Tool Remaining Life Prediction Method Based on Multi-Sensor Information Fusion

Abstract

The present invention discloses a method for predicting the remaining tool life by multi-sensor information fusion. The use of features extracted from multi-sensor signals for fusion can reduce the dependence on a single signal, avoid the prediction instability caused by relying on a single signal, and improve the robustness of the prediction model. The prediction method based on the double-exponential degradation model adopted by the present invention can avoid relying on a large amount of full-life data, and can, along with the passage of time and the sequential availability of monitoring data, use the monitoring data to update the degradation model parameters in real time online to gradually approximate the tool wear degradation trend. At the same time, the remaining life at each moment is iteratively estimated, showing good adaptability to new prediction tasks.

Description

Technical Field

[0001] The present invention belongs to the field of tool life prediction methods for numerical control machine tools, and particularly relates to a tool remaining life prediction method based on multi-sensor information fusion. Background Art

[0002] As the part that directly contacts the workpiece during the machining process of a numerical control machine tool, the tool plays a directly crucial role in the product quality. Tool wear and degradation will cause changes in the tool shape and performance, which will directly affect the workpiece quality and accuracy. Therefore, monitoring the tool state during the machining process and accurately predicting its available remaining life to ensure timely tool replacement before tool failure is of great practical significance for cost reduction, efficiency improvement, and ensuring machining quality at the same time.

[0003] Remaining life prediction can be roughly divided into two categories: data-driven and degradation model-based. Data-driven remaining life prediction mainly refers to establishing the corresponding relationship between the equipment monitoring data or the extracted features at a certain moment and the remaining life through self-learning based on deep learning theory. The Institute of Electrical Engineering Cloud (Beijing) Technology Co., Ltd. proposed a tool remaining life prediction method based on the CDBN-BiLSTM model in its patent document "Tool Remaining Life Prediction Method, Device and Medium" (Patent Application No.: CN202011430089.7, Publication No.: CN114676716A). The steps of this method are as follows: First, collect tool state signals to obtain a training set containing training samples and a test set containing test samples. Then, extract training samples from the training set and input them into a hybrid model initialized with a convolutional deep belief network CDBN model and a bidirectional long short-term memory BiLSTM model to train the hybrid model. Finally, input the test samples into the trained hybrid model to output the prediction result of the tool remaining life. The deep learning model in this method can automatically extract features from a large amount of data, and the deep network structure endows it with powerful non-linear learning ability, improving the prediction accuracy of the tool remaining life. However, the remaining life prediction method based on deep learning needs to rely on the full-life data of a large number of same-type devices under similar working conditions to train a prediction model that meets the accuracy requirements. In actual situations, it is difficult to obtain full-life data, and once the external environment, working conditions, etc. change, whether the offline-trained prediction model can well adapt to the new prediction task remains to be further studied and verified. At the same time, only the vibration signal during the machining process is considered in this method, and the wear conditions of the tool reflected by different signals are not considered. There are problems such as low data utilization rate and poor model generalization ability when establishing a prediction model with a single sensor signal. Summary of the Invention

[0004] The object of the present invention is to provide a method for predicting the remaining tool life by multi-sensor information fusion, aiming to solve the problem of poor generalization ability of the prediction model established by using a single sensor signal and deep learning method in the prior art.

[0005] The technical solution adopted by the present invention is as follows: a method for predicting the remaining tool life by multi-sensor information fusion, and the specific operation steps are as follows:

[0006] Step 1: Collect signals during the tool cutting process. The signals include cutting force signals in the X, Y, and Z directions, vibration signals in the X, Y, and Z directions, and acoustic emission signals, a total of 7 channels of signals.

[0007] Step 2: Preprocess the collected signals. Among them, directly delete invalid data, and the invalid data includes invalid data during tool feeding and tool retracting; process outliers by using a median filtering method based on a sliding window.

[0008] Step 3: From three aspects of time domain, frequency domain, and time-frequency domain, divide the time series data of the 7 channels obtained in Step 1 into 315 equal-length data samples, and respectively extract 15 features including mean, standard deviation, skewness, kurtosis, impulse factor, crest factor, shape factor, margin factor, peak-to-peak value, root mean square, sk mean, sk standard deviation, sk skewness, sk kurtosis, and wavelet packet energy, and construct a time series of 105 features to form a feature matrix X T×105 , denoted by the i-th column feature time series of X 315×105 .

[0009] Step 4: Use the monotonicity evaluation index to calculate the Spearman rank correlation coefficient between each feature time series obtained in Step 3 and the corresponding time vector as the monotonicity score of this feature, and select the features with a monotonicity score greater than 0.8 to form a screened feature matrix X T×S ;

[0010] Step 5: Perform fusion on the features screened in Step 4 based on principal component analysis PCA, take the first principal component and perform triple first-order exponential smoothing on it, with a smoothing coefficient α = 0.1, to construct a health index; fit the health index by using a double-exponential model to establish a double-exponential degradation model describing the tool degradation process.

[0011] Step 6: According to the double-exponential degradation model, combine the particle filter algorithm and Bayesian theory to predict the remaining tool life.

[0012] The present invention is also characterized in that

[0013] The specific process of the median filtering in Step 2 is as follows:

[0014] 1) Set the number of samples on both sides of the sample to k, then the window size is 2k + 1, and set the upper and lower bound coefficients n δ ;

[0015] 2) Calculate the local standard deviation x of each sample based on the sliding window δ and the local estimated median x m ;

[0016] 3) Calculate the upper and lower bounds of the outliers of the sample:

[0017] Upper bound of outliers: upbound = x m + n δ × x δ

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

[0019] 4) If the sample value is greater than the upper bound of outliers or less than the lower bound of outliers, then use the estimated median x m to replace the sample.

[0020] The monotonicity evaluation index in step 4 is specifically:

[0021] For The monotonicity score is calculated as shown in the following formula:

[0022]

[0023] In the formula, r k is an element in the vector R, and T represents the time unit;

[0024] In the formula, is sorting of.

[0025] The feature fusion based on the principal component analysis PCA in step 5 is specifically as follows:

[0026] 1) Before fusing the screened feature matrix X 315×S using PCA, perform standardization processing on X 315×S using the z - score method, denoted as

[0027] 2) Calculate the covariance matrix of

[0028] 3) Find the eigenvalues of the covariance matrix ;

[0029] The eigenvalues of the covariance matrix are assumed to be λ​i , then λ i represents the variance of the i-th principal component, and λ i corresponding eigenvector is the coefficient of the principal component with respect to the original variables;

[0030] 4) Take the eigenvector corresponding to the maximum variance For reconstruction, that is At this point, the original M S dimensional features are reduced to one dimension;

[0031] 5) Perform triple first-order exponential smoothing on x 315×1 =[x0, x1,..., x 315 , as shown in the following formula (2), where α is the smoothing coefficient, y t are the smoothing values of the first, second, and third times respectively. The time series composed of the third smoothing value, and the health index values of the corresponding 315 data samples are Y {1:315} ={y1,..., y t ,..., y 315 ;

[0032]

[0033] Double-exponential degradation model, as follows:

[0034] y = ae bt + ce dt (3)

[0035] In the formula: y is the health index representing tool degradation constructed from the monitoring signal; a, b, c, d are the model parameters of the double-exponential degradation model, controlling the trend of the degradation trajectory; t is the time.

[0036] The tool remaining life prediction process is as follows:

[0037] 1) Establish the state equation and the observation equation

[0038] Take the Gaussian random walk model as the state equation, as shown in the following formula (1):

[0039]

[0040] In the formula: ω a , ω b , ω c , ω d are the modeling errors of the state equation respectively, conforming to a mean of 0 and variances of σ a , σ b , σ c , σ dGaussian distribution; then according to the double-exponential degradation model parameter modeling method, the observation equation is obtained as shown below:

[0041] HI(t) = a(t)e b(t)·t + c(t)·e d(t)·t + v(t), v(t) ~ N(0, σ t )(5)

[0042] where ν(t) is the observation noise of the system, assumed to be Gaussian white noise, conforming to a Gaussian distribution with a mean of 0 and a variance of σ t ; then the tool wear condition at any time t can be represented by the observation function HI(t) of the tool wear degradation trend;

[0043] 2) Offline training: Use the raw data collected from the first 150 cuttings as training data, perform signal preprocessing, then perform feature extraction, feature screening, and health index construction, and bring the constructed health index representing tool wear degradation into the double-exponential degradation model for model parameter initialization; Use the least squares function to fit the observed data to roughly determine the parameter distribution range; At the same time, initialize all parameters, generate the probability distribution of each parameter according to its initial distribution type, and obtain the initial distribution particles;

[0044] 3) Online update: Use the particle filter algorithm to perform real-time updates on the double-exponential degradation model parameters and the health status evaluation results according to the real-time monitoring data of the health index, and adjust the double-exponential degradation model parameter distribution according to the weight size;

[0045] 4) Life prediction: Calculate the tool wear health index according to the real-time update results of the double-exponential degradation model parameters, predict the degradation trend of the tool wear health index from 151 to 315 times by setting the failure threshold, and calculate the remaining life of the tool.

[0046] The specific implementation steps of the particle filter algorithm are as follows:

[0047] (1) Particle set initialization, t = 1: For i = 1, 2,..., N, generate N sampling particles from the prior distribution p(x0)

[0048] (2) For t = 2, 3,..., loop and execute the following steps:

[0049] 1) Importance sampling: For i = 1, 2,..., N, generate sampling particles from the importance probability density function Calculate the particle weights and normalize them;

[0050] 2) Resampling: Resample the particle set The resampled particle set is

[0051] 3) Output: Calculate the state estimate value at time k:

[0052] Based on the above two types of existing problems, the present invention proposes a tool remaining life prediction method based on multi - sensor information fusion and Bayesian theory. Using the features extracted from multi - sensor signals for fusion can reduce the dependence on a single signal, avoid the prediction instability caused by relying on a single signal, and improve the robustness of the prediction model. Using the prediction method based on the degradation model can avoid relying on a large amount of full - life data, and can, along with the passage of time and the sequential availability of monitoring data, use the monitoring data to update the degradation model parameters in real - time online to gradually approximate the tool wear degradation trend, and at the same time iteratively estimate the remaining life at each moment, showing good adaptability to new prediction tasks.

[0053] The beneficial effects of the present invention are:

[0054] The tool remaining life prediction method with multi - sensor information fusion provided by the present invention relates to a tool remaining life prediction method technology based on multi - sensor information fusion and Bayesian theory in the field of numerical control machine tool tool life prediction, and can be used to predict the remaining life of numerical control machine tool tools, solving the problems that the tool remaining life prediction model has poor generalization ability and limited adaptability to new prediction tasks.

[0055] The present invention collects cutting force, vibration, and acoustic emission signals during the working process of the numerical control machine tool as monitoring signals, and uses multi - sensor information to establish a tool remaining life prediction model, fully considering the tool wear conditions reflected by different types of signals, effectively overcoming the limitations of using a single signal to establish a prediction model in the prior art, and improving the generalization ability of the tool remaining life prediction model.

[0056] The present invention comprehensively extracts the features of cutting force, vibration, and acoustic emission signals from three aspects: time domain, frequency domain, and time - frequency domain, uses the monotonicity evaluation index, and conducts feature screening based on the Spearman rank correlation coefficient, improving the utilization efficiency of multi - sensor signals. At the same time, the present invention adopts a life prediction method of double - exponential particle filter, conducts state tracking modeling on historical samples through Bayesian theory, and updates the state transition function in real - time to realize tool wear degradation trend prediction and remaining life assessment. The proposed method can avoid the limitations of the deep - learning - based method that requires relying on a large amount of full - life data for offline training of the prediction model and the limited adaptability of the model to new prediction tasks. Brief Description of the Drawings

[0057] Figure 1 is the flowchart of the tool remaining life prediction method in the embodiment of the present invention;

[0058] Figure 2 It is a schematic diagram of particle filtering in an embodiment of the present invention;

[0059] Figure 3 It is the remaining life prediction curve of Tool No. 01 in an embodiment of the present invention.

[0060] Figure 4 It is the remaining life prediction curve of Tool No. 02 in an embodiment of the present invention.

[0061] Figure 5 It is the remaining life prediction curve of Tool No. 03 in an embodiment of the present invention.

[0062] Figure 6 It is the remaining life prediction curve of Tool No. 04 in an embodiment of the present invention. Specific embodiments

[0063] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0064] The present invention is a method for predicting the remaining life of a tool by multi-sensor information fusion. As Figure 1 shown, it specifically includes the following steps:

[0065] (1) Signal acquisition and processing: Collect the cutting force, vibration, and acoustic emission signals during the cutting process of the tool, and preprocess the collected signals.

[0066] (2) Signal feature extraction: Extract features from the preprocessed signals in three aspects: time domain, frequency domain, and time-frequency domain.

[0067] (3) Feature ranking and selection: Use the monotonicity evaluation index to calculate the Spearman rank correlation coefficient between the time vector and the feature vector as the feature monotonicity score, and select the features with a monotonicity score greater than 0.8 to construct a screening feature matrix.

[0068] (4) Feature fusion to construct a health index: Perform PCA-based feature fusion on the selected features, take the first principal component and perform triple first-order exponential smoothing on it to construct a health index representing tool wear degradation.

[0069] (5) Modeling and prediction: First, establish a double-exponential tool wear degradation model based on the constructed health index. Then, use the least squares method to fit the observed data and initialize the model parameters. Finally, perform state tracking modeling on the historical samples through Bayesian theory, update the state transfer function, and realize tool degradation trend prediction and remaining life assessment.

[0070] As Figure 2 shown is a schematic diagram of the particle filtering method in an embodiment of the present invention;

[0071] The specific implementation steps of particle filtering in an embodiment of the present invention are as follows:

[0072] (1) Initialize the particle set, t = 1:

[0073] For i = 1, 2, …, N, generate sampling particles from the prior distribution p(x0)

[0074] (2) For t = 2, 3, …, loop and execute the following steps:

[0075] 1) Importance sampling: For i = 1, 2, …, N, generate sampling particles from the importance probability density function Calculate the particle weights And normalize them;

[0076] 2) Resampling: Resample the particle set The resampled particle set is

[0077] 3) Output: Calculate the state estimation value at time k:

[0078] The technical solution of the present invention will be further described in detail below in conjunction with embodiments.

[0079] The data set used in the embodiments of the present invention is the public data of the PHM2010 data challenge. The data is collected from the real machining process of high-speed CNC machine tools. Each brand-new tool starts machining normally until the tool life ends and then the data collection stops.

[0080] (1) Signal acquisition and processing

[0081] 1) Collect the signals during the tool cutting process, namely: cutting force signals (in the X, Y, Z directions), vibration signals (in the X, Y, Z directions), and acoustic emission signals.

[0082] Use the signals of 4 tools for cutting, namely Tool No. 01, Tool No. 02, Tool No. 03, and Tool No. 04. During the cutting process, collect the signals of 7 channels including cutting force signals (three directions), acceleration signals (three directions), and acoustic emission signals at a sampling frequency of 50 kHz.

[0083] 2) Preprocess the collected signals. First, perform invalid data processing. The invalid data specifically includes invalid data during tool feeding and tool retracting, and it is processed by directly deleting. Then, perform outlier processing using the median filtering method based on a sliding window - hampel filtering.

[0084] Among them, in the method of direct deletion, the deletion point is located using the third quartile method. Specifically: First, calculate the third quartile Q3 of the data collected during the milling process as the critical value of the ineffective data for the feed-in and feed-out; then, starting from the first value of the data, compare the process data sequentially backward until the first value greater than Q3 appears, record the current position, and then truncate the data from the first value to this position; for the ineffective data of the feed-out, compare forward from the last value.

[0085] The specific process of median filtering is as follows:

[0086] 1) Set the number of samples k on both sides of the sample, then the window size is 2k + 1, and set the upper and lower bound coefficients n δ , where k = 4000 is set, and n δ = 3.

[0087] 2) Calculate the local standard deviation x δ and the local estimated median x m of each sample based on the sliding window;

[0088] 3) Calculate the upper and lower bounds of the outliers of the sample:

[0089] Upper bound of outliers: upbound = x m + n δ × x δ

[0090] Lower bound of outliers: downbound = x m - n δ × x δ

[0091] 4) If the sample value is greater than the upper bound of outliers or less than the lower bound of outliers, then use the estimated median x m to replace the sample.

[0092] (2) Signal feature extraction

[0093] From the three aspects of time domain, frequency domain, and time-frequency domain, divide the time series data of the 7 channels obtained in (1) into 315 data samples of equal length, and extract 15 features such as mean, standard deviation, skewness, kurtosis, impulse factor, crest factor, shape factor, margin factor, peak-to-peak value, root mean square, sk mean, sk standard deviation, sk skewness, sk kurtosis, and wavelet packet energy respectively, and construct a time series of 105 features (abbreviated as feature time series) to form a feature matrix X 315×105 . Let represent the i-th column feature time series of X 315×105 .

[0094] (3) Feature ranking and selection

[0095] Based on the monotonicity evaluation index, calculate the Spearman rank correlation coefficient between each characteristic time series obtained in (2) and the corresponding time vector as the monotonicity score of this characteristic, and select the characteristics with a monotonicity score greater than 0.8 to form the screening characteristic matrix X 315×S For The monotonicity score is calculated as shown in the following formula:

[0096]

[0097] In the formula, r k is an element in vector R,

[0098]

[0099] In the formula, is the sorting of

[0100] (4) Feature fusion to construct a health index

[0101] Fuse the features screened in (3) based on principal component analysis (PCA), take the first principal component and perform triple first-order exponential smoothing on it with a smoothing coefficient α = 0.1 to construct a health index. The feature fusion based on PCA is as follows

[0102] 1) Since there are great differences in the dimensions and magnitudes of the signal monitoring data of each channel, before fusing the screening characteristic matrix X 315×S using PCA, the z-score method is used to standardize X 315×S and denote it as

[0103] 2) Calculate the covariance matrix of

[0104] 3) Find the eigenvalues of

[0105] Assume the eigenvalues of the covariance matrix are λ i , then λ i represents the variance of the i-th principal component, and the eigenvector corresponding to λ i is the coefficient of the principal component with respect to the original variables.

[0106] 4) Take the eigenvector corresponding to the maximum variance to reconstruct , that is At this point, the original M S -dimensional features are reduced to one dimension.

[0107] 5) For x 315×1 = [x0, x1,..., x 315 ​​​Perform triple first-order exponential smoothing as shown in Equation (2) below, where α is the smoothing coefficient. y t are the smoothing values for the first, second, and third times respectively. The time series formed by the third smoothing value and the corresponding health index values of 315 data samples are Y {1:315} =[y1,…,y t ,…,y 315 ;

[0108]

[0109] (5) Modeling and prediction

[0110] Use a double-exponential model to fit the health index obtained in (4) and establish a double-exponential degradation model to describe the tool degradation process as follows:

[0111] y=ae bt +ce dt (3)

[0112] In the formula: y is the health index representing tool degradation constructed from the monitoring signal; a, b, c, d are model parameters that control the trend of the degradation trajectory; t is time.

[0113] According to the tool wear degradation model and combined with the particle filter algorithm and Bayesian theory, the tool remaining life prediction process is as follows:

[0114] 1) Establish the state equation and the observation equation

[0115] Use the Gaussian random walk model as the state equation to model the exponential degradation model parameters. The state equation and the observation equation are shown as follows:

[0116]

[0117] In the formula: ω a , ω b , ω c , ω d are the modeling errors of the state equation, which conform to the Gaussian distribution with a mean of 0 and variances of σ a , σ b , σ c , σ d respectively; then according to the double-exponential degradation model parameter modeling method, the observation equation is obtained as shown below:

[0118] HI(t)=a(t)·e b(t)·t +c(t)·e d(t)·t +v(t), v(t)~N(0, σ t ) (5)

[0119] where \(v(t)\) is the observation noise of the system, assumed to be Gaussian white noise, following a Gaussian distribution with a mean of 0 and a variance of \(\sigma\). t The tool wear condition at any time \(t\) can be represented by the observation function \(HI(t)\) of the tool wear degradation trend.

[0120] 2) Offline training: A total of 315 cuttings were completed for the 4 tools used in the experiment from the start of machining until the end of their service life. The original data collected from the first 150 cuttings was used as training samples for feature screening and constructing health indicators. The least - squares function was used to fit the observed data to roughly determine the parameter distribution range. At the same time, all parameters were initialized, and according to their initial distribution types, the probability distributions of each tool parameter were generated to obtain the initial distribution particles, where the number of particles was set to 1000.

[0121] 3) Online update: The particle filter algorithm was used to update the model parameters and health status in real - time according to the real - time monitoring data of the health indicators, and the parameter distribution of the double - exponential degradation model was adjusted according to the weight values.

[0122] 4) Life prediction: According to the real - time update results of the model parameters, the tool wear health indicators were calculated. By setting the failure threshold, the degradation trend of the tool wear health indicators for the 151 - 315th cuttings was predicted, and the remaining life of the tool was calculated.

[0123] The verification results of the full - life data of the No. 01 - 04 tools in the experiment are as Figures 3 to 6 shown. The figure gives the predicted values and the true values of the remaining life. As Figures 3 to 6 can be seen from it, in the first 200 cuttings, due to the small amount of observed data, the Bayesian particle filter algorithm failed to estimate the parameters of the double - exponential degradation model well, resulting in the predicted values of the remaining life not being able to follow the true values well. However, starting from the 200th cutting, as time progresses and the observed data continuously increases, the predicted values get closer and closer to the true values. For the No. 01 and No. 04 tools, in the later stage of cutting, the predicted values of the remaining life deviate from the true values somewhat, but they are always in an under - prediction state, that is, the available remaining cutting times of the tool are conservatively estimated. This situation will only lead to early tool change and will not affect the machining accuracy of the product, which is an ideal situation in the actual machining scenario.

Claims

1. A method for predicting the remaining tool life by multi-sensor information fusion, characterized in that The specific operation steps are as follows: Step 1: Collect signals during the cutting process of the tool. The signals include cutting force signals in three directions of X, Y, and Z, vibration signals in three directions of X, Y, and Z, and acoustic emission signals, a total of 7 channels of signals; Step 2: Preprocess the collected signals. Among them, directly delete the invalid data, and the invalid data includes invalid data during tool feed and invalid data during tool retraction; process the outliers using the median filtering method based on a sliding window; Step 3: For the signals preprocessed in Step 2, from the three aspects of time domain, frequency domain, and time-frequency domain, divide the time series data of the 7 channels obtained in Step 1 into 315 equally long data samples, and extract 15 features including mean, standard deviation, skewness, kurtosis, impulse factor, crest factor, shape factor, margin factor, peak-to-peak value, root mean square, sk mean, sk standard deviation, sk skewness, sk kurtosis, and wavelet packet energy respectively, to construct a time series of 105 features in total, forming a feature matrix , denoted by the th column feature time series; Step 4: Using the monotonicity evaluation index, calculate the Spearman rank correlation coefficient between each feature time series obtained in Step 3 and the corresponding time vector as the monotonicity score of this feature, and select the features with a monotonicity score greater than 0.8 to form a screened feature matrix ; Step 5: Fuse the features screened in Step 4 based on principal component analysis (PCA), take the first principal component and perform triple first-order exponential smoothing on it with a smoothing coefficient , construct a health index; fit the health index using a double-exponential model to establish a double-exponential degradation model describing the tool degradation process; Step 6: According to the double-exponential degradation model, combine the particle filter algorithm and Bayesian theory to predict the remaining tool life.

2. The method for predicting the remaining tool life by multi-sensor information fusion according to claim 1, characterized in that The specific process of the median filtering described in step 2 is as follows: 1) Set the number of samples on both sides of the sample to k, then the window size is 2k + 1, and set the upper and lower bound coefficients ; 2) Calculate the local standard deviation of each sample based on the sliding window , the local estimated median ; 3) Calculate the upper and lower bounds of the outlier of the sample: Upper bound of the outlier: Lower bound of the outlier: 4) If the sample value is greater than the upper bound of the outlier or less than the lower bound of the outlier, then use the estimated median to replace the sample.

3. The method for predicting the remaining tool life by multi-sensor information fusion according to claim 1, characterized in that The monotonicity evaluation index in Step 4 is specifically: For , the monotonicity score is calculated as follows: （1） wherein, is an element in vector R, represents a time unit; ; In the formula, is sorting of.

4. The method for predicting the remaining tool life by multi-sensor information fusion according to claim 1, characterized in that The feature fusion based on principal component analysis (PCA) in Step 5 is specifically as follows: 1) Before fusing the screened feature matrix using PCA perform standardization on it using the z-score method and denote it as ; 2) Calculate covariance matrix of ; 3) Calculate the covariance matrix eigenvalues; Assume the eigenvalues of the covariance matrix are , then represents the variance of the th principal component, and the corresponding eigenvector is the coefficient of the principal component with respect to the original variables; 4) Take the eigenvector corresponding to the maximum variance For Reconstruction, that is , thus far, the original dimensional features are reduced to one dimension; 5) For Perform triple first-order exponential smoothing as shown in the following formula (2), where is the smoothing coefficient, , , are the smoothing values for the first, second, and third times respectively. The time series formed by the third smoothing value and the corresponding health index values of 315 data samples are ; （2）。 5. The method for predicting the remaining tool life by multi-sensor information fusion according to claim 1, wherein, The double-exponential degradation model is as follows: （3） Wherein: is a health indicator characterizing tool degradation constructed from the monitoring signal; are the model parameters of the double exponential degradation model, controlling the trend of the degradation trajectory; is the time.

6. The method for predicting the remaining tool life by multi-sensor information fusion according to claim 5, wherein The process for predicting the remaining tool life is as follows: 1) Establish the state equation and the observation equation. Take the Gaussian random walk model as the state equation, as shown in the following formula (4): （4） Wherein: , , , are respectively the modeling errors of the state equation, conforming to the Gaussian distribution with a mean of 0 and variances of , , , ; then according to the modeling method of the double-exponential degradation model parameters, the observation equation is obtained as shown below: （5） In the formula, is the observation noise of the system, assumed to be Gaussian white noise, conforming to a Gaussian distribution with a mean of 0 and a variance of ; then the tool wear condition at any moment can be represented by the observation function of the tool wear degradation trend; 2) Offline training: Use the original data collected in the first 150 cuts as the training data, perform signal preprocessing, then perform feature extraction, feature screening, and construction of the health index. Substitute the constructed health index representing tool wear degradation into the double-exponential degradation model for model parameter initialization; use the least squares function to fit the observed data to roughly determine the parameter distribution range; at the same time, initialize all parameters, generate the probability distribution of each parameter according to its initial distribution type, and obtain the initial distribution particles; 3) Online update: Use the particle filter algorithm to update the double-exponential degradation model parameters and the health status evaluation results in real time according to the real-time monitoring data of the health index, and adjust the double-exponential degradation model parameter distribution according to the weight size; 4) Life prediction: Calculate the tool wear health index according to the real-time update results of the double-exponential degradation model parameters. By setting the failure threshold, predict the degradation trend of the tool wear health index for the 151st to 315th cuts, and calculate the remaining tool life.

7. The method for predicting the remaining tool life by multi-sensor information fusion according to claim 6, characterized in that, The specific implementation steps of the particle filter algorithm are: (1) Initialize the particle set, : For , generate from the prior distribution sampled particles ; (2) For , loop and execute the following steps: 1) Importance sampling: For , generate sampling particles from the importance probability density function , calculate the particle weights , and normalize them; 2) Resampling: Resample the particle set to obtain the resampled particle set ; 3) Output: Calculate the state estimate value at time: .

Images