A tool life real-time monitoring and early warning method and system

Abstract

The present application relates to the technical field of machining tools, and proposes a tool life real-time monitoring and early warning method and system. First, the original acoustic emission signal of the tool working site is obtained, adaptive cavitation noise cancellation processing is performed, and comparison index calculation is performed based on the processed signal, including entropy change and kurtosis calculation, frequency dispersion-attenuation coupling analysis is performed in combination with short-time Fourier transform, and frequency dispersion slope change rate and energy attenuation ratio are extracted. Subsequently, the current tool individualized threshold vector identified by the deep learning model is obtained from the cloud, compared with the comparison index, the life remaining proportion is determined, and early warning is implemented. Through millisecond acoustic emission signal analysis, microscopic secondary detection and cloud model self-learning, the present application realizes millisecond response, extremely low false alarm rate and seamless system integration under the premise of no shutdown and unattended, and breaks through the traditional fixed tool changing period, and approaches the real service life limit of the tool.

Description

Technical Field

[0001] This invention relates to the field of machining cutting tools, specifically to a method and system for real-time monitoring and early warning of tool life. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Real-time tool life monitoring is a core component in intelligent manufacturing, ensuring product quality and equipment efficiency. With the widespread application of high-speed cutting, five-axis linkage, and automatic tool changing in industries such as aerospace, automotive, and energy, the evolution of tool wear and sudden chipping has been shortened to the millisecond level, posing multiple challenges to monitoring systems, including rapid response, accurate identification, and system integration. Traditional strategies such as "timed tool changing" or "manual inspection" are not only inadequate for complex machining conditions but also lead to the premature scrapping of a large number of tools, resulting in significant resource waste. Therefore, developing real-time monitoring technologies with high response speed, low false alarm rate, and the ability to coordinate with production line systems has become crucial for improving tool utilization and manufacturing efficiency.

[0004] Currently, various monitoring methods exist, but they generally suffer from three major problems: response delay, high false alarm rate, and system isolation. First, the response delay primarily stems from the system's reliance on offline modeling or low-frequency sampling signals. For example, some solutions use current or torque signals combined with wavelet decomposition and neural networks for wear prediction, requiring servo system control parameters, resulting in long response paths. Updating the model necessitates downtime for collecting a large number of samples, making millisecond-level responses difficult. Vision and infrared detection solutions suffer from air path lag or image processing delays, also struggling to cope with sudden chipping in high-speed cutting scenarios. Second, the high false alarm rate arises from the overly simplistic monitoring signals and poor anti-interference capabilities. Taking acoustic emission signals as an example, in high-speed milling, they are affected by coolant cavitation and electromagnetic interference, resulting in a signal-to-noise ratio below 12dB and an effective signal loss rate exceeding 45%. Image recognition solutions are susceptible to interference from chip occlusion and thermal drift, leading to drastic fluctuations in recognition accuracy, poor model generalization ability, and actual performance far below laboratory results. Furthermore, traditional timed tool changing strategies often result in tools being scrapped before reaching their lifespan limit, wasting resources and increasing production costs. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a real-time tool life monitoring and early warning method and system. By leveraging millisecond-level acoustic emission signal analysis, secondary microscopic detection, and cloud-based model self-learning, it achieves millisecond-level response, low false alarm rate, and seamless system integration without downtime or unattended operation. Simultaneously, it optimizes the traditional fixed tool change cycle, bringing the tool life closer to its actual service life, thus providing a solution for tool cutting force control and tool life monitoring.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The first aspect of this invention provides a method for real-time monitoring and early warning of tool life, comprising the following steps:

[0008] The original acoustic emission signal at the tool working site is acquired and preprocessed, and the preprocessed signal is subjected to adaptive cavitation noise cancellation processing.

[0009] Comparison metrics were calculated, and bandpass filtering and kurtosis statistical analysis were performed on the acoustic emission signal after adaptive cavitation noise cancellation processing to calculate entropy change. with kurtosis Combining short-time Fourier transform with dispersion-attenuation coupling analysis, the rate of change of the slope of the modal dispersion curve is calculated. And the energy decay ratio ΔE / Δf;

[0010] The personalized threshold vector of the current tool, which is learned and identified based on a deep learning model, is obtained and compared with various calculated comparison indicators to determine the remaining lifespan ratio and issue an early warning.

[0011] The deep learning model acquires the machining data of the current tool in real time, identifies the individual feature offset of the tool relative to the standard tool, and adds the set general benchmark threshold to the predicted individual feature offset as the personalized threshold vector of the current tool.

[0012] A second aspect of the present invention provides a real-time tool life monitoring and early warning system, comprising:

[0013] An acoustic emission sensor and a control terminal, wherein the control terminal is configured to perform the steps of the above-described method for real-time monitoring and early warning of tool life.

[0014] A third aspect of the present invention provides a real-time tool life monitoring and early warning system, comprising:

[0015] The noise processing module is configured to acquire the original acoustic emission signal at the tool working site, preprocess it, and perform adaptive cavitation noise cancellation processing on the preprocessed signal.

[0016] The comparison index calculation module is configured to calculate comparison indices. Based on the acoustic emission signal after adaptive cavitation noise cancellation processing, it performs bandpass filtering and kurtosis statistical analysis to calculate entropy change. with kurtosis Combining short-time Fourier transform with dispersion-attenuation coupling analysis, the rate of change of the slope of the modal dispersion curve is calculated. And the energy decay ratio ΔE / Δf;

[0017] The threshold update and early warning module is configured to obtain the personalized threshold vector of the current tool after learning and recognition based on the deep learning model, compare it with the calculated comparison indicators, determine the remaining life ratio and issue an early warning.

[0018] The deep learning model acquires the machining data of the current tool in real time, identifies the individual feature offset of the tool relative to the standard tool, and adds the set general benchmark threshold to the predicted individual feature offset as the personalized threshold vector of the current tool.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] The monitoring and early warning method of this invention can significantly reduce the response latency of traditional monitoring methods in high-speed machining scenarios, shorten the entire process time of signal acquisition, feature calculation, and threshold comparison, and meet the millisecond-level real-time requirements. Adaptive cavitation noise cancellation improves the signal-to-noise ratio, reduces the signal loss rate, and effectively enhances the system's anti-interference capability. The combination of dispersion-attenuation coupling characteristics and a dual-threshold discrimination algorithm reduces the false alarm rate of initial wear. By identifying individual tool offsets through a deep learning model and dynamically correcting the comparison threshold, it has the ability to adapt to individual tools, effectively suppressing false alarms and missed alarms caused by manufacturing tolerances and different working conditions, and greatly improving monitoring accuracy and system adaptability.

[0021] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0023] Figure 1 This is a flowchart of a real-time tool life monitoring and early warning method according to Embodiment 1 of the present invention;

[0024] Figure 2 This is a flowchart of the deep learning model construction and threshold recognition in Embodiment 1 of the present invention;

[0025] Figure 3 This is a schematic flowchart of the acoustic emission signal preprocessing in Embodiment 1 of the present invention;

[0026] Figure 4 This is an interface diagram of the real-time tool life monitoring and early warning system of Embodiment 3 of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0029] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0030] Example 1

[0031] In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 4 As shown, a method for real-time monitoring and early warning of tool life includes the following steps:

[0032] Step 1: Obtain the original acoustic emission signal from the tool's working environment and preprocess it. Then, perform adaptive cavitation noise cancellation on the preprocessed signal.

[0033] Step 2: Calculate the comparison index. Based on the acoustic emission signal after adaptive cavitation noise cancellation processing, perform bandpass filtering and kurtosis statistical analysis, and calculate the entropy change. with kurtosis Combining short-time Fourier transform (STFT) with dispersion-attenuation coupling analysis, the rate of change of the slope of the modal dispersion curve is calculated. And the energy decay ratio ΔE / Δf;

[0034] Step 3: Obtain the personalized threshold vector of the current tool after learning and recognition based on the deep learning model, compare it with the calculated comparison indicators, determine the remaining life ratio and issue an early warning.

[0035] The deep learning model acquires the machining data of the current tool in real time, identifies the individual feature offset of the tool relative to the standard tool, and adds the set general benchmark threshold to the predicted individual feature offset as the personalized threshold vector of the current tool.

[0036] Optionally, the deep learning model can be deployed in the cloud or on an edge device; preferably, it can be deployed in the cloud.

[0037] In this embodiment, an adaptive cavitation noise cancellation unit is integrated into the acoustic emission signal acquisition link to suppress coolant cavitation noise and effectively improve the signal-to-noise ratio. Wear-related entropy changes are extracted based on bandpass filtering and kurtosis analysis. with kurtosis The metrics, combined with Short-Time Fourier Transform (STFT) for dispersion-attenuation coupling analysis, extract the slope of the dispersion curve and the energy attenuation ratio ΔE / Δf of the Lamb wave mode to dynamically characterize the tool wear evolution process. Subsequently, a cloud-based deep learning model continuously receives the machining history data of each tool, identifies the individual feature offset of each tool relative to the standard model, and generates personalized threshold vectors based on a general benchmark threshold, enabling accurate comparison and real-time early warning of the life status of each tool.

[0038] This implementation significantly reduces the response latency of traditional monitoring methods in high-speed machining scenarios, achieving a total time of less than 5ms for signal acquisition, feature calculation, and threshold comparison, meeting millisecond-level real-time requirements. The adaptive cavitation noise cancellation module improves the signal-to-noise ratio by more than 18dB, reducing signal loss rate and effectively enhancing interference resistance. The combination of dispersion-attenuation coupling characteristics and a dual-threshold discrimination algorithm reduces the false alarm rate of initial wear. By identifying individual tool offsets through a deep learning model and dynamically correcting the comparison threshold, the system possesses individual adaptability, effectively suppressing false alarms and missed alarms caused by manufacturing tolerances and different working conditions, greatly improving monitoring accuracy and system adaptability.

[0039] Further technical solutions also include step 4: when the remaining lifespan ratio is lower than the set value, a tool replacement operation is performed, the replaced tool is microscopically identified to determine whether the tool can be reused, and the comparison index calculated by the tool is used as the input value, and the result of microscopic identification is used as the output to construct new training data for rolling training of the deep learning model in the cloud.

[0040] The current tool processing data includes: comparison indicators calculated from acoustic emission signals collected during tool operation, identification results of whether the replaced tool is reusable obtained by microscopic identification, and the remaining life ratio of reusable tools.

[0041] Optionally, the training interval can be set, such as one day, two days or longer, to construct a new training set from the tool operation data within the corresponding time period, perform real-time training, and update the personalized threshold vector of the current tool after training.

[0042] In step 1, acoustic emission signals can be acquired using an acoustic emission sensor; the acoustic emission sensor is installed within ±15° of the machine tool spindle axis and 60±5mm away from the cutting edge;

[0043] In step 1, such as Figure 3The diagram shows the preprocessing steps for acoustic emission signals, including the following:

[0044] Step 11: Acquire the original acoustic emission signal and perform two-stage processing: charge amplification and voltage amplification;

[0045] Step 12: Extract the effective cutting signal segment of the tool based on the root mean square value and signal-to-noise ratio of the window signal;

[0046] Specifically, acoustic emission signals are captured in real time by setting a sliding window of size such as 100 milliseconds. The window slides once every 10 milliseconds, and the root mean square value (RMS) and signal-to-noise ratio (SNR) of each window are calculated. The RMS is used to determine whether the signal energy has reached the characteristic threshold of the cutting state, and the machine tool spindle speed and feed rate are used for verification. Low-quality signal segments are then filtered out by the SNR to ensure that the extracted data reflects the cutting state and has sufficient signal quality.

[0047] Step 13: Match the acquired acoustic emission signal with the corresponding tool ID to determine the acoustic emission signal of the current tool;

[0048] Specifically, during initialization or each tool change operation, a unique tool ID (such as T100, T101, etc.) is assigned to each physical tool, and this ID, along with metadata such as the tool's model, geometric parameters, and material information, is entered into the database to complete the registration and binding of the tool's identity. During actual machining, real-time data interaction is performed with the CNC machine tool control system through the OPC UA interface to dynamically read the active tool number (ID) on the current spindle.

[0049] Step 11 is implemented using amplifiers, including a charge amplifier, a voltage amplifier, and an output amplifier; optionally, the first-stage charge amplifier has a gain of 40dB, the second-stage voltage amplifier has a gain of 20dB, and the total gain is 60dB, to avoid oscillation caused by excessive single-stage gain; the output uses a high-speed video amplifier, such as the AD8397.

[0050] In the above embodiments, by introducing a two-stage front end for charge amplification and voltage amplification, the quality of the acoustic emission signal can be effectively enhanced, allowing weak tool wear signals to be clearly presented, thereby improving the overall sensitivity and accuracy of the system. Background noise removal and signal segment extraction processes reduce the system's false recognition rate, ensuring that the extracted data is highly targeted and effective. The main frequency calculation and tool signal mapping mechanism ensure accurate positioning of the monitored object, providing support for multi-tool parallel monitoring and improving the overall monitoring accuracy and reliability of the system.

[0051] In step 1, the preprocessed acoustic emission signal is subjected to adaptive cavitation noise cancellation processing through the constructed adaptive cavitation noise cancellation unit; taking the cavitation background noise collected by the reference microphone as input, the cancellation signal with phase reversal characteristics is generated through adaptive filtering methods such as feature basis construction and Hilbert transform, so as to suppress the cavitation component in the original acoustic emission signal and improve the signal-to-noise ratio.

[0052] Specifically, the method for adaptive cavitation noise cancellation processing of the preprocessed signal includes the following steps:

[0053] Step 101: Acquire the preprocessed acoustic emission signal and reference signal r(n), perform synchronous acquisition and frequency band filtering to obtain time domain data with phase alignment and frequency band limited to the set frequency range;

[0054] Optionally, the frequency range for the frequency band limitation is 220-320kHz;

[0055] Specifically, the reference signal r(n) represents the cavitation noise signal collected by the reference microphone; the reference microphone is placed near the coolant pipeline or outside the machine tool housing to detect the background noise generated by coolant cavitation.

[0056] It is feasible to use a dual-channel synchronous acquisition card to perform hardware-triggered aligned synchronous sampling of the acoustic emission signal and the reference signal r(n), ensuring that the phase error between the two is less than 0.5µs; for example, a dual-channel synchronous acquisition card with a sampling rate of 1MHz and a resolution of 16bit can be used.

[0057] Step 102, Feature basis construction: Perform spectrum analysis and spectral kurtosis discrimination on the reference signal, extract a set number of main frequency components, and construct an orthogonal feature basis vector set to achieve dynamic tracking of the cavitation-dominant frequency band;

[0058] The process of constructing an orthogonal feature basis vector set is as follows:

[0059] Step 1021: Perform a Fast Fourier Transform on the reference signal r(n);

[0060] Specifically, perform a 512-point Fast Fourier Transform (FFT).

[0061] Step 1022: Calculate spectral kurtosis to locate the frequency band where cavitation noise energy is most concentrated;

[0062] Step 1023: Extract the maximum number of spectral lines with the highest energy within the most concentrated frequency band with a set step size u1;

[0063] Specifically, the five spectral lines with the highest energy are extracted with a step size of 4 kHz u1;

[0064] Step 1024: The frequency components corresponding to the extracted spectral lines are subjected to Gram-Schmidt orthogonalization to generate a set of feature basis vectors;

[0065] Specifically, when five spectral lines are extracted, the frequency components corresponding to the extracted spectral lines are orthogonalized by Gram-Schmidt to generate a five-dimensional feature basis vector set φ=[φ1,φ2,...,φ5], which can be updated every 0.512ms.

[0066] Step 103: Perform characteristic basis projection and Hilbert transform on the reference signal to generate an inverse time-domain signal with the opposite phase to the original cavitation noise. Specifically:

[0067] Step 1031: Project the reference signal r(n) onto the characteristic basis φ to obtain the projection coefficients c(m);

[0068] Step 1032: For each basis vector φ of the characteristic basis φ m Perform the Hilbert transform to obtain the corresponding complex analytic form;

[0069] Step 1033: Based on the projection coefficients c(m), perform a weighted combination of the complex analytic forms of the Hilbert transform of the corresponding basis vectors to obtain the cavitation noise complex envelope signal. The instantaneous amplitude A(n) and instantaneous phase θ(n) of the cavitation noise are obtained.

[0070] Step 1034: Construct the inverse signal of the cavitation noise complex envelope signal. And generate an inverse cancellation signal through inverse mapping of characteristic basis. That is, the cavitation noise inverted time-domain signal;

[0071] Step 104: Subtract the cavitation noise inverse time domain signal from the synchronously acquired preprocessed acoustic emission signal to obtain the acoustic emission signal after adaptive cavitation noise cancellation processing;

[0072] The above steps overcome the limitations of traditional filters that only perform linear superposition, achieving waveform-level reconstruction and inversion of non-stationary cavitation noise. By constructing a dynamic cavitation frequency tracking and feature back-projection mechanism, adaptive cancellation of coolant cavitation noise is achieved. Compared with traditional static filters or fixed reference signal cancellation methods, this method has a high adaptability to frequency drift and amplitude fluctuations, improving the signal-to-noise ratio by more than 18dB in the critical noise range of 200–400kHz, effectively reducing the signal loss rate, improving the system's anti-interference capability and wear feature extraction accuracy, and providing a more stable and reliable signal foundation for subsequent lifespan assessment.

[0073] In step 2, bandpass filtering and feature statistical analysis are performed on the acoustic emission signal after adaptive cavitation noise cancellation processing to calculate the entropy change. with kurtosis The process includes the following steps:

[0074] Step 21: Apply a bandpass filter to the acoustic emission signal after adaptive cavitation noise cancellation processing; a bandpass filter of 50–400 kHz can be used.

[0075] Step 22: Set up a sliding window and calculate the kurtosis within each sliding window. To measure the sharpness of the signal; to calculate the entropy change. To measure signal complexity, higher entropy indicates more dispersed energy;

[0076] cliff The calculation formula is:

[0077] ;

[0078] in, This represents the number of sampling points within the sliding window. Let be the signal amplitude at the i-th sampling point within the window. This represents the mean of the signal within the window. The standard deviation of the signal within the window;

[0079] cliff A significantly increased value (e.g., >5.0) indicates a large number of sudden impacts in the signal, typically corresponding to rapid damage such as chipping or microcracks in the tool. Kurtosis A significant increase in value can be achieved by setting a threshold, such as setting a threshold of not less than 5.0. When the value is greater than the set threshold, it is judged as a significant increase.

[0080] Based on the complexity and disorder of signal energy quantized by wavelet packet entropy, entropy change... The calculation formula is:

[0081] ;

[0082] ;

[0083] Among them, the wavelet packet energy probability distribution This represents the proportion of energy of the k-th wavelet packet node (i.e., frequency band) to the total energy of the j-th layer decomposition; wavelet packet entropy. Represents the Shannon entropy calculated at the j-th level decomposition; the baseline entropy. This represents the wavelet packet entropy of the j-th layer calculated under a baseline or normal state.

[0084] A sustained increase indicates that the signal energy has shifted from concentrated to dispersed, reflecting progressive tool wear. When When the wavelet packet entropy threshold is exceeded, it indicates that the tool has entered the abnormal wear stage. The wavelet packet entropy threshold can be set to be greater than 40%. In this embodiment, 40% can be used as the general benchmark threshold for the wavelet packet entropy threshold.

[0085] Optionally, if the sliding window is set to 100ms, a set will be obtained every 100ms. It is used to identify changes in cutting conditions or abnormal vibration events.

[0086] In step 2, based on the acoustic emission signal after adaptive cavitation noise cancellation processing, dispersion-attenuation coupling analysis is performed using short-time Fourier transform (STFT) to calculate the slope change rate of the modal dispersion curve. and wavelet packet energy attenuation ratio The process includes the following steps:

[0087] Step 201: Perform a short-time Fourier transform (STFT) on the acoustic emission signal to obtain the frequency-time distribution spectrum curve;

[0088] Specifically, the key analysis parameters for the Short Time Fourier Transform (STFT) are set as follows: Hanning window (256 points in length) and 75% overlap. This configuration provides a frequency resolution of approximately 1.95 kHz at a sampling rate of 2 MHz, which is sufficient to capture subtle spectral structure changes caused by tool wear.

[0089] Step 202: Extract the S0 / A0 mode dispersion curve of the Lamb wave from the spectrum curve, and calculate the rate of change of the slope of the mode dispersion curve. :

[0090] ;

[0091] in, This represents the slope of the dispersion curve of the Lamb wave; Indicates the time interval; S0 is the symmetric mode, and A0 is the antisymmetric mode;

[0092] Rate of change of slope Characterizing the drastic evolution of the dispersion properties of stress wave propagation over time, when k disp A value >0.18ms / kHz indicates that the tool has entered the abnormal wear stage.

[0093] Step 203: Calculate the wavelet packet energy attenuation ratio within the set frequency band. The frequency band can be set to 300–400kHz.

[0094] ;

[0095] in, Indicates the energy difference within a specified frequency band; Indicates the frequency difference of a set frequency band;

[0096] Rate of change of slope of modal dispersion curve Used to represent changes in modal propagation characteristics, wavelet packet energy attenuation rate ratio This is used to quantize the linear decay rate of signal energy as frequency increases.

[0097] Step 3, the method for generating the personalized threshold vector of the current tool, includes the following steps:

[0098] Step 31: Based on the current tool's model, coating, and material hardness information, obtain the current tool's baseline threshold vector from the cloud. ;

[0099] ;

[0100] in, Indicates the kurtosis baseline threshold. Indicates the entropy change baseline threshold. The threshold value representing the rate of change of the slope of the modal dispersion curve. This indicates that the wavelet packet energy attenuation rate is compared to the baseline threshold.

[0101] Optionally, a baseline threshold vector is set for each type of tool, a baseline table is constructed, and the MySQL baseline table is stored in the cloud. Alternatively, the baseline threshold vector can be established offline. For each type of tool, the acoustic emission signals of the entire process from new tool to complete wear under standard working conditions are collected. The moment when the wear of the back face of each tool reaches the critical value (VB=0.15 mm) is determined by offline microscopy measurement. The acoustic emission feature value corresponding to the critical point is extracted, and the feature values ​​of all samples of the same model are statistically analyzed (e.g., average value is taken) to obtain the baseline threshold vector of the tool model.

[0102] A specific setup example, ;

[0103] Step 32: Obtain all comparison indices of acoustic emission signals of the current tool before the current time, calculate the statistics of each comparison index, input them into the pre-trained deep learning model for recognition, and obtain the individual feature offset.

[0104] Specifically, cloud-based deep learning models can employ lightweight CNN models; individual feature offset, i.e., the individual feature offset of an individual tool relative to a standard tool, is expressed as:

[0105] ;

[0106] in, Indicates kurtosis offset; Indicates entropy shift; Indicates the shift in dispersion slope; Represents the energy decay rate offset; offset vector Each component in the equation can be positive or negative. A positive offset indicates that the tool has higher tolerance in the corresponding feature dimension compared to a standard tool, requiring a higher alarm threshold to fully utilize its service life. A negative offset indicates that the tool has poor tolerance, requiring a lower alarm threshold to ensure machining safety.

[0107] Step 33: Add the baseline threshold to the predicted individual feature offset to generate a personalized threshold vector suitable for the current tool. :

[0108] ;

[0109] The obtained personalized threshold vector The data is sent to the tooling work site and compared with the calculated comparison indicators. Based on the comparison results, an early warning is issued.

[0110] The above implementation method can fully exploit the characteristic offset information generated by individual tools in actual applications, overcoming the incompatibility of traditional fixed threshold methods with manufacturing tolerances, coating differences, and material fluctuations, and realizing tool-level differentiated judgment logic. By dynamically generating personalized threshold vectors, the system can accurately identify the life stage based on actual wear evolution characteristics, significantly reducing false alarms and missed alarms caused by inaccurate thresholds, improving the intelligence and adaptability of the overall monitoring system, and ensuring maximum utilization of tool resources and high-reliability machining.

[0111] To achieve intelligent prediction of the offset trend of individual tool features, a supervised learning-based model for predicting individual feature offsets was built and trained in the cloud. This model can predict the multidimensional threshold offset vector of a new tool relative to a standard tool based on its acoustic emission characteristics during the initial machining stage, supporting subsequent adaptive threshold setting and personalized lifetime modeling.

[0112] Furthermore, step 32 also includes the process of training the deep learning model in the cloud, including the following steps:

[0113] Step 321: Obtain acoustic emission signals from the tool's historical operation and construct a training dataset;

[0114] Specifically, it acquires a large amount of historical tool status data from multiple production lines and under various working conditions, covering the entire lifecycle records of different models and batches, with each data point corresponding to an independent tool;

[0115] Step 322, Input Feature Construction: Calculate the corresponding comparison indicators, including entropy change, from the acquired acoustic emission signals of the tool during operation. , cliff Rate of change of slope of modal dispersion curve And the wavelet packet energy attenuation ratio ΔE / Δf; the calculation process for this step is the same as in step 2;

[0116] Step 323: Calculate the statistics for each comparison indicator feature, and concatenate all the statistics into a high-dimensional vector as input data for the deep learning model;

[0117] Optional statistics include mean, variance, maximum, minimum, skewness, and trend slope;

[0118] Step 324, Label Construction: For each tool, obtain the time when it reaches the wear critical value using offline microscopic measurement methods. The stable value of the feature within the sliding window is then compared with the standard reference threshold vector of the same type of tool. Subtracting them creates the offset label:

[0119] The wear threshold can be set as the tool wear amount. ;

[0120] Specifically, the method for calculating the eigenvalue stability is to extract from... The characteristic stability value is obtained by averaging the comparison index values ​​of each acoustic emission signal within a time window around a given time. ;

[0121] The formula for calculating the offset label is:

[0122] ;

[0123] Step 324: Input the input data into the deep learning model and train the deep learning model with the offset labels as the output.

[0124] Step 325: Calculate the training loss based on the training results, iterate the training until the training cutoff condition is met, and then perform cross-validation to obtain the trained deep learning model; optionally, 5-fold cross-validation can be used.

[0125] Specifically, such as Figure 2 The diagram shows the online monitoring closed loop, demonstrating the real-time monitoring and adaptation process executed by the industrial control computer at the edge: After a new tool begins machining, the edge node collects its initial signals and extracts features. On one hand, it calls the baseline model distributed from the cloud; on the other hand, it inputs the initial features into a deep learning model to predict the individual feature offset Δ of the tool. Finally, the general benchmark thresholds distributed from the cloud will be... Compared with the predicted Δ Add them together to generate a personalized threshold that applies only to the current tool. This is used for subsequent real-time threshold determination and lifetime prediction. The arrows indicate the data interaction between the cloud database and edge nodes, including the distribution of model parameters and thresholds, realizing the working principle of cloud-based collaboration.

[0126] A further technical solution, in step 3, involves comparing the current tool's personalized threshold vector with the calculated comparison indicators to determine the remaining tool life ratio. This method includes the following steps:

[0127] Step 301: When the comparison index exceeds the corresponding threshold in the personalized threshold vector, and it is determined to be abnormal, calculate the normalized overscaling amount of each comparison index value relative to the personalized threshold.

[0128] Specifically, when and ;or, and When an anomaly is detected, the system immediately captures the out-of-range characteristic value of the current window.

[0129] The normalized overscaling of each current feature value relative to its personalized threshold is calculated using the following formula:

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] Step 302: Based on the set weights, the obtained normalized hyperscalar values ​​are weighted and fused to calculate the comprehensive score. ;

[0135] ;

[0136] ;

[0137] in, These are the weighting coefficients;

[0138] Step 303: Calculate the overall score. Mapping yields the percentage of remaining tool life;

[0139] Specifically, a mapping function can be constructed to calculate the comprehensive score. The mapping yields the remaining tool life percentage; the mapping function is as follows:

[0140] ;

[0141] In addition, to improve the robustness of the algorithm, constraints can be introduced:

[0142] ;

[0143] The maximum security score threshold set by the constraint conditions is used to forcibly truncate abnormal score values.

[0144] In the above embodiments, to achieve accurate assessment of the remaining life of the tool during the machining process, the system constructs a comprehensive scoring mechanism based on the difference between the current window acoustic emission feature value and the personalized threshold vector, and maps the score to the remaining life percentage R through a piecewise linear function to drive the early warning system and assist the operator in decision-making;

[0145] In this embodiment, an early warning is issued based on the obtained remaining lifetime ratio, employing a three-level response mechanism:

[0146] Level III Warning: If the remaining lifespan is no more than 30%, yellow LED flashing can be activated at a frequency of 2Hz and a duty cycle of 50%.

[0147] Level II Warning: If the remaining tool life is no more than 10%, a Level II warning will be issued; the remaining tool life and tool coordinates will be pushed to the mobile device.

[0148] Level I Warning: When the lifespan reaches zero, a Level I warning is triggered; an emergency stop command is sent to the PLC to lock the machine tool and prevent workpiece damage; simultaneously, a CSV event log is automatically generated, including timestamp, tool ID, personalized threshold, and entropy change. , cliff Information such as the cause of downtime is recorded and the event log is encrypted and backed up to the cloud daily, supporting quality traceability.

[0149] In step 4, when the remaining lifespan ratio is lower than the set value, a tool replacement operation is performed. The replaced tool is microscopically identified to determine whether the tool can be reused. The comparison index calculated by the tool is used as the input value, and the result of the microscopic identification is used as the output to construct new training data for rolling training of the deep learning model in the cloud.

[0150] When the remaining tool life percentage falls below a set value, a tool change operation is performed. The method for determining whether the replaced tool is reusable by microscopic identification includes the following steps:

[0151] Step 41: For the replaced tools, obtain images of the replaced tools using a photomicroscope;

[0152] Specifically, when a Level I warning triggers a tool change, the old tool is transferred by the engineer to an offline workstation. A 4K high-definition microscope (48MP resolution, 60FPS low latency) is used to photograph the wear area on the flank face, and the wear amount is automatically measured (accuracy ±0.02mm). Wear measurement algorithm:

[0153] Step 42: Convert the tool image to grayscale;

[0154] Step 43: Perform Gaussian filtering on the grayscale image;

[0155] Step 44: Perform Canny edge detection and contour extraction on the filtered image, mark the smallest bounding rectangle, and obtain the target area image of the tool to be identified.

[0156] Step 45: Calculate the length of the wear zone based on the obtained target area image. With depth ;

[0157] Calculation of wear band length ( The method, specifically: establishes a mapping relationship between the image pixel coordinate system and the physical space coordinate system using a preset calibration plate, and determines the conversion coefficient from pixels to millimeters. In the image Cartesian coordinate system, using the original cutting edge profile of the tool as the geometric reference, the starting boundary point of the wear area is automatically identified. and termination boundary point And calculate the Euclidean distance as the length of the wear band. :

[0158] ;

[0159] Depth computing ( The method, specifically:

[0160] Based on digital image processing and curve fitting techniques, precise quantification of depth parameters is performed. Using a discrete set of points in the unworn area of ​​the tool, the standard profile curve of the original cutting edge is generated through least-squares fitting. Feature points are extracted from the lower boundary contour of the wear area, and the actual contour curve of the wear area is obtained through curve fitting. ; at equal intervals along the length of the wear band Set the sampling point sequence { For each sampling point, calculate the wear depth in the vertical direction:

[0161] ;

[0162] The maximum value is used to calculate the depth, and this value is taken as the depth calculation value.

[0163] ;

[0164] in, This is the conversion factor from pixels to millimeters.

[0165] Specifically, five frames of images can be acquired consecutively and calculated separately. and Bilinear interpolation is used to improve calculation accuracy;

[0166] Step 46: Based on the length and depth of the wear band, determine the remaining life and calculate the actual remaining life percentage;

[0167] ;

[0168] in, The maximum permissible wear depth can be set to 0.15mm; This is the total length of the cutting edge of the tool;

[0169] After the above image processing, if >10%, mark as "reusable tool" and put back into inventory; if Once a tool is deemed unusable, it is marked as "scrapped." Data such as the remaining lifespan, model, usage time, and measured wear values ​​of reusable tools are written to a cloud-based MySQL database via an API interface. This microscopic measured data, together with macroscopic acoustic emission characteristic data, constitutes the training samples for the deep learning model, used to optimize the prediction accuracy of individual feature offsets.

[0170] Optionally, new data can be further utilized each day at midnight to fine-tune the learning process using a lightweight fine-tuning mechanism, such as setting the learning rate to 1×10 using the PyTorch framework. -3 The batch size is set to 32, which performs rolling optimization on the deep learning model in the cloud, continuously reducing the prediction error and dynamically updating the personalized threshold vector.

[0171] This embodiment constructs an integrated closed-loop logic architecture for real-time tool life monitoring and early warning, encompassing sensing, analysis, judgment, control, and learning. This differs from the loosely structured, isolated functional modules in existing technologies, which suffer from delayed responses and require manual intervention. An edge industrial control computer continuously collects acoustic emission signals, performing local computation to instantly extract features, determine thresholds, and trigger three-level early warning commands. After an early warning, a 4K microscopic vision module collects old tool wear data and transmits it back to a cloud database. Daily automatic transfer learning model fine-tuning is performed to dynamically evolve the baseline threshold. This architecture forms a continuous, millisecond-level, and adaptive closed loop between physical signal acquisition, real-time judgment logic, equipment control response, and cloud model updates, achieving a leap from static judgment to dynamic feedback learning in tool life monitoring.

[0172] To illustrate the effectiveness of the aforementioned early warning method, a simulation experiment was conducted, as detailed below:

[0173] Traditional dual thresholds are set as follows: K > 5.0, ΔHj > 40%. This embodiment adds a dispersion-attenuation coupling threshold, set as: the rate of change of the dispersion curve slope. >0.18ms / kHz, ΔE / Δf>4.5dB / 100kHz.

[0174] Experimental data: When the cutting length reaches 5500m, the tool wear V B It is approximately equal to 0.062 mm. At this time, K=4.1, ΔHj=33%, and the traditional algorithm did not trigger the warning; however, this embodiment added a comparison index, and the rate of change of the slope of the dispersion curve reached 0.18 ms / kHz, ΔE / Δf=5.1dB / 100kHz, successfully triggering the Level II warning (lifetime ≤10%), proving that the new model has higher sensitivity to initial wear.

[0175] Then, three new cutting tools of the same model (T100, T101, T102) were selected for testing:

[0176] First, the baseline is established: Under standard operating conditions, the general baseline threshold is measured.

[0177] k dispth =0.18ms / kHz, (ΔE / Δf) th =5.0dB / 100kHz;

[0178] Initial data collection: After running the three tools for 10 minutes, their initial characteristic mean values ​​were as follows:

[0179] Standard tool T100: [0.02ms / kHz, 0.8dB / 100kHz];

[0180] The inherently high-frequency T101: [0.08ms / kHz, 2.1dB / 100kHz];

[0181] The inherently low T102 level: [-0.01ms / kHz, -0.5dB / 100kHz];

[0182] The initial features are input into the deep learning model to predict the offset. :

[0183] T101: ;

[0184] The personalized threshold of T101 = [0.18 + 0.06, 5.0 + 1.3] = [0.24ms / kHz, 6.3dB / 100kHz];

[0185] T102: ;

[0186] The personalized threshold for T102 is [0.18-0.03, 5.0-0.7] = [0.15ms / kHz, 4.3dB / 100kHz].

[0187] T101 cutting tool at wear amount V B When the wear is 0.10 mm, the eigenvalue is [0.22, 5.9]. Using a general threshold (0.18, 5.0) will result in serious false alarms; using its specific threshold (0.24, 6.3) will be considered normal until the wear worsens to V. B The warning is accurately triggered only when the feature value is 0.14mm and the threshold value is [0.25, 6.5]. Similarly, the T102 tool avoids the risk of missed detections due to excessively high threshold values. It is evident that the deep learning correction mechanism in this embodiment can adapt to the characteristics of different individuals, significantly improving monitoring accuracy without drastically changing the system architecture.

[0188] Example 2

[0189] Based on Embodiment 1, this embodiment of the present invention provides a real-time tool life monitoring and early warning system, including: an acoustic emission sensor, a signal processing module, a database, an early warning execution module, a microscopic vision device, and a control terminal, wherein the control terminal is configured to execute the steps of the real-time tool life monitoring and early warning method described in Embodiment 1.

[0190] Acoustic emission sensor: used to collect processing signals;

[0191] Signal processing module: Includes a two-stage design of charge amplifier and voltage amplifier, 1V / pC sensitivity, 1.6kHz-1MHz bandwidth, and single power supply of 12-24V;

[0192] Database: Pre-stores baseline thresholds and personalized thresholds for different tool models to provide a basis for early warning;

[0193] Early warning execution module: Triggers corresponding alarms and control actions based on early warning instructions.

[0194] Microscopic vision device: Used for non-contact measurement of the wear area on the flank face of the old tool after tool replacement; this module can include a 48MP industrial camera, telecentric lens and ring light source, with 60FPS low-latency image acquisition capability, measurement accuracy ±0.02mm, and supports automatic identification of wear band length and wear depth; images and measurement results are uploaded to an industrial control computer via USB 3.0 interface. If the wear amount is <0.15mm, it is marked as "reusable" and the data is written to a cloud MySQL database for model fine-tuning and tool life limit exploration.

[0195] Example 3

[0196] Based on Embodiment 1, this embodiment provides a real-time tool life monitoring and early warning system, including:

[0197] The noise processing module is configured to acquire the original acoustic emission signal at the tool working site, preprocess it, and perform adaptive cavitation noise cancellation processing on the preprocessed signal.

[0198] The comparison index calculation module is configured to calculate comparison indices. Based on the acoustic emission signal after adaptive cavitation noise cancellation processing, it performs bandpass filtering and kurtosis statistical analysis to calculate entropy change. with kurtosis Combined with short-time Fourier transform, dispersion-attenuation coupling analysis was performed to calculate the energy attenuation ratio ΔE / Δf.

[0199] The threshold update and early warning module is configured to obtain the personalized threshold vector of the current tool after learning and recognition based on the deep learning model, compare it with the calculated comparison indicators, determine the remaining life ratio and issue an early warning.

[0200] The deep learning model acquires the machining data of the current tool in real time, identifies the individual feature offset of the tool relative to the standard tool, adds the set general benchmark threshold to the predicted individual feature offset vector, and generates a personalized threshold vector applicable to the current tool.

[0201] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0202] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0203] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring and early warning of tool life, characterized in that, Includes the following steps: The original acoustic emission signal at the tool working site is acquired and preprocessed, and the preprocessed signal is subjected to adaptive cavitation noise cancellation processing. Comparison metrics were calculated, and bandpass filtering and kurtosis statistical analysis were performed on the acoustic emission signal after adaptive cavitation noise cancellation processing to calculate entropy change. with kurtosis Combining short-time Fourier transform with dispersion-attenuation coupling analysis, the rate of change of the slope of the modal dispersion curve is calculated. And the energy decay ratio ΔE / Δf; the entropy change , cliff Rate of change of slope And the energy decay ratio ΔE / Δf is used as a comparison index; The personalized threshold vector of the current tool, determined based on a deep learning model, is obtained, compared with various calculated comparison indicators, and an early warning is issued after determining the remaining lifespan ratio. The deep learning model acquires the current tool's machining data in real time, predicts the individual feature offset vector of the tool relative to the standard tool, and adds the set benchmark threshold vector to the predicted individual feature offset vector as the personalized threshold vector of the current tool. A method for adaptive cavitation noise cancellation of preprocessed signals includes the following steps: The preprocessed acoustic emission signal and the reference signal r(n) are synchronously acquired and frequency band filtered to obtain time-domain data with phase alignment and frequency band limited to a set frequency range; Perform spectral analysis and spectral kurtosis discrimination on the reference signal, extract a set number of main frequency components, and construct an orthogonal feature basis vector set; The reference signal is subjected to characteristic basis projection and Hilbert transform to generate an inverse time-domain signal with the opposite phase to the original cavitation noise, i.e., the cavitation noise inverse time-domain signal. The acoustic emission signal after adaptive cavitation noise cancellation is obtained by subtracting the cavitation noise inverse time domain signal from the synchronously acquired preprocessed acoustic emission signal.

2. The method for real-time monitoring and early warning of tool life as described in claim 1, characterized in that: When the remaining lifespan ratio is lower than the set value, a tool replacement operation is performed. The replaced tool is microscopically identified to determine whether it can be reused. The tool's calculated comparison index is used as the input value, and the microscopic identification result is used as the output to construct new training data for rolling training of the deep learning model in the cloud. The method for performing microscopic identification on replaced cutting tools to determine whether they are reusable includes the following steps: For the replaced cutting tools, obtain images of the replaced cutting tools using a photomicroscope; The image of the tool changer is converted to grayscale. Apply a Gaussian filter to the grayscale image; For the filtered image, Canny edge detection and contour extraction are performed, and the minimum bounding rectangle is marked to obtain the target area image of the tool to be identified; Based on the obtained target area image, the length and depth of the wear zone are calculated; Based on the length and depth of the wear band, calculate the actual remaining life percentage, and based on the actual remaining life percentage, determine whether the tool can be reused.

3. The method for real-time monitoring and early warning of tool life as described in claim 1, characterized in that: The process of constructing an orthogonal feature basis vector set is as follows: Perform a Fast Fourier Transform on the reference signal r(n); Calculate spectral kurtosis to locate the frequency band where cavitation noise energy is most concentrated; Within the most concentrated frequency band, extract a set number of spectral lines with the highest energy using a set step size; The frequency components corresponding to the extracted spectral lines are orthogonalized using Gram-Schmidt to generate a set of feature basis vectors.

4. The method for real-time monitoring and early warning of tool life as described in claim 1, characterized in that: By performing characteristic basis projection and Hilbert transform on the reference signal, an inverse time-domain signal with the opposite phase to the original cavitation noise is generated. Specifically: The reference signal r(n) is projected onto the characteristic basis φ to obtain the projection coefficients c(m); For each basis vector φ of the characteristic basis φ m Perform the Hilbert transform to obtain the corresponding complex analytic form; We obtain the complex analytic form of the Hilbert transform of the corresponding basis vectors by weighted combination based on the projection coefficients c(m), and obtain the cavitation noise complex envelope signal, and obtain the instantaneous amplitude A(n) and instantaneous phase θ(n) of the cavitation noise. Construct the inverse signal of the cavitation noise complex envelope signal, and generate an antiphase cancellation signal through inverse mapping of the characteristic basis. That is, the cavitation noise inverted time domain signal.

5. The method for real-time monitoring and early warning of tool life as described in claim 1, characterized in that: Based on the acoustic emission signal after adaptive cavitation noise cancellation processing, bandpass filtering and feature statistical analysis are performed to calculate the entropy change. with kurtosis The process includes the following steps: Apply a bandpass filter to the acoustic emission signal after adaptive cavitation noise cancellation processing; Set a sliding window, and calculate the kurtosis within each sliding window. and wavelet packet entropy change .

6. The method for real-time monitoring and early warning of tool life as described in claim 1, characterized in that: Based on the acoustic emission signal after adaptive cavitation noise cancellation processing, dispersion-attenuation coupling analysis is performed using short-time Fourier transform to calculate the rate of change of the slope of the modal dispersion curve. and wavelet packet energy attenuation ratio The process includes the following steps: A short-time Fourier transform is performed on the acoustic emission signal to obtain the frequency-time distribution spectrum curve; Extract the S0 / A0 mode dispersion curves of the Lamb wave from the spectral curve, calculate the rate of change of the slope of the mode dispersion curves, and calculate the wavelet packet energy attenuation rate ratio within the set frequency band. S0 is the symmetric mode, and A0 is the antisymmetric mode.

7. The method for real-time monitoring and early warning of tool life as described in claim 1, characterized in that: The method for generating the personalized threshold vector of the current cutting tool includes the following steps: Based on the current tool model, coating, and material hardness information, obtain the current tool's baseline threshold vector. ; The comparison indexes of all acoustic emission signals of the current tool before the current time are obtained, the statistics of each comparison index are calculated, and the data are input into the pre-trained deep learning model for prediction to obtain the individual feature offset vector. The baseline threshold vector is added to the predicted individual feature offset vector to generate a value suitable for the current tool T. new Personalized threshold vectors.

8. A real-time tool life monitoring and early warning system, characterized in that, include: An acoustic emission sensor and a control terminal, wherein the control terminal is configured to perform the steps of a real-time tool life monitoring and early warning method according to any one of claims 1-7.

9. A real-time tool life monitoring and early warning system, characterized in that, include: The noise processing module is configured to acquire the original acoustic emission signal at the tool working site, preprocess it, and perform adaptive cavitation noise cancellation processing on the preprocessed signal. The comparison index calculation module is configured to calculate comparison indices. Based on the acoustic emission signal after adaptive cavitation noise cancellation processing, it performs bandpass filtering and kurtosis statistical analysis to calculate entropy change. with kurtosis Combining short-time Fourier transform with dispersion-attenuation coupling analysis, the rate of change of the slope of the modal dispersion curve is calculated. And the energy decay ratio ΔE / Δf; the entropy change , cliff Rate of change of slope And the energy decay ratio ΔE / Δf is used as a comparison index; The threshold update and early warning module is configured to obtain the personalized threshold vector of the current tool determined based on the deep learning model, compare it with the calculated comparison indicators, and issue an early warning after determining the remaining life ratio. The deep learning model acquires the current tool's machining data in real time, predicts the individual feature offset vector of the tool relative to the standard tool, and adds the set benchmark threshold vector to the predicted individual feature offset vector as the personalized threshold vector of the current tool. A method for adaptive cavitation noise cancellation of preprocessed signals includes the following steps: The preprocessed acoustic emission signal and the reference signal r(n) are synchronously acquired and frequency band filtered to obtain time-domain data with phase alignment and frequency band limited to a set frequency range; Perform spectral analysis and spectral kurtosis discrimination on the reference signal, extract a set number of main frequency components, and construct an orthogonal feature basis vector set; The reference signal is subjected to characteristic basis projection and Hilbert transform to generate an inverse time-domain signal with the opposite phase to the original cavitation noise, i.e., the cavitation noise inverse time-domain signal. The acoustic emission signal after adaptive cavitation noise cancellation is obtained by subtracting the cavitation noise inverse time domain signal from the synchronously acquired preprocessed acoustic emission signal.

Images