Acoustic emission signal-based grinding wheel residual life prediction method and system
By using preprocessing based on acoustic emission signals and deep learning models, the real-time and accuracy issues of judging the wear state of grinding wheels are solved, achieving high-precision prediction and early warning of the remaining life of grinding wheels, which is suitable for intelligent machining systems of high-precision coordinate grinding machines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN DIGITAL DESIGN & MANUFACTURING INNOVATION CENTER CO LTD
- Filing Date
- 2025-12-01
- Publication Date
- 2026-04-28
AI Technical Summary
In high-precision coordinate grinding, existing technologies lack real-time and accuracy in judging the wear state of grinding wheels, making it difficult to predict the remaining life of grinding wheels across different working conditions and grinding wheel types.
A method based on acoustic emission signals is adopted. By preprocessing acoustic emission signal data at different wear stages, feature vector sets are extracted, a training dataset is constructed, and a deep regression model of acoustic emission signal time-frequency features is used for training. Combined with a convolutional-bidirectional long short-term memory hybrid deep network model (CNN-BiLSTM), the remaining life of grinding wheels is predicted, so as to achieve real-time and accurate life prediction and early warning.
It achieves high-precision, real-time identification and life assessment of grinding wheel wear conditions, has rapid response and adaptive update capabilities, is applicable to different types of grinding machines and grinding wheels, and provides high-precision grinding wheel condition monitoring and early warning support.
Smart Images
Figure CN121935877A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent manufacturing and machine tool condition monitoring technology, and more specifically, to a method and system for predicting the remaining life of a grinding wheel based on acoustic emission signals. Background Technology
[0002] In high-precision coordinate grinding, the wear condition of the grinding wheel directly affects machining accuracy, surface roughness, and dimensional stability. Grinding wheel wear typically includes abrasive passivation, adhesion layer clogging, localized detachment, or crack propagation. Failure to dress or replace the wheel in a timely manner will lead to increased grinding temperature, workpiece burning, machining vibration marks, and dimensional and positional errors.
[0003] Currently, the assessment of grinding wheel condition largely relies on experience or indirect monitoring signals (such as spindle current, grinding force, vibration, etc.), lacking real-time accuracy. Acoustic emission signals can sensitively reflect transient events such as microcracks and abrasive grain breakage within the grinding zone and are considered a highly sensitive information source for identifying grinding wheel wear. However, acoustic emission signals are noisy and significantly affected by changes in operating conditions, making it difficult for traditional threshold methods and statistical feature analysis to achieve generalized predictions across operating conditions and grinding wheel types.
[0004] Achieving high-precision and interpretable prediction of the remaining life of grinding wheels to provide a basis for intelligent decision-making in the grinding process is an urgent technical problem to be solved. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for predicting the remaining life of grinding wheels based on acoustic emission signals, which can improve the accuracy and adaptability of grinding wheel remaining life prediction.
[0006] This invention provides a method for predicting the remaining life of a grinding wheel based on acoustic emission signals, comprising the following steps: S1: Preprocess the acoustic emission signal data at different wear stages to obtain a feature vector set; based on the feature vector set and the wear amount at different wear stages, obtain a training dataset; S2: Use the training dataset to train the deep regression model of the time-frequency features of the acoustic emission signal to obtain the trained deep regression model of the time-frequency features of the acoustic emission signal; S3: The trained acoustic emission signal time-frequency feature deep regression model is used to predict the acoustic emission characteristics in the actual grinding process to obtain the predicted value of the remaining life of the grinding wheel.
[0007] The present invention also provides a grinding wheel remaining life prediction system based on acoustic emission signals, the system comprising the following modules: The training dataset construction module is configured to: preprocess acoustic emission signal data at different wear stages to obtain a feature vector set; and obtain a training dataset based on the feature vector set and the wear amount at different wear stages. The model building and training module is configured to: train the deep regression model of time-frequency features of acoustic emission signals using the training dataset to obtain the trained deep regression model of time-frequency features of acoustic emission signals; The grinding wheel remaining life prediction module is configured to: use the trained acoustic emission signal time-frequency feature deep regression model to predict the acoustic emission characteristics in the actual grinding process, and obtain the predicted value of the grinding wheel remaining life.
[0008] The method and system for predicting the remaining life of a grinding wheel based on acoustic emission signals provided by this invention have the following beneficial effects: This invention addresses the problems of difficulty in real-time identification of grinding wheel wear, inaccurate life assessment, and delayed early warning response during long-term continuous machining in high-precision coordinate grinding machines. The system utilizes an AE sensor to capture weak signals during the grinding wheel wear process, collecting acoustic emission signals and machine tool process parameters. It exhibits rapid response and high sensitivity to early wear characteristics. Multi-scale time-frequency feature extraction, such as wavelet packet decomposition, is performed on the acoustic emission signals to extract multi-dimensional features including energy, envelope, and spectrum. Normalization and temporal encoding are then performed in conjunction with operating parameters. The obtained features are input into a hybrid deep network model (CNN-BiLSTM), which fully exploits the time-frequency features of the signal to achieve high-precision prediction of grinding wheel life. When the life falls below a set threshold, an early warning is triggered and output online via edge computing nodes, interacting with the machine tool control system to provide adjustment or shutdown prompts.
[0009] This invention utilizes high-frequency AE signals to reflect the wear state of grinding wheels, achieving non-contact real-time detection without altering the machining structure. It introduces time-frequency joint features and a deep learning model to accurately capture complex wear evolution patterns and achieve multi-scale feature fusion. Through closed-loop processing from signal acquisition, feature extraction, lifespan prediction to early warning control, it integrates lifespan prediction and early warning. An edge computing module enables rapid on-site inference and online model updates based on new data, allowing for online deployment and adaptive updates. The invention boasts a fast overall response speed, enabling real-time assessment of lifespan status during grinding. The model structure exhibits excellent transferability, applicable to different types of grinding machines and grinding wheels, and can be quickly adapted through retraining. This invention can intelligently identify wear characteristics, predict lifespan in real-time, and adaptively update the wear state of grinding wheels with different grinding materials and wheel types, providing early warning and effective identification of different wear stages. It offers technical support for adaptive adjustment and preventative maintenance of the grinding process, featuring high precision, robustness, and transferability, making it suitable for online monitoring and lifespan early warning in intelligent machining systems such as high-precision coordinate grinding machines. Attached Figure Description
[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the grinding wheel remaining life prediction method based on acoustic emission signals provided by the present invention; Figure 2 This is a flowchart of the overall process for predicting the remaining life of a grinding wheel based on acoustic emission signals provided by the present invention. Figure 3 This is a schematic diagram of the acoustic emission signal acquisition and monitoring system provided by the present invention; Figure 4 This is a diagram of the CNN-BiLSTM hybrid neural network structure provided by the present invention. Detailed Implementation
[0011] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0012] Figure 1 A schematic diagram of the grinding wheel remaining life prediction method based on acoustic emission signals according to this embodiment is shown. In this embodiment, the grinding wheel remaining life prediction method based on acoustic emission signals includes the following steps: S1: Preprocess the acoustic emission signal data at different wear stages to obtain a feature vector set; based on the feature vector set and the wear amount at different wear stages, obtain a training dataset; As an exemplary embodiment, acoustic emission signal data at different wear stages are collected through multiple grinding tests, and the wear amount (such as wear width, surface breakage rate, etc.) is measured by a grinding wheel profile measuring instrument and an optical microscope. The acquisition system synchronously receives signals of spindle speed, cutting force, and coolant flow rate to establish the correspondence between machining conditions and acoustic emission response; In one exemplary embodiment, the acoustic emission signal data is acquired by an acoustic emission signal acquisition module, wherein the sampling frequency of the acoustic emission signal acquisition module is not less than 5 times the system characteristic frequency and the signal-to-noise ratio is not less than 40dB. In one exemplary embodiment, step S1 specifically includes: S11: Perform feature calibration on acoustic emission signal data at different wear stages to obtain calibrated acoustic emission signals; In one exemplary embodiment, the feature calibration includes energy spectrum calibration, amplitude calibration, root mean square value calibration, and event count calibration; S12: Perform bandpass filtering on the calibrated acoustic emission signal to obtain the filtered acoustic emission signal; In one exemplary embodiment, the bandpass filter has a frequency range of 50kHz-1MHz; As an exemplary embodiment, the bandpass filter employs an adaptive bandpass filter algorithm, which automatically adjusts the filter boundary based on the real-time spectral distribution of the acoustic emission signal to improve adaptability to different operating conditions. S13: Perform feature extraction on the filtered acoustic emission signal to obtain acoustic emission signal feature parameters; In one exemplary embodiment, the feature extraction includes short-time Fourier transform, wavelet packet decomposition, and envelope analysis; In one exemplary embodiment, the acoustic emission signal characteristic parameters include time-domain characteristic parameters, frequency-domain characteristic parameters, and time-frequency-domain characteristic parameters; In one exemplary embodiment, the time-domain feature parameters include the root mean square value of acoustic emission, peak value, impulse factor, and kurtosis; the frequency-domain feature parameters include instantaneous frequency distribution, band energy ratio, spectral centroid, and bandwidth; and the time-frequency domain feature parameters include wavelet packet energy entropy and instantaneous energy distribution. S14: Normalize and reduce the dimensionality of the acoustic emission signal feature parameters to obtain a feature vector set; In one exemplary embodiment, the normalization is Z-score normalization or min-max normalization; the dimensionality reduction is principal component analysis dimensionality reduction. S15: Based on the feature vector set and the wear amount at different wear stages, construct a mapping label set between acoustic emission features and wear amount, and obtain a training dataset based on the mapping label set and the feature vector set; In one exemplary embodiment, the different wear stages include an initial wear stage, a stable wear stage, and a failure stage; the initial wear stage represents a wear amount equal to a fraction of the grinding wheel's lifespan. Hereinafter, the stable wear stage refers to a stage where the wear amount exceeds the grinding wheel's life. And not greater than the grinding wheel's lifespan. The failure stage refers to a wear level greater than or equal to the grinding wheel's lifespan. ; S2: Use the training dataset to train the deep regression model of the time-frequency features of the acoustic emission signal to obtain the trained deep regression model of the time-frequency features of the acoustic emission signal; In one exemplary embodiment, the acoustic emission signal time-frequency feature deep regression model includes an input layer, a convolutional layer, a pooling layer, a bidirectional long short-term memory network layer, a fully connected layer, and an output layer connected in sequence. The input layer receives a standardized feature sequence. The convolutional layer captures local feature patterns and suppresses noise. The bidirectional long short-term memory network layer simultaneously captures the positive and negative time dependencies of the feature sequence during wear evolution. The fully connected layer maps the features extracted by the bidirectional long short-term memory network layer to a high-dimensional feature space of the lifetime prediction output. The output layer outputs the remaining service life or wear level probability of the grinding wheel. In one exemplary embodiment, the input layer has a window length of 100 and a dimension of [missing information]. ; The convolutional layer includes 32... Convolution kernel; The pooling layer uses Max pooling operation; The bidirectional long short-term memory network layer includes 128 units each of forward and backward LSTM; The output layer uses a linear activation function to output the remaining service life or wear level probability of the grinding wheel. In one exemplary embodiment, the training of the deep regression model for the time-frequency features of the acoustic emission signal employs the Adam optimization algorithm, with a learning rate of... The loss function is the mean squared error, and the number of iterations is... wheel; In an exemplary embodiment, the training process of the acoustic emission signal time-frequency feature deep regression model further includes a verification stage, which specifically includes: using a laser interferometer to obtain the circular runout of the grinding wheel and the surface roughness of the machined surface as measured parameters, performing correlation analysis with the predicted value of the remaining life of the grinding wheel output by the acoustic emission signal time-frequency feature deep regression model, obtaining a correlation coefficient, and using the correlation coefficient to evaluate the magnitude of the prediction error; S3: The trained acoustic emission signal time-frequency feature deep regression model is used to predict the acoustic emission characteristics in the actual grinding process to obtain the predicted value of the remaining life of the grinding wheel. In one exemplary embodiment, the grinding wheel remaining life prediction method based on acoustic emission signals further includes: When the predicted remaining life of the grinding wheel is lower than the preset remaining life threshold, an early warning signal is generated and sent to the machine tool controller; the controller dynamically adjusts the grinding parameters according to the life status, or prompts the operator to perform grinding wheel dressing / replacement operations; In one exemplary embodiment, the grinding parameters include feed rate and depth of cut; In one exemplary embodiment, the preset remaining life threshold is a fraction of the grinding wheel's remaining life. ; As an exemplary embodiment, the input of the deep regression model for the time-frequency characteristics of acoustic emission signals is a sequence of feature vectors, and the output is a predicted value of the remaining life of the grinding wheel; In an exemplary embodiment, the grinding wheel remaining life prediction method based on acoustic emission signals further includes: performing Kalman filtering on the predicted grinding wheel remaining life value and fusing it with a moving average to eliminate abnormal jump values, and calculating the life confidence interval based on historical trends. In one exemplary embodiment, the confidence level of the lifetime confidence interval is: ; In one exemplary embodiment, the generation function of the preset remaining lifetime threshold is based on a lifetime decay model, which is: ,in For remaining lifespan, For the rated life of the grinding wheel, For wear loss rate factor, This is an empirical coefficient; In one exemplary embodiment, the range of the empirical coefficient is: ; In an exemplary embodiment, the grinding wheel remaining life prediction method based on acoustic emission signals further includes: when a sudden high-energy event occurs in the measured acoustic emission signal during the machining process, an emergency stop signal is generated and the original waveform of the acoustic emission signal in the past 5 seconds is saved as a diagnostic basis; the sudden high-energy event is when the amplitude of the acoustic emission signal is greater than a preset amplitude threshold.
[0013] This embodiment provides a grinding wheel remaining life prediction system based on acoustic emission signals. The system includes the following modules: The training dataset construction module is configured to: preprocess acoustic emission signal data at different wear stages to obtain a feature vector set; and obtain a training dataset based on the feature vector set and the wear amount at different wear stages. The model building and training module is configured to: train the deep regression model of time-frequency features of acoustic emission signals using the training dataset to obtain the trained deep regression model of time-frequency features of acoustic emission signals; The grinding wheel remaining life prediction module is configured to: use the trained acoustic emission signal time-frequency feature deep regression model to predict the acoustic emission characteristics in the actual grinding process, and obtain the predicted value of the grinding wheel remaining life.
[0014] In some embodiments, the above-described method for predicting the remaining life of a grinding wheel based on acoustic emission signals can also be implemented in the following ways.
[0015] In this embodiment, the method for predicting the remaining life of a grinding wheel based on acoustic emission signals includes the following steps: S01: During the machining process of the coordinate grinding machine spindle, a broadband acoustic emission sensor is arranged in the contact area between the grinding wheel and the workpiece; it is connected to a preamplifier and a data acquisition module (sampling rate...). The system acquires acoustic emission signals in real time and performs characteristic calibrations such as energy spectrum, amplitude, root mean square value, and event count. The acquisition system simultaneously receives spindle speed, cutting force, and coolant flow signals to establish the correspondence between machining conditions and acoustic emission responses.
[0016] As an exemplary embodiment, the sampling frequency of the acoustic emission signal acquisition module is not less than 5 times the system characteristic frequency, and the signal-to-noise ratio is not less than... This is to ensure the accuracy of energy and spectral feature extraction.
[0017] S02: Bandpass filtering (frequency range) of the original acoustic emission signal This removes electromagnetic noise and background interference from machine tools.
[0018] Short-time Fourier transform, wavelet packet decomposition, and envelope analysis were used to extract time-domain, frequency-domain, and time-frequency-domain feature parameters, including peak energy, band energy ratio, spectral centroid, kurtosis, and instantaneous frequency distribution. All feature parameters were normalized by Z-score and reduced in dimensionality by principal component analysis to form a feature vector set.
[0019] As an exemplary embodiment, the signal preprocessing employs an adaptive bandpass filtering algorithm, which automatically adjusts the filtering boundary based on the real-time spectral distribution of the acoustic emission signal to improve adaptability to different operating conditions.
[0020] As an exemplary embodiment, the feature extraction module includes at least the following three types of features: time-domain features: root mean square value of acoustic emission, peak value, impulse factor, kurtosis; frequency-domain features: main band energy ratio, spectral centroid, bandwidth; time-frequency domain features: wavelet packet energy entropy, instantaneous energy distribution; the features are input into the depth model after min-max normalization.
[0021] S03: Acoustic emission signal data at different wear stages were collected through multiple grinding tests. Combined with wear measurements (such as wear width and surface breakage rate) obtained using a grinding wheel profile measuring instrument and an optical microscope, a mapping label set between acoustic emission characteristics and wear amount was constructed. The labels are divided according to the proportion of grinding wheel life: initial wear stage (…). Stable wear stage () ) and failure stage ( ).
[0022] S04: A deep regression model for the time-frequency features of acoustic emission signals is constructed using a joint structure of convolutional neural networks and bidirectional long short-term memory networks. The convolutional neural network layer is used for local time-frequency feature extraction, while the bidirectional long short-term memory network layer is used to capture the long-term dependencies of features over time. The model input is a sequence of feature vectors, and the output is the predicted remaining life of the grinding wheel. The Adam optimization algorithm is used during training, with a learning rate of [missing information]. The loss function is the mean squared error, and the number of iterations is... wheel.
[0023] As an exemplary embodiment, the structure of the convolutional neural network-bidirectional long short-term memory model includes: an input layer (dimension...) Two convolutional layers (convolutional kernels) ReLU activation); pooling layer ( Max pooling); bidirectional long short-term memory network layer (64 hidden units); fully connected layer and output layer (output is lifetime prediction value).
[0024] As an exemplary embodiment, during the verification phase, the circular runout of the grinding wheel and the surface roughness Ra of the machined surface are obtained using a laser interferometer as measured parameters. Correlation analysis is then performed between these parameters and the predicted remaining life of the grinding wheel output by the model. The resulting correlation coefficient is... A value close to 1 indicates a small average prediction error.
[0025] S05: During actual grinding, acoustic emission characteristics are input into the trained model in real time, and the model outputs a predicted value for the remaining life of the grinding wheel. When the predicted life is lower than a set threshold (e.g., remaining life), the model outputs a predicted value for the remaining wheel life. When the grinding wheel wears out, the system automatically triggers a warning signal and sends it to the machine tool controller via the TCP / Modbus bus. The controller dynamically adjusts grinding parameters (such as feed rate and depth of cut) based on the wheel's lifespan, or prompts the operator to perform wheel dressing / replacement operations.
[0026] As an exemplary embodiment, the model prediction output is fused with Kalman filtering and moving average to eliminate anomalous jump values, and the lifetime confidence interval is calculated based on historical trends. (Confidence level).
[0027] As an exemplary embodiment, the warning threshold is automatically and adaptively adjusted according to the grinding wheel type and machining conditions, and the threshold generation function is based on a life decay model. ,in For the rated life of the grinding wheel, For wear loss rate factor, For empirical coefficients ( ).
[0028] As an exemplary embodiment, during the processing, if a sudden high-energy event (amplitude) occurs in the acoustic emission signal... Set threshold The system automatically identifies the grinding wheel as broken or abnormally impacted, immediately triggers an emergency stop signal, and saves the original waveform of the acoustic emission signal within the past 5 seconds as a diagnostic basis.
[0029] In some embodiments, the above-described method for predicting the remaining life of a grinding wheel based on acoustic emission signals can also be implemented in the following ways.
[0030] like Figure 2 The diagram shown is the overall flowchart. In this embodiment, the method for predicting the remaining life of a grinding wheel based on acoustic emission signals includes the following steps: Step 1: Acoustic Emission Signal Acquisition and System Construction: A broadband acoustic emission sensor is placed near the grinding wheel spindle of the coordinate grinding machine to acquire high-frequency acoustic emission signals from the grinding wheel-workpiece contact interface during grinding. The sensor is selected with a center frequency of... A piezoelectric AE sensor within the range, through a preamplifier and a high sampling rate ( The data acquisition system is connected to an industrial computer and features data synchronization and real-time caching capabilities. To reduce environmental noise interference, a vibration isolation layer is installed on the inner wall of the spindle housing, and a time-window synchronous triggering mechanism is used to synchronously record the AE signal along with the grinding wheel speed and feed rate, establishing a multi-source coupled dataset.
[0031] Step 2: Signal Preprocessing and Multi-Scale Feature Extraction: The original acoustic emission signal is preprocessed, including denoising (based on wavelet threshold filtering), envelope extraction, and energy normalization. Subsequently, short-time Fourier transform and continuous wavelet transform are used to extract the time-frequency energy distribution of the acoustic emission signal. Eight statistical features are extracted in the time domain, including root mean square, peak factor, kurtosis, and impulse frequency; ten features are extracted in the frequency domain, including dominant frequency, bandwidth-to-energy ratio, and spectral entropy; and the low, medium, and high-frequency energy ratios of the continuous wavelet transform energy matrix are extracted in the time-frequency domain, forming a complete multi-scale feature vector set. Furthermore, principal component analysis is used to reduce the dimensionality of the features, retaining features with a cumulative variance contribution rate exceeding [a certain percentage]. The feature components are used to construct a feature set for model input.
[0032] Step 3: Deep Learning Model Construction and Training: A convolutional-bidirectional long short-term memory network model is established to achieve temporal correlation learning and lifetime prediction of acoustic emission features. The model structure includes: 1. Convolutional layer: One-dimensional convolution is used ( Extracting local patterns from AE features; 2. Pooling layer: Performs feature dimensionality reduction and noise suppression; 3. Bidirectional Long Short-Term Memory Network Layer: Captures the time dependence of AE signals during wear evolution; 4. Fully connected layer: Maps the extracted features to the lifetime prediction output space; 5. Output layer: Outputs the remaining service life or wear level probability of the grinding wheel.
[0033] During model training, a labeled grinding experiment dataset is used, with actual wear or number of processed parts as the lifetime label. The mean squared error loss function is employed, and the network weights are iteratively updated using the Adam optimizer. To prevent overfitting, a Dropout mechanism (proportional to...) is introduced. The model performance was evaluated using 5-fold cross-validation after training.
[0034] Step 4: Online Lifespan Prediction and Multi-Level Early Warning Mechanism: Deploy the trained model to the edge computing module of the grinding machine control system. The system acquires AE signal segments in real time (window length...). It completes rapid feature extraction and model inference, and outputs the current grinding wheel's life prediction value and confidence interval.
[0035] Based on lifespan prediction results and set thresholds, different warning states are triggered in stages: Level 1 Warning (Minor Wear): Remaining Service Life The system indicates that processing can continue; Level 2 Warning (Moderate Wear): Remaining Service Life It is recommended to dress the grinding wheel; Level 3 Warning (Severe Wear and Tear): Remaining Lifespan It automatically sends a stop signal and prevents processing from starting.
[0036] In addition, the system uploads early warning information and life curves to the host computer monitoring platform for historical life trend analysis and equipment health management.
[0037] Step 5: Experimental Verification and System Calibration: A coordinate grinding machine grinding experiment platform was built to collect and calibrate AE signals under different working conditions (depth of cut, feed, and speed). By comparing and analyzing the detection results of machining effects and changes in grinding wheel condition (such as changes in grinding wheel morphology and surface quality) with the model prediction output, the prediction accuracy and stability of the method were verified.
[0038] In some embodiments, the above-described method for predicting the remaining life of a grinding wheel based on acoustic emission signals can also be implemented in the following ways.
[0039] To facilitate understanding, before describing a method and system for predicting the remaining life of grinding wheels based on acoustic emission signals, we will first provide a detailed explanation of the system structure, signal processing flow, deep learning modeling, and life warning strategy.
[0040] 1. Acoustic emission signal acquisition and synchronization mechanism The acoustic emission signal acquisition system provided in this embodiment consists of an acoustic emission sensor, a signal conditioning circuit, a data acquisition card, a process parameter acquisition module, and a synchronization buffer module. Figure 3 The diagram shows the structure of an acoustic emission signal acquisition and monitoring system.
[0041] (1) Hardware structure: The AE sensor adopts a wideband piezoelectric structure with a center frequency of Frequency response range It is mounted on the end face of the grinding wheel spindle or on the support base, and is fixed by bonding with acoustic coupling adhesive to ensure high-fidelity signal transmission. The signal is then passed through a low-noise preamplifier (gain). ) and bandpass filter ( After adjustment, input the data to the data acquisition card (sampling frequency). , (Resolution). Simultaneously, the process parameter acquisition module reads information such as spindle speed, feed rate, depth of cut, and coolant status in real time via the machine tool PLC interface, with a sampling frequency of [missing information]. .
[0042] (2) Synchronization and caching mechanism: The system uses a unified timestamp to align the AE signal with the machine tool parameters, and the data is cached in real time through a circular buffer. Window. Before grinding begins, self-calibration is performed, recording the AE baseline during idle operation for subsequent normalization. During grinding... For sampling windows, each window contains There are 10 sampling points. The data entry format is as follows: .
[0043] This mechanism ensures the synchronization and traceability of multi-source data, providing a high-quality data foundation for subsequent modeling.
[0044] 2. AE signal preprocessing and feature extraction The AE signal contains rich grinding status information, but it is also affected by noise, system response and environmental interference. Therefore, it needs to be systematically preprocessed and feature extracted.
[0045] (1) Signal preprocessing: Bandpass filtering: Fourth order is used Filter, passband Effectively suppresses low-frequency mechanical vibration noise; Envelope demodulation: via The envelope of the AE signal is extracted by transformation to reflect the energy changes in abrasive particle breakage and crack propagation; Wavelet packet decomposition: using The wavelet basis is decomposed into four levels to obtain 16 frequency band energy coefficients; Energy spectrum calculation: Calculate the energy of each sub-band to form an energy distribution vector, which is used for frequency domain feature analysis.
[0046] (2) Feature Extraction and Fusion: Time-domain statistical features were extracted from the envelope signal, including 10 items such as mean, variance, kurtosis, skewness, RMS, energy entropy, spectral centroid, and spectral width. The 16-dimensional wavelet energy features and 10-dimensional statistical features were fused into a 26-dimensional feature vector, and Z-score normalization was performed. This fused feature vector can comprehensively reflect the transient changes and frequency domain energy distribution characteristics of grinding wheel wear.
[0047] 3. CNN-BiLSTM hybrid neural network lifetime prediction model; like Figure 4 The diagram shows the CNN-BiLSTM hybrid neural network structure. To achieve high-precision prediction of the remaining life of the grinding wheel, this embodiment proposes a deep hybrid structure that combines a convolutional neural network and a bidirectional long short-term memory network.
[0048] (1) Network structure design: The input layer receives a standardized 26-dimensional feature sequence (window length). ).
[0049] Convolutional layer: uses 32 convolutional kernels ( ), used to capture local feature patterns and suppress noise; Pooling layer: using max pooling operation ( (), reduce feature dimensionality, and retain key features; BiLSTM layer: 128 units each for forward and backward LSTM, which can simultaneously capture the forward and backward dependencies of time series. Fully connected layer: maps the BiLSTM output to a high-dimensional feature space; Output layer: Using a linear activation function, outputs the predicted remaining life of the grinding wheel.
[0050] (2) Model Training and Validation: The Adam optimizer (learning rate 0.001) was used with mean squared error as the loss function. The training data came from the calibration dataset of multi-round grinding experiments, covering different rotational speeds, feed rates, and cooling conditions. The data was processed during training. train, After 200 iterations, the model performed well on the validation set. A value close to 1 indicates that the model has high prediction accuracy and strong stability.
[0051] (3) Deployment and online prediction: After training, the model is deployed on an embedded industrial PC and communicates with the machine tool control system via TCP / IP or Modbus protocol. The system inputs the latest AE signal window in real time, and the model outputs the remaining life value of the grinding wheel to realize online status prediction.
[0052] 4. Lifespan Early Warning and System Operation Mechanism Predictive system lifespan threshold setting ,when When necessary, the system automatically triggers the alarm module, issuing a maintenance prompt signal or adjusting machining parameters (such as reducing the feed rate or lowering the rotational speed). Simultaneously, the system records the life decay curve, enabling visualized management of the grinding wheel's performance.
[0053] Alarm signals can be displayed on the machine tool's human-machine interface or uploaded to the host computer monitoring system to achieve closed-loop management of "prediction-early warning-maintenance".
[0054] In addition, this system supports multi-grinding wheel number management and can dynamically adjust model weights according to different batches and different grinding conditions to ensure the adaptability and versatility of predictions.
[0055] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A method for predicting the remaining life of a grinding wheel based on acoustic emission signals, characterized in that, Includes the following steps: S1: Preprocess the acoustic emission signal data at different wear stages to obtain a feature vector set; based on the feature vector set and the wear amount at different wear stages, obtain a training dataset; S2: Use the training dataset to train the deep regression model of the time-frequency features of the acoustic emission signal to obtain the trained deep regression model of the time-frequency features of the acoustic emission signal; S3: The trained acoustic emission signal time-frequency feature deep regression model is used to predict the acoustic emission characteristics in the actual grinding process to obtain the predicted value of the remaining life of the grinding wheel.
2. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, Step S1 specifically includes: S11: Perform feature calibration on acoustic emission signal data at different wear stages to obtain calibrated acoustic emission signals; S12: Perform bandpass filtering on the calibrated acoustic emission signal to obtain the filtered acoustic emission signal; S13: Perform feature extraction on the filtered acoustic emission signal to obtain acoustic emission signal feature parameters; S14: Normalize and reduce the dimensionality of the acoustic emission signal feature parameters to obtain a feature vector set; S15: Based on the feature vector set and the wear amount at different wear stages, construct a mapping label set between acoustic emission features and wear amount, and obtain a training dataset based on the mapping label set and the feature vector set.
3. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, The acoustic emission signal time-frequency feature deep regression model includes an input layer, a convolutional layer, a pooling layer, a bidirectional long short-term memory network layer, a fully connected layer, and an output layer connected in sequence. The input layer is used to receive the standardized feature sequence. The convolutional layer is used to capture local feature patterns and suppress noise. The bidirectional long short-term memory network layer is used to simultaneously capture the positive and negative time dependencies of the feature sequence during wear evolution. The fully connected layer is used to map the features extracted by the bidirectional long short-term memory network layer to the high-dimensional feature space of the life prediction output. The output layer is used to output the remaining service life or wear level probability of the grinding wheel.
4. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, The input layer has a window length of 100 and a dimension of [missing information]. The convolutional layer comprises 32 layers. Convolution kernel; the pooling layer uses Max pooling operation; the bidirectional long short-term memory network layer includes 128 units each of forward and backward LSTM; the output layer uses a linear activation function to output the remaining service life or wear level probability of the grinding wheel.
5. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, The deep regression model for the time-frequency features of the acoustic emission signal is trained using the Adam optimization algorithm, with a learning rate of [missing information]. The loss function is the mean squared error, and the number of iterations is... wheel.
6. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, The training process of the acoustic emission signal time-frequency feature deep regression model also includes a verification stage, which specifically includes: using a laser interferometer to obtain the circular runout of the grinding wheel and the surface roughness of the machined surface as measured parameters, performing correlation analysis with the predicted value of the remaining life of the grinding wheel output by the acoustic emission signal time-frequency feature deep regression model, obtaining the correlation coefficient, and using the correlation coefficient to evaluate the magnitude of the prediction error.
7. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, The method for predicting the remaining life of a grinding wheel based on acoustic emission signals further includes: when the predicted remaining life of the grinding wheel is lower than a preset remaining life threshold, a warning signal is generated and sent to the machine tool controller; the controller dynamically adjusts the grinding parameters according to the life status, or prompts the operator to perform grinding wheel dressing / replacement operations.
8. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, The method for predicting the remaining life of a grinding wheel based on acoustic emission signals further includes: performing Kalman filtering on the predicted remaining life of the grinding wheel and fusing it with a moving average to eliminate abnormal jump values, and calculating the life confidence interval based on historical trends.
9. The method for predicting the remaining life of a grinding wheel based on acoustic emission signals according to claim 1, characterized in that, The grinding wheel remaining life prediction method based on acoustic emission signals further includes: when a sudden high-energy event occurs in the measured acoustic emission signal during the machining process, an emergency stop signal is generated and the original waveform of the acoustic emission signal in the past 5 seconds is saved as a diagnostic basis; the sudden high-energy event is when the amplitude of the acoustic emission signal is greater than a preset amplitude threshold.
10. A grinding wheel remaining life prediction system based on acoustic emission signals, characterized in that, The system includes the following modules: The training dataset construction module is configured to: preprocess acoustic emission signal data at different wear stages to obtain a feature vector set; and obtain a training dataset based on the feature vector set and the wear amount at different wear stages. The model building and training module is configured to: train the deep regression model of time-frequency features of acoustic emission signals using the training dataset to obtain the trained deep regression model of time-frequency features of acoustic emission signals; The grinding wheel remaining life prediction module is configured to: use the trained acoustic emission signal time-frequency feature deep regression model to predict the acoustic emission characteristics in the actual grinding process, and obtain the predicted value of the grinding wheel remaining life.