A tool state detection method based on spindle current signal

Abstract

The application provides a tool state detection method based on a main shaft current signal, comprising the following steps: designing an orthogonal experiment related to a cutting speed, a feed rate and a cutting depth; collecting a single-phase current of a main shaft motor driver by using a current sensor; collecting a wear amount of a rear tool face of a milling cutter after each experiment, fitting a tool wear curve according to the collected wear amount; performing denoising on the collected original signal of the single-phase current of the main shaft motor driver based on a third-fourth quantile method of a sliding window; extracting time-frequency domain features and time-frequency joint domain features of the single-phase current signal of the main shaft motor driver based on the preprocessed data; selecting n features from the time-frequency domain features and the time-frequency joint domain features as input of a prediction model by using grey correlation degree analysis; establishing a prediction model based on a least square support vector machine regression model, optimizing the prediction model by using a genetic algorithm, and predicting the tool wear amount based on the optimized prediction model.

Description

Technical Field

[0001] This invention relates to the field of tool condition detection technology, and specifically to a tool condition detection method based on spindle current signals. Background Technology

[0002] With the continuous development of the manufacturing industry, modern machinery manufacturing needs to meet the requirements of higher quality, higher efficiency, lower cost, and lower energy consumption. Metal cutting is the foundation of modern manufacturing, and tool failure is one of the most common and costly failures in metal cutting. Tools can fail in various forms, leading to a decline in workpiece quality and production efficiency, and in severe cases, even causing workpiece scrap or machine tool damage. Tool failure accounts for 20% of total machine tool downtime. According to relevant statistics, my country's annual expenditure of tens of billions of yuan on tools significantly impacts machine tool costs and labor costs, which account for 20% and 38% of manufacturing costs, respectively. However, for machine tools equipped with a mature and accurate tool condition monitoring system, both downtime and manufacturing costs can be significantly reduced. Existing methods for monitoring tool condition using sensor signals generally employ cutting force sensors or vibration and acoustic emission sensors. While cutting force sensors directly reflect tool wear, they suffer from drawbacks such as high cost and poor industrial applicability. Vibration and acoustic emission sensors, although economical and widely applicable, cannot meet the conditions of operation when cutting fluid is turned on. Dry cutting is rare in actual production, making it difficult to generalize the resulting predictive models to large-scale industrial production. Furthermore, existing vibration and cutting force signal data have high dimensionality and large data volume, failing to meet the real-time requirements for tool condition monitoring in actual production. Summary of the Invention

[0003] This invention addresses the problem that current tool monitoring technology prediction models are not widely applicable and are based on experimental data from a single tool under a single working condition. It proposes a tool condition detection method based on spindle current signals.

[0004] To solve the above problems, the present invention adopts the following technical solution:

[0005] A tool condition detection method based on spindle current signal includes the following steps:

[0006] Step 1: Design an orthogonal experiment related to cutting speed, feed rate, and depth of cut:

[0007] Step 2: Use a current sensor to collect the single-phase current of the spindle motor driver;

[0008] Step 3: After each experiment, collect the wear amount of the cutting edge back face of the milling cutter, fit the tool wear curve based on the collected wear amount, and divide it into three wear stages: initial rapid wear stage, normal wear stage, and rapid wear stage;

[0009] Step 4: Denoise the raw single-phase current signal of the acquired spindle motor driver based on the third quartile method using a sliding window;

[0010] Step 5: Extract the time-domain, frequency-domain, and time-frequency joint-domain features of the single-phase current signal of the spindle motor driver based on the preprocessed data;

[0011] Step 6: Using grey relational analysis, select n features from the time-domain features, frequency-domain features, and time-frequency joint domain features as inputs to the prediction model;

[0012] Step 7: Establish a prediction model based on the least squares support vector machine regression model, optimize the prediction model through a genetic algorithm, and predict the tool wear based on the optimized prediction model.

[0013] Furthermore, in step 3, the time-domain features include root mean square, variance, maximum value, skewness, kurtosis and peak-to-peak value; the frequency-domain features include spectral skewness coefficient, spectral kurtosis coefficient and spectral power; and the time-frequency joint domain features include wave packet energy basis.

[0014] Furthermore, step 4 is detailed below:

[0015] 4.1 Sort all values ​​from smallest to largest, divide all data into 4 equal parts, take the value at the 25th percentile as the upper quartile value, denoted as Q1, the value at the 50th percentile as the median, denoted as Q2, and the value at the 75th percentile as the lower quartile, denoted as Q3; set Q3 as the invalid data threshold, and then compare sequentially from the beginning of the data to the first data larger than Q3;

[0016] 4.2 Set the sample size to k, the window size to 2k+1, and the upper and lower bound coefficients to n. δ Calculate the sample local standard deviation x based on the sliding window. δ Compared with the local estimated median x m Calculate the upper and lower bounds of outliers in the sample:

[0017] Outlier upper bound: upbound = x m +n δ ×x δ

[0018] Lower bound of outliers: downbound = x m -n δ ×x δ

[0019] If the sample value is not within the upper or lower bound, the estimated median value is used to replace the value.

[0020] Furthermore, step 5 consists of the following specific steps:

[0021] 5.1 Extracting Time-Domain Features Based on Statistics:

[0022] Root mean square:

[0023] variance:

[0024] Maximum value: Max = max(X);

[0025] Minimum value: Min = min(X);

[0026] Peak-to-peak value: PP = max(X) - min(X);

[0027] Skewness:

[0028] Kuroshi: μ4 is the fourth-order center distance, δ 4 Standard deviation;

[0029] 5.2 Processing current signals using Discrete Fourier Transform: The amplitude of a current signal can be expressed as: The phase is: Frequency domain features extracted from amplitude spectrum: spectral skewness coefficient: Spectral peak state coefficient: Spectral power:

[0030] 5.3 The current signal is passed through a set of wavelet packet function bases and then through a low-pass filter h. k and high-pass filter g k The low-frequency and high-frequency components are obtained by decomposing and downsampling. The low-frequency components are used as input signals and wavelet decomposition is performed to obtain the low-frequency and high-frequency components of the next layer. As the number of wavelet decomposition layers increases, the frequency domain resolution also increases. At the same time, wavelet decomposition is performed on the detail band obtained by high-pass filtering. The optimal decomposition path of the signal is calculated through the minimum cost function, and the original signal is decomposed according to the path.

[0031] Choosing the Dobesi wavelet as the wavelet function basis, we obtain the joint time-frequency domain features;

[0032] The adjustment function in multiscale analysis is expressed by the following formula for each level:

[0033]

[0034] Where (a0,…,a N-1 ) is a finite-length real number sequence, and is an adjustment coefficient;

[0035] Meanwhile, the wavelet function can also be used to adjust the linear combination of functions as follows:

[0036]

[0037] Where (b0,…,b) M-1 ) is a finite-length real number sequence, and is the wavelet coefficient.

[0038] Furthermore, step 6 is detailed below:

[0039] 6.1 Identify the sequence to be analyzed:

[0040] The tool flank wear amount is selected as the parent sequence, and the time domain characteristics, frequency domain characteristics, and time-frequency joint domain characteristics of the current signal are selected as the sub-sequences.

[0041] 6.2 Normalization of Data Sequences:

[0042] Each number in the data sequence is the first data element, resulting in a new sequence: Given a data sequence X = {x(k) / k = 1, 2, ..., n}, the initial values ​​are:

[0043] 6.3 Calculate the grey relational coefficient:

[0044] Substitute the subsequence and the parent sequence into the formula:

[0045]

[0046] 6.4 Calculate the mean of the correlation coefficients to form a correlation sequence, and select the top n features of the correlation sequence as input to the prediction model.

[0047] Furthermore, step 7 is detailed below:

[0048] 7.1 Splitting the dataset: The selected n features are used as the dataset, and the dataset is divided into a training set and a validation set according to the proportions.

[0049] 7.2 Determine the value ranges of the penalty coefficient C and kernel parameter σ of the least squares support vector regression machine model: the penalty parameter ranges from [0.1, 1000], and the radial basis function kernel parameter ranges from [0.01, 100].

[0050] 7.3 Using binary encoding, the fitness function is determined as follows: Where: len is the number of samples, y ij For training output values, y j The expected output value;

[0051] 7.4 Randomly generate N individuals to form the initial training population;

[0052] 7.5 The initialized population was trained using a least squares support vector regression machine program, and the fitness of the trained model was evaluated using the LS-SVM model.

[0053] 7.6 Determine the termination condition. If the iteration termination requirement is met, stop; if the termination condition is not met, perform binary league selection, crossover, and mutation on the population, and repeat step 7.5.

[0054] 7.7 Substitute the optimal parameter penalty coefficient C and kernel function σ into the least squares support vector regression machine program, use the tool wear data in the validation set for validation, use the validated least squares support vector regression machine program as the prediction model, and use the prediction model to predict the tool wear.

[0055] The beneficial effects of this invention are as follows:

[0056] By repeatedly conducting experiments to collect a large amount of data related to the cutting tool under different machining conditions, the spindle current signal features were extracted and the grey relational algorithm was used to fully explore the optimal feature parameters related to tool wear in the spindle current signal. This improved the prediction accuracy of the support vector machine prediction model in fitting the tool failure curve and shortened the prediction time. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating an embodiment of the method of the present invention.

[0058] Figure 2 This is a flowchart illustrating the prediction of tool wear using spindle current according to an embodiment of the present invention.

[0059] Figure 3 This is a flowchart of the least squares support vector regression model optimized based on genetic algorithm in an embodiment of the present invention. Detailed Implementation

[0060] To further understand the present invention, preferred embodiments of the present invention are described below in conjunction with examples. However, it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, and are not intended to limit the scope of the claims of the present invention.

[0061] This invention provides a tool condition detection method based on spindle current signals. The method involves repeatedly performing trial machining with multiple tools under various working conditions on a machine tool equipped with sensors. By collecting spindle current signals and measuring tool wear data during machining, and performing data preprocessing, the method extracts data features from the preprocessed data and fully mines the optimal feature parameters related to tool wear in the current signals. This improves the prediction accuracy and real-time performance of the tool wear prediction model based on support vector machine regression.

[0062] The aforementioned sensors are Hall current sensors, which are respectively installed at the three-phase current input terminals of the spindle motor servo driver.

[0063] The aforementioned data preprocessing mainly involves two categories: invalid data and outlier data. Invalid data specifically includes invalid data from tool entry and exit, while outlier data includes noise, missing values, and data inconsistencies caused by factors such as sensor malfunctions, human error, system errors, multiple heterogeneous data sources, and network transmission disorder. For invalid data, since the sampling frequency is high and the data volume is sufficient for training the model, the third quartile method can be used to determine the critical value. Data before the critical value is considered invalid and can be directly deleted. Outlier data can be categorized based on its characteristics, such as point outliers, fluctuating values, collective outliers, and significant noise signals. The key to handling outliers lies in how to identify them, and the median filtering (Hampel filtering) method based on a sliding window is selected.

[0064] The aforementioned third quartile method refers to: calculating the third quartile of the data collected during the cutting process as the threshold for invalid data; starting from the first data collected, comparing the process data sequentially until the first value greater than the threshold appears; data from the first number up to that position is considered invalid data and can be directly deleted.

[0065] The above-described median filtering (Hampel filtering) method based on a sliding window: sets the number of samples to k, the window size to 2k+1, and the upper and lower bound coefficients to n. δ Calculate the local standard deviation x for each sample. δ Local estimated median x m Calculate the upper and lower limits of outliers for the sample. If the sample value is greater than or less than the upper or lower limit, then use the estimated median x. m Replace the sample data.

[0066] The above data feature extraction is as follows: extracting time-domain features based on statistics, frequency-domain features based on spectrum analysis, and time-frequency joint-domain features based on wavelet packet energy from the current signal.

[0067] The aforementioned statistical time-domain features are: new data extracted by using statistical methods to calculate the root mean square, variance, maximum value, minimum value, skewness, kurtosis, peak-to-peak value, etc. of the sample dataset.

[0068] The frequency domain characteristics based on the above spectrum analysis are as follows: spectrum analysis is the most commonly used and most important analysis method in frequency domain analysis. The amplitude spectrum is obtained by performing Fourier transform on the acquired current signal to obtain the spectral skewness coefficient, spectral peak coefficient and spectral power.

[0069] The above-mentioned time-frequency joint domain features based on wavelet packet energy are as follows: based on wavelet transform, the detail subbands are further decomposed, and finally the optimal signal decomposition path is calculated through the minimum cost function. The original input Sinha is decomposed and features are extracted using this path, and the Dobesi wavelet is selected as the wavelet function basis for decomposing the signal.

[0070] The aforementioned optimal feature parameters refer to the ten data features extracted through time-frequency domain analysis using the grey relational analysis algorithm, which are designated as subsequences. Tool wear is taken as the parent sequence. The mean of each feature is calculated, and then each element in the feature is divided by the mean of that index. The next step is to calculate the correlation coefficient between each feature in the subsequence and the parent sequence. The top five data features with the highest correlation coefficients are then selected.

[0071] The aforementioned support vector machine regression prediction model is preferably a least squares support vector machine regression model optimized by a genetic algorithm. A portion of the tool travel data is used as the training set, and the remaining portion as the test set. The mean square error of the tool wear prediction model is used as the population fitness. Selection, crossover, and mutation operations are performed to generate a new population, and the fitness of each individual is calculated. The optimal individual is selected and compared with the optimal fitness, and iterative training is conducted to obtain the optimal prediction model.

[0072] Example:

[0073] like Figure 1 and Figure 2 The diagram shows the specific flow of a tool condition detection method based on spindle current signals, which includes the following steps:

[0074] Step 1: Design an orthogonal experiment related to cutting speed, feed rate, and depth of cut: Four cutting speed settings were used: 100m / min, 200m / min, 300m / min, and 400m / min; four feed rate settings were used: 0.1mm / r, 0.2mm / r, 0.3mm / r, and 0.4mm / r; and the depth of cut was set to 0.5mm, 1mm, 1.5mm, and 2mm. Under different parameter settings, three Huhao 8mm four-flute end mills were used, with each feed length being 100mm. A total of 16 sets of experiments were set up, with each set of experiments performed 5 times, resulting in a total of 240 sets of experimental data.

[0075] Step 2: Use the Hangzhi AIT60-10V current sensor to collect the single-phase current of the spindle motor driver. The current signal sampling frequency is 5kHz.

[0076] Step 3: After each experiment, stop the machine and remove the tool once. Use a microscope to measure the wear amount on the flank face of the milling cutter's cutting edge. Based on the trend of the tool wear curve fitted to the wear amount, it can be divided into three wear stages: the initial rapid wear stage, the normal wear stage, and the acute wear stage.

[0077] Step 4: Data preprocessing is required for the acquired raw spindle current signal:

[0078] (1) The acquired spindle current signal needs to be processed using the third quartile method. First, arrange all values ​​from smallest to largest and divide all data into four equal parts. Take the value at the 25th percentile as the upper quartile value, denoted as Q1; the value at the 50th percentile as the median, denoted as Q2; and the value at the 75th percentile as the lower quartile, denoted as Q3. Set Q3 as the invalid data threshold and then compare the data sequentially from the beginning (end) to the end (front), truncating the first data up to the first data larger than Q3.

[0079] (2) After deleting invalid data, abnormal data needs to be processed. First, set the number of samples to k, the window size to 2k+1, and the upper and lower bound coefficients to n. δ， Then calculate the sample local standard deviation x based on the sliding window. δ Compared with the local estimated median x m Calculate the upper and lower bounds of outliers in the sample:

[0080] Outlier upper bound: upbound = x m +n δ ×x δ

[0081] Lower bound of outliers: downbound = x m -n δ ×x δ

[0082] If the sample value is not within the upper or lower bound, the estimated median value is used to replace the value.

[0083] Step 5: Extract the time-frequency domain features and time-frequency joint domain features of the spindle current signal. The time-domain features include root mean square, variance, maximum value, skewness, kurtosis, and peak-to-peak value. The frequency domain features are first obtained by performing a discrete Fourier transform to obtain the spectrum, and the extracted features are spectral skewness coefficient, spectral kurtosis coefficient, and spectral power.

[0084] (1) Statistical-based temporal feature extraction:

[0085] Root mean square:

[0086] variance:

[0087] Maximum value: Max = max(X);

[0088] Minimum value: Min = min(X);

[0089] Peak-to-peak value: PP = max(X) - min(X);

[0090] Skewness:

[0091] Kuroshi: μ4 is the fourth-order center distance, δ 4 Standard deviation;

[0092] (2) Processing current signals using Discrete Fourier Transform: The amplitude of a current signal can be expressed as: The phase is: The following feature is extracted from the amplitude spectrum: spectral skewness coefficient: Spectral peak state coefficient: Spectral power:

[0093] (3) Pass the current signal through a set of wavelet packet function bases and a low-pass filter h k and high-pass filter g k The signal is decomposed and downsampled to obtain low-frequency and high-frequency components. The low-frequency components are then used as input signals for wavelet decomposition to obtain the next layer of low-frequency and high-frequency components. As the number of wavelet decomposition layers increases, the frequency domain resolution also increases. Simultaneously, wavelet decomposition is performed on the detail bands obtained from high-pass filtering. The optimal decomposition path is calculated using the minimum cost function, and the original signal is decomposed according to this path. This invention selects the Dobese wavelet as the wavelet function basis to obtain the joint time-frequency domain features.

[0094] First, the adjustment function in multi-scale analysis can be expressed as follows for each level:

[0095]

[0096] Where (a0,…,a N-1 ) is a finite-length real number sequence, and is the adjustment coefficient.

[0097] Meanwhile, the wavelet function can also be used to adjust the linear combination of functions as follows:

[0098]

[0099] Where (b0,…,b) M-1 ) is a finite-length real number sequence, and is the wavelet coefficient.

[0100] Step 6: Using grey relational analysis, select five features as inputs for the prediction model: (1) Establish the sequence to be analyzed:

[0101] The tool flank wear amount is selected as the parent sequence, and the time domain characteristics, frequency domain characteristics, and time-frequency joint domain characteristics of the current signal are selected as the sub-sequences.

[0102] (2) Normalization of data sequences:

[0103] Each number in the data sequence is the first data element, resulting in a new sequence: Given a data sequence X = {x(k) / k = 1, 2, ..., n}, the initial values ​​are:

[0104] (3) Calculate the grey relational coefficient:

[0105] Substitute the subsequence and the parent sequence into the formula:

[0106]

[0107] (4) Calculate the mean of the correlation coefficients to form a correlation sequence, and select the top five features of the correlation sequence as inputs to the prediction model.

[0108] Step 7: Establish a least-squares support vector machine regression model based on genetic algorithm optimization

[0109] (1) Split the dataset: Divide the dataset into 70% training set and 30% validation set according to the proportion;

[0110] (2) Determine the value ranges of the penalty coefficient C and kernel parameter σ in the least squares support vector regression model: the penalty parameter ranges from [0.1, 1000], and the radial basis function kernel parameter ranges from [0.1, 1000].

[0111] [0.01,100];

[0112] (3) Using binary encoding, the fitness function is determined as follows: Where: len is the number of samples, y ij For training output values, y j This is the expected output value.

[0113] (4) Randomly generate N individuals to form the initial training population.

[0114] (5) The initialized population was trained using the least squares support vector regression machine program, and the fitness of the trained model was evaluated using the LS-SVM model.

[0115] (6) Determine the termination condition. If the iteration termination requirement is met, stop; if the termination condition is not met, perform binary league selection, crossover and mutation on the population and repeat (5).

[0116] (7) Substitute the optimal parameter penalty coefficient C and kernel function σ into the least squares-based support vector regression machine program to predict the tool wear data in the validation set.

[0117] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A tool condition detection method based on spindle current signal, characterized in that, Includes the following steps: Step 1: Design orthogonal experiments related to cutting speed, feed rate, and depth of cut; Step 2: Use a current sensor to collect the single-phase current of the spindle motor driver; Step 3: After each experiment, collect the wear amount of the cutting edge back face of the milling cutter, fit the tool wear curve based on the collected wear amount, and divide it into three wear stages: initial rapid wear stage, normal wear stage, and rapid wear stage; Step 4: Denoise the raw single-phase current signal of the acquired spindle motor driver based on the third quartile method using a sliding window; Step 5: Extract the time-domain, frequency-domain, and time-frequency joint-domain features of the single-phase current signal of the spindle motor driver based on the preprocessed data; Step 6: Using grey relational analysis, select n features from the time-domain features, frequency-domain features, and time-frequency joint domain features as inputs to the prediction model. The specific steps are as follows: 6.1 Identify the sequence to be analyzed: The tool flank wear amount is selected as the parent sequence, and the time domain characteristics, frequency domain characteristics, and time-frequency joint domain characteristics of the current signal are selected as the sub-sequences. 6.2 Normalization of Data Sequences: Each number in the data sequence is the first data element, resulting in a new sequence: Given a data sequence X = {x(k) / k = 1, 2, ..., n}, the initial values ​​are: ; 6.3 Calculate the grey relational coefficient: Substitute the subsequence and the parent sequence into the formula: 6.4 Calculate the mean of the correlation coefficients to form a correlation sequence, and select the top n features of the correlation sequence as input to the prediction model; Step 7: Establish a prediction model based on the least squares support vector machine regression model, optimize the prediction model through a genetic algorithm, and predict the tool wear based on the optimized prediction model.

2. The tool condition detection method based on spindle current signal as described in claim 1, characterized in that, In step 5, the time-domain features include root mean square, variance, maximum value, skewness, kurtosis and peak-to-peak value; the frequency-domain features include spectral skewness coefficient, spectral kurtosis coefficient and spectral power; and the time-frequency joint domain features include wave packet energy basis.

3. The tool condition detection method based on spindle current signal as described in claim 1, characterized in that, Step 4 is detailed below: 4.1 Sort all values ​​from smallest to largest, divide all data into 4 equal parts, take the value at the 25th percentile as the upper quartile value, denoted as Q1, the value at the 50th percentile as the median, denoted as Q2, and the value at the 75th percentile as the lower quartile value, denoted as Q3; set Q3 as the invalid data threshold, and then compare sequentially from the beginning of the data to the first data larger than Q3; 4.2 Set the sample size to k, the window size to 2k+1, and the upper and lower bound coefficients to k. Calculate the sample local standard deviation based on the sliding window. Compared with the local estimated median Calculate the upper and lower bounds of outliers in the sample: Upper bound of outliers: Lower bound of outliers: If the sample value is not within the upper or lower bound, the estimated median value is used to replace the value.

4. The tool condition detection method based on spindle current signal as described in claim 1, characterized in that, Step 5 is detailed below: 5.1 Extracting Time-Domain Features Based on Statistics: Root mean square: , variance: ; Maximum value: ; Minimum value: ; Peak-to-peak value: ; ； Kuroshi: The fourth-order center distance, Standard deviation; 5.2 Processing current signals using Discrete Fourier Transform: The amplitude of the current signal can be expressed as: The phase is: Frequency domain features are extracted using the amplitude spectrum. Spectral skewness coefficient: Spectral peak state coefficient: Spectral power: ; 5.3 Pass the current signal through a set of wavelet packet function bases and then through a low-pass filter. and high-pass filter Decompose the downsampling to obtain low-frequency and high-frequency components; The low-frequency component is used as the input signal and wavelet decomposition is performed to obtain the low-frequency and high-frequency components of the next layer. As the number of wavelet decomposition layers increases, the frequency domain resolution also increases. At the same time, wavelet decomposition is performed on the detail band obtained by high-pass filtering. The optimal decomposition path of the signal is calculated through the minimum cost function, and the original signal is decomposed according to the path. Choosing the Dobesi wavelet as the wavelet function basis, we obtain the joint time-frequency domain features; The adjustment function in multiscale analysis is expressed by the following formula for each level: Where (a0,…,a) N-1 ) is a finite-length real number sequence, and is an adjustment coefficient; Meanwhile, the wavelet function can also be used to adjust the linear combination of functions as follows: Where (b0,…,b) M-1 ) is a finite-length real number sequence, and are wavelet coefficients.

5. The tool condition detection method based on spindle current signal as described in claim 1, characterized in that, Step 7 is detailed below: 7.1 Splitting the dataset: The selected n features are used as the dataset, and the dataset is divided into a training set and a validation set according to the proportions. 7.2 Determine the value ranges of the penalty coefficient C and kernel parameter σ of the least squares support vector regression machine model: the penalty parameter ranges from [0.1, 1000], and the radial basis function kernel parameter ranges from [0.01, 100]. 7.3 Using binary encoding, the fitness function is determined as follows: in: len is the number of samples. To train the output value, The expected output value; 7.4 Randomly generate N individuals to form the initial training population; 7.5 The initialized population was trained using a least squares support vector regression machine program, and the fitness of the trained model was evaluated using the LS-SVM model. 7.6 Determine the termination condition. If the iteration termination requirement is met, stop; if the termination condition is not met, perform binary league selection, crossover, and mutation on the population, and repeat step 7.

5. 7.7 Substitute the optimal parameter penalty coefficient C and kernel function σ into the least squares support vector regression machine program, use the tool wear data in the validation set for validation, use the validated least squares support vector regression machine program as the prediction model, and use the prediction model to predict the tool wear.

Images