient and maximum mutual information correlation coefficient are used as feature selection criteria, with the following mathematical formulas: In the formula, The Spearman correlation coefficient. The difference in grade between the eigenvalue and the corresponding actual wear value. The number of samples; In the formula, For mutual information, For joint probability density, These are the characteristic value and the corresponding actual wear value, respectively. The maximum mutual information correlation coefficient is obtained from mutual information; The thresholds for Spearman correlation coefficient and maximum mutual information correlation coefficient were set to 0.95, and 24 feature values ​​were finally selected as inputs to the prediction model.

6. The tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network according to claim 1, characterized in that: In step 4), 80% of the model building dataset is randomly divided into training dataset and 20% into validation dataset.

7. The tool wear prediction method based on an improved subtraction optimizer and an improved bidirectional long short-term memory neural network according to claim 1, characterized in that: The ISABO algorithm is configured with a population size of 25; a maximum number of iterations of 50; an upper limit for the parameter optimization range of [0.1, 0.1, 300]; and a lower limit for the parameter optimization range of [1×10]. -8 1×10 -4 [50]; the maximum number of training iterations is 1000; the retention ratio of the Dropout layer is 0.

4.

8. A computer-readable storage medium, characterized in that: The computer program stores a method for predicting tool wear based on any one of the improved subtraction optimizers and improved bidirectional long short-term memory neural networks as described in claims 1 to 7.

9. A tool wear prediction system implementing any one of claims 1 to 7, characterized in that: It includes a data acquisition unit, a processor, an input unit, an output unit, and a storage unit. The data acquisition unit monitors the vibration signals of the worktable and the cutting force signals of the tool under test during operation and processes them into a real-time signal feature set. The input unit can receive the real-time signal feature set. The output unit displays the dynamic identification results of the wear state of the tool under test in real time. The storage unit is used to store data and the instruction sequence executed by the processor to implement the tool wear prediction model.

Images