A signal analysis-based machine tool cutting tool on-line monitoring system and a monitoring method thereof

Abstract

This invention discloses an online monitoring system and method for machine tool machining based on signal analysis. The system includes a vibration signal acquisition module, a signal transmission module, a data storage module, and a data processing and visualization module. The method includes: acquiring vibration signals at an 8kHz sampling rate using a piezoelectric accelerometer installed on the tool holder; transmitting the signals to a host computer via TCP / IP; constructing a tool life database; extracting 10 time-frequency domain feature parameters (time-domain statistics, dimensionless indices, envelope analysis, and frequency-domain analysis features) to construct a feature matrix; using PCA dimensionality reduction to extract the first two principal components (cumulative explained variance > 92.1%); and using a weighted KNN classifier (K=7, distance inverse weighting) to identify and classify five states: normal cutting, tool wear, chipping, built-up edge, and chatter fault. The results are displayed in real time and trigger threshold alarms via host computer software. This invention achieves 100% accuracy in real-time online monitoring of the five tool states, with advantages such as comprehensive feature extraction, efficient dimensionality reduction, accurate classification, and timely early warning.

Description

Technical Field

[0001] This invention belongs to the field of machining tools, and particularly relates to an online monitoring system and method for machine tool machining based on signal analysis. Background Technology

[0002] During CNC machining, when the cutting tool is cutting at high speed on the machine tool, the temperature in the workpiece contact area rises sharply with the machining time, leading to accelerated tool wear. Some wear phenomena cannot be observed or detected by the naked eye, and will result in poor workpiece accuracy and surface finish in the next machining operation due to tool wear or abnormalities. This can even lead to workpiece scrap, damage to expensive equipment, and machine tool downtime, directly impacting the accuracy, efficiency, and economic benefits of machining. Statistics show that approximately 75% of equipment failures in production are caused by tool failure, making online monitoring of machine tool cutting tools particularly important. With the improvement of intelligent production lines and the degree of production automation, especially with the emergence of flexible manufacturing system technology in China, online monitoring of the machining process will become increasingly important in the future. The importance of cutting tools to the machining process is self-evident; their condition largely determines product quality and the stability and reliability of the machine tool. Traditional tool wear and abnormal condition detection mainly relies on experienced machining personnel to estimate wear and abnormalities based on machining noise, chip color, cutting vibration, and tool cutting time. This method depends on human experience and is not conducive to widespread adoption. Therefore, online monitoring of machine tool cutting tools plays a crucial role in promoting the automation and unmanned operation of machining processes. Currently, tool wear detection methods can be mainly divided into two categories: direct detection methods and indirect detection methods. Direct detection methods usually require machine shutdown for inspection, making real-time online monitoring impossible to ensure timely monitoring. Indirect detection methods indirectly analyze tool wear status by monitoring physical signals related to tool wear (such as cutting force, vibration, current, acoustic emission, etc.), and can generally achieve real-time online monitoring, but have limitations such as complex sensor installation, high signal processing algorithm complexity, and difficulty in distinguishing multiple fault states. For example, CN114650471B discloses an online tool detection method beyond the field of view, which uses the machine tool control system to drive the machine tool's translation table to move a CCD camera above the tool to be inspected, focusing on the tool and acquiring images, thereby realizing the online monitoring function of the tool cutting edge during complex surface machining. However, this method mainly achieves timely modification and optimization of the machining path. Its vision-based detection method is essentially a static offline detection, which cannot monitor and warn of the real-time dynamic machining status of the tool. Furthermore, it does not involve the acquisition and analysis of vibration signals, fails to use piezoelectric accelerometers to collect vibration signals of the tool holder position in real time, and does not build an intelligent monitoring system based on time-frequency domain feature extraction, principal component analysis dimensionality reduction, and KNN classification. Therefore, it cannot achieve real-time and accurate identification of various states such as tool wear, tool chipping, built-up edge, and chatter faults.

[0003] For example, CN113267138A discloses an online tool inspection device and its detection method. This device detects the tool's motion signal via a photoelectric switch and transmits the signal to a control component. This causes a drive unit to move a moving block, which in turn moves a laser displacement sensor and the tool in a reciprocating linear motion, thus enabling real-time detection of tool deformation. However, this device is suitable for industrial robots but not for rotary cutting machining scenarios such as machining centers and machine tools. Furthermore, it cannot achieve online monitoring during the machining process. In addition, this solution does not use vibration signals as the monitoring object, does not utilize multi-channel dynamic acquisition at an 8kHz sampling rate to obtain vibration signals, does not transmit the signals to a host computer for data analysis via TCP / IP, and does not combine principal component analysis dimensionality reduction and weighted KNN classification algorithms to identify and classify the five types of tool states. Therefore, it has shortcomings in terms of applicability and real-time performance for monitoring the condition of machining tools. CN112621382B discloses an online tool wear monitoring method based on cutting vibration damping characteristics. Targeting the characteristics of milling machining, it uses the cutting vibration response signal and employs generalized zero-crossing and Hilbert transform to calculate the local mean frequency and instantaneous amplitude, ultimately calculating the local damping ratio of the cutting vibration to achieve tool wear monitoring. While this method utilizes vibration signals, it only monitors wear based on a single damping ratio feature, failing to integrate multi-domain feature information such as time-domain statistical features (root mean square value, peak value, peak-to-peak value, kurtosis, skewness), dimensionless index features (peak factor, waveform factor), envelope analysis features (root mean square value of the envelope), and frequency-domain analysis features (spectral centroid, spectral entropy) for comprehensive characterization. Furthermore, this method does not use principal component analysis to reduce the dimensionality of multi-dimensional features, nor does it utilize a weighted KNN classifier to identify and classify five states: normal cutting, tool wear, tool chipping, built-up edge, and chatter faults. Therefore, it has limitations in terms of simultaneous monitoring of multiple fault states and classification accuracy. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies and provide an online monitoring system and method for machine tool processing tools based on signal analysis. It collects vibration signals in real time through a piezoelectric accelerometer and utilizes the organic combination of multi-domain feature extraction, principal component analysis dimensionality reduction, and weighted KNN classification to achieve high-accuracy real-time online monitoring of five states of machine tool processing tools.

[0005] The technical solution of this invention: An online monitoring system for machine tool processing tools based on signal analysis includes: a vibration signal acquisition module, a signal transmission module, a data storage module, and a data processing and visualization module.

[0006] The vibration signal acquisition module collects vibration signals from the cutting process of the machine tool using a piezoelectric accelerometer. The piezoelectric accelerometer is installed at the tool holder to collect vibration signals transmitted near the tool tip. The module also includes a data acquisition card that supports multi-channel dynamic acquisition with a sampling rate of 8kHz and a signal acquisition duration of 1 second. The technical principle is as follows: During machine tool processing, friction is generated due to the relative motion between the tool and the workpiece, inevitably causing machine tool vibration. As tool wear increases, the cutting force increases, and the amplitude and frequency of the vibration also increase. Therefore, the vibration signal can characterize the degree of tool wear. Installing the piezoelectric accelerometer at the tool holder effectively captures vibration signals transmitted near the tool tip, and the application of vibration signals in tool wear detection has the advantages of simple measurement and low instrument cost. By performing multi-channel dynamic acquisition at a sampling rate of 8kHz and a signal acquisition time of 1 second, the vibration signal, including the spindle rotation frequency (40Hz), the tooth passage frequency (200Hz), and their harmonic components, can be completely captured, providing a high-quality data foundation for subsequent feature extraction. The technical benefits of this feature are: stable and high-fidelity acquisition of tool vibration signals during machining, avoiding the impact of traditional downtime detection methods on production efficiency, and ensuring the real-time nature and continuity of monitoring.

[0007] The signal transmission module uses TCP / IP to transmit the vibration signal to the host computer. The technical principle is as follows: the collected vibration signal has a high sampling frequency and a large number of sampling points, therefore a reliable data transmission method is required; TCP / IP transmission has advantages such as stable transmission, strong anti-interference capability, and long transmission distance, which can meet the reliable transmission requirements of the vibration signal from the machine tool to the host computer. The technical effect of this feature is: it ensures the integrity and real-time performance of the high-sampling-rate vibration signal during transmission, providing a reliable data link for subsequent data storage and online analysis.

[0008] The data storage module, deployed on the host computer, receives and saves the vibration signals, constructing a tool lifespan database. This database is used to extract characteristic information representing tool wear, thereby improving the accuracy of tool wear condition detection. The technical principle is as follows: by saving the collected tool vibration signals, a full-lifespan database covering the entire process of tool wear from initial wear to normal wear and then to rapid wear is constructed. This allows for the accumulation of characteristic patterns at different wear stages, providing rich data samples and reference benchmarks for subsequent feature extraction and classification. The establishment of the full-lifespan database ensures that monitoring results do not rely solely on signals at a single moment, but are based on the statistical regularities and trend analysis of historical data, thus improving the accuracy and reliability of detection. The technical effect of this feature is that through the accumulation of full-lifespan data, a comprehensive understanding of the tool wear evolution process is achieved, providing data support for intelligent diagnosis and significantly improving the accuracy of tool wear condition detection.

[0009] The data processing and visualization module is deployed on the host computer and includes host computer software for performing feature extraction, dimensionality reduction analysis, and classification recognition on the vibration signal. The host computer software also displays the diagnostic results and issues threshold alarms in real time. Specifically, the data processing and visualization module includes a feature extraction unit, a dimensionality reduction analysis unit, a classification recognition unit, and a status display and alarm unit.

[0010] The feature extraction unit extracts time-frequency domain feature parameters from the vibration signal, organizes the extracted time-frequency domain feature parameters into feature vectors, and constructs a feature matrix for principal component analysis dimensionality reduction based on the feature vectors. The time-frequency domain feature parameters include time-domain statistical features, dimensionless index features, envelope analysis features, and frequency-domain analysis features. The time-domain statistical features include root mean square value, peak value, peak-to-peak value, kurtosis, and skewness; the dimensionless index features include peak factor and waveform factor; the envelope analysis features include the root mean square value of the envelope; and the frequency-domain analysis features include the spectral centroid and spectral entropy. The underlying principle is that different tool states exhibit different statistical characteristics in the time domain, frequency domain, and envelope domain. Under normal cutting conditions, the vibration signal is a stable periodic signal, dominated by the rotational frequency and the tooth frequency. When the tool wears, higher harmonics are enhanced, frictional impacts occur, noise levels increase, and the root mean square value significantly increases. When the tool chips, a sudden energy impact occurs, with kurtosis and peak factor significantly higher than in other states. With built-up edge, low-frequency amplitude modulation occurs, and sideband components appear. During chatter faults, the self-excited vibration energy continuously increases over time, and the spectral entropy decreases, reflecting a high concentration of spectral energy. By extracting 10 feature parameters from four dimensions—time domain, dimensionless index, envelope, and frequency domain—five types of tool states can be comprehensively and multi-dimensionally characterized, forming a highly discriminative feature vector. The technical benefits of this feature are: through the fusion of multi-domain features, the hidden information of different tool states in the vibration signal is fully explored; the constructed feature matrix provides high-quality input for subsequent dimensionality reduction and classification, significantly improving the discriminativeness and accuracy of state recognition.

[0011] The dimensionality reduction analysis unit is configured to perform principal component analysis (PCA) on the feature matrix to extract the first two principal components, ensuring that the cumulative variance explained by these two principal components is greater than 92.1%. The underlying principle is that redundancy and correlation exist among the 10-dimensional features; directly inputting them into a classifier would lead to excessively high feature dimensionality, increased computational complexity, and even overfitting. PCA projects the original features onto a new coordinate system through orthogonal transformation, maximizing the variance along the first principal component (PC1) and second-largest along the second principal component (PC2), with no correlation between the principal components. PC1 primarily contributes from impact-sensitive features such as kurtosis, crest factor, and skewness, which are key indicators for distinguishing different tool failure types. By extracting the first two principal components, more than 92.1% of the original variance information can be explained, reducing the feature dimension from 10 to 2 dimensions while retaining most of the class discrimination information. The technical effects of this feature are: it significantly reduces data dimensionality and computational complexity, eliminates feature redundancy, avoids overfitting, and retains sufficient class discrimination information. This results in a clear clustering distribution of the five types of tool states after dimensionality reduction in two-dimensional space, with distinct boundaries between different categories, laying a good foundation for the construction of subsequent classifiers.

[0012] The classification and recognition unit is loaded with a pre-trained KNN classifier, used to identify and classify tool states based on dimensionality-reduced feature parameters. Tool states include normal cutting states as well as tool wear, tool chipping, built-up edge, and chatter fault states. The KNN classifier employs a weighted K-nearest neighbor algorithm with K=7, using the inverse distance as the weight for weighted voting. The KNN classifier is trained on a dataset consisting of five types of tool state samples, divided into training and test sets in a 7:3 ratio, with 50 samples collected for each class. Its technical principle is that in the two-dimensional feature space after PCA dimensionality reduction, sample points of the same type of tool state cluster together, while samples of different types form clear boundaries. For a sample to be identified, the weighted KNN algorithm finds the 7 nearest training samples in the feature space, using the inverse distance as the weight for weighted voting; the closer the sample, the greater its weight, thus assigning the sample to the class with the highest weight. Because the kurtosis and peak factor of edge breakage faults are significantly higher than other states, and the spectral entropy of flutter faults is the lowest, these key features are still preserved after dimensionality reduction using PCA. This results in clear clustering boundaries for the five states in two-dimensional space, allowing the KNN algorithm to accurately distinguish between them. A 7:3 ratio was used to divide the training set (175 groups) and the test set (75 groups), ensuring that samples of each class maintained the same proportion in both sets, guaranteeing good generalization ability of the classifier. The technical effects of this feature are: 100% classification accuracy on the test set; all diagonal elements in the confusion matrix are maximum values, and all off-diagonal elements are 0; precision, recall, and F1 score are all 1.000, demonstrating excellent model performance. Furthermore, all five sets of novel, unknown data not involved in training were correctly identified, validating the generalization ability and practical engineering application potential of this method.

[0013] The status display and alarm unit is used to visualize the identification and classification results in real time through the host computer software, and to issue an alarm when at least one of the time-frequency domain feature parameters exceeds the corresponding set threshold. Specifically, the root mean square (RMS) value corresponds to the tool wear threshold, the kurtosis and maxima factor correspond to the tool chipping threshold, and the spectral entropy corresponds to the chatter fault threshold. The technical principle is as follows: the RMS value characterizes the intensity of the vibration signal. As tool wear intensifies, the cutting force increases, and the vibration energy strengthens, resulting in a monotonically increasing RMS value. Therefore, tool wear early warning can be achieved by setting a wear threshold for the RMS value. The kurtosis and maxima factor are extremely sensitive to impact signals. Tool chipping generates a sudden energy impact, causing a sharp increase in kurtosis and maxima factor. Setting a chipping threshold can promptly detect chipping anomalies. The spectral entropy reflects the uniformity of the spectral energy distribution. During chatter faults, energy is concentrated at a specific frequency, and the spectral entropy decreases significantly. Setting a chatter threshold can achieve early warning of chatter faults. The host computer software displays the diagnostic results in real time, allowing operators to intuitively grasp the tool status. The technical benefits of this feature are: it enables real-time visual monitoring of tool status and multi-level threshold early warning. When any feature parameter exceeds the corresponding threshold, an alarm is triggered in time, reminding the operator to stop the machine for tool maintenance and replacement. This effectively avoids workpiece scrap, equipment damage, and machine tool downtime, significantly improving the safety and economic efficiency of machining.

[0014] A method for online monitoring of machine tool machining tools based on signal analysis includes the following steps: S1. Vibration signal acquisition: During the machine tool processing, vibration signals during the cutting process of the machine tool are acquired in real time through a piezoelectric accelerometer. S2, Signal Transmission: The vibration signal is transmitted to the host computer; S3. Data storage: The vibration signal is saved and a tool life database is constructed. The tool life database is used to extract feature information characterizing tool wear in order to improve the accuracy of tool wear condition detection. S4. Feature extraction: Extract time-frequency domain feature parameters from the vibration signal, organize the extracted time-frequency domain feature parameters into feature vectors, and construct a feature matrix for dimensionality reduction of principal component analysis based on the feature vectors; S5. Dimensionality Reduction Analysis: Perform principal component analysis to reduce the dimensionality of the feature matrix, extracting the first two principal components, such that the cumulative explained variance contribution rate of the first two principal components to the feature matrix is ​​greater than 92.1%. S6. Classification and Recognition: Using a pre-trained KNN classifier, the tool status is identified and classified based on the dimensionality-reduced feature parameters. The tool status includes normal cutting status as well as tool wear, tool chipping, built-up edge, and chatter fault status. S7. Status display and alarm: The identification and classification results are displayed in real time through the host computer software, and an alarm is triggered when at least one of the time-frequency domain feature parameters exceeds the corresponding set threshold.

[0015] The piezoelectric accelerometer is installed at the tool holder to collect vibration signals transmitted from near the tool tip. A data acquisition card is used for multi-channel dynamic acquisition at a sampling rate of 8kHz, with a signal acquisition duration of 1 second. Data transmission is performed via TCP / IP. The time-frequency domain feature parameters include time-domain statistical features, dimensionless index features, envelope analysis features, and frequency-domain analysis features. The time-domain statistical features include root mean square (RMS), peak value, peak-to-peak value, kurtosis, and skewness. The dimensionless index features include peak factor and waveform factor. The envelope analysis features include the RMS envelope value. The frequency-domain analysis features include the spectral centroid and spectral entropy. The RMS value corresponds to the tool wear threshold, the kurtosis and peak factor correspond to the tool chipping threshold, and the spectral entropy corresponds to the chatter fault threshold. The KNN classifier uses a weighted K-nearest neighbor algorithm with K = 7, weighted by the reciprocal of the distance. The KNN classifier is trained by dividing a dataset consisting of five types of tool state samples into a training set and a test set in a 7:3 ratio, with 50 samples collected for each state type. It also includes an unknown data verification step: extracting the time-frequency domain feature parameters from the real-time acquired vibration signals that did not participate in the KNN classifier training, and inputting them into the KNN classifier for identification and classification after dimensionality reduction by principal component analysis.

[0016] The beneficial effects of this invention are: By installing a piezoelectric accelerometer at the tool holder, multi-channel dynamic data acquisition is performed at an 8kHz sampling rate, and the data is transmitted to the host computer in real time via TCP / IP. This constructs a tool life database, enabling real-time online monitoring of tool status during machining. This transforms traditional periodic maintenance into real-time detection, effectively improving production efficiency, reducing downtime for maintenance, and preventing tool breakage due to severe wear, thus improving tool utilization efficiency.

[0017] (2) By extracting 10 time-frequency domain feature parameters, such as time-domain statistical features, dimensionless index features, envelope analysis features and frequency domain analysis features, the hidden information of different tool states in the time domain, frequency domain and envelope domain of vibration signals is fully explored from the technical principle. This enables comprehensive and high-precision characterization of five types of states: normal cutting, tool wear, tool chipping, built-up edge and chatter fault, which significantly improves the distinguishability of state recognition.

[0018] (3) By performing principal component analysis on the 10-dimensional feature matrix to reduce the dimensionality, the first two principal components were extracted to make the cumulative explained variance contribution rate greater than 92.1%. This eliminated feature redundancy from a technical perspective, reduced data dimensionality and computational complexity, avoided overfitting, and retained sufficient class discrimination information. This enabled the five types of tool states to form a clear cluster distribution in two-dimensional space, laying a good foundation for the subsequent classifier construction. Attached Figure Description

[0019] Figure 1 This is a structural diagram of the online monitoring system for machine tool processing based on signal analysis according to the present invention; Figure 2 A schematic diagram of time-domain vibration signals for five types of tool states; Figure 3 This is a schematic diagram of the feature mean matrix composed of 10 time-frequency domain feature parameters extracted from the vibration signal; Figure 4 PCA scatter plots of the five tool states after dimensionality reduction by principal component analysis; Figure 5 This is a schematic diagram of the real-time monitoring and alarm interface of the host computer software of the present invention; Figure 6 A schematic diagram of the confusion matrix for the KNN classifier to classify and identify five types of tool states. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] Example 1: Reference Figure 1 This embodiment provides an online monitoring system for machine tool processing tools based on signal analysis, including a vibration signal acquisition module 1, a signal transmission module 2, a data storage module 3, and a data processing and visualization module 4.

[0022] The vibration signal acquisition module 1 includes a piezoelectric accelerometer and a data acquisition card installed at the tool holder. During machine tool processing, the piezoelectric accelerometer acquires vibration signals transmitted near the tool tip in real time, and the data acquisition card performs multi-channel dynamic acquisition at a sampling rate of 8kHz, with a signal acquisition time of 1 second. In this embodiment, during rotary mechanical cutting, five typical tool states are acquired: normal cutting state, tool wear state, tool chipping state, built-up edge state, and chatter fault state. 50 samples are acquired for each state, for a total of 250 samples. The acquired vibration signals include the spindle rotation frequency (40Hz), the tooth passage frequency (200Hz), and their harmonic components.

[0023] Reference Figure 2 The time-domain signal characteristics of various tool states are as follows: Under normal cutting conditions, the vibration signal is a stable periodic signal, mainly composed of rotational frequency and tooth frequency; under tool wear conditions, higher harmonics are enhanced, frictional impact occurs, and noise level increases; under tool chipping conditions, sudden energy impact is generated, and kurtosis increases significantly; under built-up edge conditions, low-frequency amplitude modulation occurs, and sideband components appear; under chatter fault conditions, self-excited vibration energy continues to increase over time.

[0024] The signal transmission module 2 uses TCP / IP to transmit the collected vibration signals to the host computer.

[0025] The data storage module 3 is deployed on the host computer to receive and save vibration signals and build a tool life database to extract feature information characterizing tool wear, thereby improving the accuracy of tool wear state detection.

[0026] The data processing and visualization module 4 is deployed on a host computer and includes host computer software for feature extraction, dimensionality reduction analysis, and classification of vibration signals. The software also displays diagnostic results and triggers threshold alarms in real time. The data processing and visualization module 4 includes a feature extraction unit 41, a dimensionality reduction analysis unit 42, a classification and recognition unit 43, and a status display and alarm unit 44.

[0027] The feature extraction unit 41 extracts 10 time-frequency domain feature parameters for each vibration signal, organizes the extracted time-frequency domain feature parameters into feature vectors, and constructs a feature matrix for principal component analysis dimensionality reduction based on the feature vectors. The 10 time-frequency domain feature parameters include: time-domain statistical features (root mean square value, peak value, peak-to-peak value, kurtosis, skewness), dimensionless index features (peak factor, waveform factor), envelope analysis features (root mean square value of envelope), and frequency domain analysis features (spectral centroid, spectral entropy). (Refer to...) Figure 3 The color intensity of the feature mean matrix intuitively reflects the differences in various features of different tool states: the kurtosis and crest factor of chipping failure are significantly higher than those of other states; the spectral entropy of chatter failure is the lowest, reflecting the high concentration of spectral energy; and the root mean square value of tool wear state is significantly higher than that of normal cutting state.

[0028] The dimensionality reduction analysis unit 42 performs principal component analysis (PCA) to reduce the dimensionality of the 10-dimensional feature matrix. Since there is redundancy among the 10-dimensional features, directly inputting them into the classifier would lead to overfitting. By performing orthogonal transformation on the features through PCA, extracting the first two principal components can cumulatively explain 92.1% of the original variance information. (Refer to...) Figure 4The PCA scatter plot shows that after dimensionality reduction, the five tool states form a clear cluster distribution in two-dimensional space, with distinct boundaries between different categories. PC1 mainly contributes from impact-sensitive features such as kurtosis, peak factor, and skewness, indicating that these features are key indicators for distinguishing different tool failure types. This demonstrates that the 10 extracted features, after dimensionality reduction using PCA, retain sufficient class discriminative information, laying a solid foundation for subsequent classifier construction.

[0029] The classification and recognition unit 43 is loaded with a pre-trained KNN classifier. A weighted K-Nearest Neighbors (K=7) algorithm is used to construct the fault classifier. The core principle is: for a sample to be identified, the seven nearest training samples are found in the feature space, and weighted voting is performed using the inverse of the distance as the weight, assigning the sample to the category with the highest weight. The 250 sets of samples are randomly divided into a training set (175 sets) and a test set (75 sets) in a 7:3 ratio, ensuring that samples of each category maintain the same proportion in the training and test sets. The KNN classifier does not require explicit training, but all training samples must be saved as a reference benchmark. Figure 6 The test results show that the classification accuracy is 100%, and the diagonal elements in the confusion matrix are all maximum values, while the off-diagonal elements are all 0, indicating that all five types of tool states were accurately identified without any misclassification. The precision, recall, and F1 score are all 1.000, demonstrating excellent model performance.

[0030] To verify the model's practical diagnostic capability, five sets of novel vibration signals were collected—signals that were not used in training or testing and were completely independent of existing datasets. Ten identical feature values ​​were extracted from these unknown signals, and after PCA dimensionality reduction, they were input into a pre-trained KNN classifier for fault identification. The results showed that all five sets of unknown data were correctly identified, validating the generalization ability and practical engineering application potential of the proposed method.

[0031] The status display and alarm unit 44 displays the identification and classification results in real time via host computer software, and issues an alarm when the time-frequency domain feature parameters exceed the corresponding set thresholds. Specifically, the root mean square value corresponds to the tool wear threshold, kurtosis and peak factor correspond to the tool chipping threshold, and spectral entropy corresponds to the chatter fault threshold. (Refer to...) Figure 5 The host computer software interface displays the current tool status, characteristic parameter values, and historical trend curves in real time. When any characteristic parameter exceeds the corresponding threshold, the interface will issue an alarm by flashing red and making an audible prompt, reminding the operator to stop the machine in time for tool maintenance and replacement.

[0032] Example 2: This embodiment provides an online monitoring method for machine tool machining based on signal analysis. The specific steps are as follows: S1. Vibration signal acquisition: During the machine tool processing, the vibration signal of the cutting process of the machine tool is acquired in real time by a piezoelectric accelerometer installed at the tool holder. The data acquisition card is used to perform multi-channel dynamic acquisition at a sampling rate of 8kHz, and the signal acquisition time is 1 second.

[0033] S2. Signal transmission: The vibration signal is transmitted to the host computer using TCP / IP.

[0034] S3. Data storage: Save vibration signals and build a tool life database to extract characteristic information of tool wear, so as to improve the accuracy of tool wear condition detection.

[0035] S4. Feature Extraction: Extract 10 time-frequency domain feature parameters from the vibration signal, including time-domain statistical features (root mean square value, peak value, peak-to-peak value, kurtosis, skewness), dimensionless index features (peak factor, waveform factor), envelope analysis features (root mean square value of envelope), and frequency domain analysis features (spectral centroid, spectral entropy). Organize the extracted time-frequency domain feature parameters into feature vectors, and construct a feature matrix for dimensionality reduction of principal component analysis based on the feature vectors.

[0036] S5. Dimensionality Reduction Analysis: Perform principal component analysis on the feature matrix to reduce its dimensionality, extract the first two principal components, and ensure that the cumulative variance explained by the first two principal components is greater than 92.1% of the feature matrix.

[0037] S6. Classification and Recognition: A pre-trained weighted KNN classifier (K=7, weighted voting based on the inverse of distance) is used to classify and identify tool states based on dimensionality-reduced feature parameters. Tool states include normal cutting states as well as tool wear, tool chipping, built-up edge, and chatter fault states. The KNN classifier is trained on a dataset consisting of five types of tool state samples, divided into training and test sets in a 7:3 ratio, with 50 samples collected for each state. For real-time acquired vibration signals that were not involved in KNN classifier training, time-frequency domain feature parameters are extracted, and after dimensionality reduction using principal component analysis, they are input into the KNN classifier for identification and classification.

[0038] S7. Status Display and Alarm: The identification and classification results are displayed in real time via host computer software, and an alarm is triggered when at least one of the time-frequency domain feature parameters exceeds the corresponding set threshold. The root mean square value corresponds to the tool wear threshold, kurtosis and peak factor correspond to the tool chipping threshold, and spectral entropy corresponds to the chatter fault threshold.

Claims

1. An online monitoring system for machine tool machining tools based on signal analysis, characterized in that, include: The vibration signal acquisition module is used to acquire vibration signals during the cutting process of machine tool processing tools through a piezoelectric accelerometer. A signal transmission module is used to transmit the vibration signal to a host computer; a data storage module is deployed on the host computer to receive and save the vibration signal, construct a tool life database, which is used to extract feature information characterizing tool wear to improve the accuracy of tool wear state detection; a data processing and visualization module is deployed on the host computer, including host computer software, used to perform feature extraction, dimensionality reduction analysis, and classification recognition on the vibration signal, and to display the diagnostic results and trigger threshold alarms in real time through the host computer software. The data processing and visualization module includes: a feature extraction unit, used to extract time-frequency domain feature parameters from the vibration signal, organize the extracted time-frequency domain feature parameters into feature vectors, and construct a feature matrix for principal component analysis dimensionality reduction based on the feature vectors; a dimensionality reduction analysis unit, configured to perform principal component analysis dimensionality reduction on the feature matrix, extracting the first two principal components such that the cumulative explained variance contribution rate of the first two principal components to the feature matrix is ​​greater than 92.1%; a classification and recognition unit, loaded with a pre-trained KNN classifier, used to identify and classify the tool status based on the dimensionality-reduced feature parameters, the tool status including normal cutting status and tool wear, tool chipping, built-up edge, and chatter fault status; and a status display and alarm unit, used to visualize the identification and classification results in real time through the host computer software, and to issue an alarm when at least one of the time-frequency domain feature parameters exceeds a corresponding set threshold.

2. The online monitoring system for machine tool processing tools based on signal analysis according to claim 1, characterized in that, The piezoelectric accelerometer is installed at the tool holder to collect vibration signals transmitted from near the tool tip. The vibration signal acquisition module also includes a data acquisition card, which supports multi-channel dynamic acquisition, with a sampling rate of 8kHz and a signal acquisition duration of 1 second.

3. The online monitoring system for machine tool processing tools based on signal analysis according to claim 1, characterized in that, The signal transmission module uses TCP / IP for data transmission. The time-frequency domain characteristic parameters include time-domain statistical features, dimensionless index features, envelope analysis features, and frequency-domain analysis features. The time-domain statistical features include root mean square (RMS), peak value, peak-to-peak value, kurtosis, and skewness. The dimensionless index features include peak factor and waveform factor. The envelope analysis features include the root mean square (RMS) value of the envelope. The frequency-domain analysis features include the spectral centroid and spectral entropy. The RMS value corresponds to the tool wear threshold, the kurtosis and peak factor correspond to the tool chipping threshold, and the spectral entropy corresponds to the chatter fault threshold.

4. The online monitoring system for machine tool processing tools based on signal analysis according to claim 1, characterized in that, The KNN classifier uses a weighted K-nearest neighbor algorithm with a K value of 7, and performs weighted voting based on the reciprocal of the distance.

5. The online monitoring system for machine tool processing tools based on signal analysis according to any one of claims 1-4, characterized in that, The classification and recognition unit is loaded with a pre-trained KNN classifier. The KNN classifier is trained based on a dataset consisting of five types of tool state samples, which are divided into a training set and a test set in a 7:3 ratio, with 50 samples collected for each state type. The classification and recognition unit is also used to identify and classify real-time acquired vibration signals that were not involved in the training.

6. A method for online monitoring of machine tool machining tools based on signal analysis, characterized in that, Includes the following steps: S1. Vibration Signal Acquisition: During machine tool processing, vibration signals from the cutting process of the machining tool are acquired in real time using a piezoelectric accelerometer. S2. Signal Transmission: The vibration signals are transmitted to a host computer. S3. Data Storage: The vibration signals are saved, and a tool life database is constructed. This database is used to extract feature information characterizing tool wear to improve the accuracy of tool wear state detection. S4. Feature Extraction: Time-frequency domain feature parameters are extracted from the vibration signals. The extracted time-frequency domain feature parameters are organized into feature vectors, and a feature matrix for dimensionality reduction using principal component analysis is constructed based on these feature vectors. S5. Dimensionality Reduction Analysis: Principal component analysis is performed on the feature matrix to reduce its dimensionality. The first two principal components are extracted, ensuring that the cumulative explained variance contribution rate of the first two principal components to the feature matrix is ​​greater than 92.1%. S6. Classification and Recognition: A pre-trained KNN classifier is used to classify and identify the tool state based on the dimensionality-reduced feature parameters. The tool state includes normal cutting state, tool wear, tool chipping, built-up edge, and chatter fault states. S7. Status display and alarm: The identification and classification results are displayed in real time through the host computer software, and an alarm is triggered when at least one of the time-frequency domain feature parameters exceeds the corresponding set threshold.

7. The online monitoring method for machine tool processing tools based on signal analysis according to claim 6, characterized in that, The piezoelectric accelerometer is installed at the tool holder to collect vibration signals transmitted from near the tool tip. In the vibration signal acquisition step, a data acquisition card is used to perform multi-channel dynamic acquisition at a sampling rate of 8kHz, and the signal acquisition time is 1 second.

8. The online monitoring method for machine tool processing tools based on signal analysis according to claim 6, characterized in that, In the signal transmission step, data transmission is performed using TCP / IP. The time-frequency domain characteristic parameters include time-domain statistical features, dimensionless index features, envelope analysis features, and frequency-domain analysis features. The time-domain statistical features include root mean square (RMS) value, peak value, peak-to-peak value, kurtosis, and skewness. The dimensionless index features include peak factor and waveform factor. The envelope analysis features include the root mean square (RMS) value of the envelope. The frequency-domain analysis features include the spectral centroid and spectral entropy. The RMS value corresponds to the tool wear threshold, the kurtosis and peak factor correspond to the tool chipping threshold, and the spectral entropy corresponds to the chatter fault threshold.

9. The online monitoring method for machine tool processing tools based on signal analysis according to claim 6, characterized in that, The KNN classifier uses a weighted K-nearest neighbor algorithm with a K value of 7, and performs weighted voting based on the reciprocal of the distance.

10. The online monitoring method for machine tool processing tools based on signal analysis according to any one of claims 6-9, characterized in that, The KNN classifier is trained by dividing a dataset consisting of five types of tool state samples into a training set and a test set in a 7:3 ratio, with 50 samples collected for each state type. It also includes an unknown data verification step: extracting the time-frequency domain feature parameters from the real-time acquired vibration signals that did not participate in the KNN classifier training, and inputting them into the KNN classifier for identification and classification after dimensionality reduction by principal component analysis.

Images