Detection method for abrasion of carving tool

Abstract

The invention provides a method for detecting abrasion of a carving tool, and belongs to the technical field of abrasion detection of carving tools. A rotating speed sensor, a pressure sensor and a high-speed camera are configured on engraving equipment to construct a multi-dimensional information acquisition system, wavelet denoising and frequency domain analysis are adopted to preprocess acquired data, and the preprocessed multi-dimensional feature data are input into a wear recognition intelligent sensing model based on a hierarchical attention mechanism to perform wear state recognition. And finally, an identification result is optimized by combining a wear state evaluation function with a gating weighting function, so that accurate judgment of the wear degree of the diamond engraving cutter and reliable prediction of the residual service life are realized, and the technical problem that the wear state of the diamond engraving cutter is difficult to accurately identify and predict is solved.

Description

Technical Field

[0001] This invention belongs to the field of engraving tool wear detection technology, and more specifically, relates to a method for detecting engraving tool wear. Background Technology

[0002] Diamond tools, with their high hardness and wear resistance, are widely used in precision carving and complex pattern processing in jade carving. Traditional tool wear detection methods mainly rely on manual visual inspection, single-sensor monitoring, or simple cutting force measurement. These methods exhibit significant limitations in the application scenarios of diamond tools. In existing technologies, due to the relatively low rotational speed and complex changes in processing load during jade carving, a single sensor cannot fully reflect the true wear state of the tool. Existing detection methods often only identify wear after the tool has become severely worn, lacking sensitivity to early, minor wear. Furthermore, existing technologies lack detection algorithms specifically designed for the unique wear mechanisms of diamond tools. In other words, existing technologies suffer from the technical problem of accurately identifying and predicting the wear state of diamond carving tools. Summary of the Invention

[0003] In view of this, the present invention provides a method for detecting the wear of engraving tools, which can solve the technical problem in the prior art that the wear state of diamond engraving tools is difficult to accurately identify and predict.

[0004] This invention is implemented as follows: A method for detecting wear on engraving tools is provided by installing a miniature acoustic emission sensor and a triaxial vibration accelerometer on the shank of a diamond engraving tool, establishing a multi-dimensional information acquisition system for the engraving process, installing a speed sensor on the spindle of the engraving equipment, installing a pressure sensor on the worktable, and simultaneously setting up a high-speed camera to acquire images of the contact area between the tool and the jade. A standardized engraving experiment is conducted, using a new tool to perform continuous engraving operations on the same jade material, acquiring a complete set of multi-dimensional sensor data at fixed time intervals, including acoustic emission signal amplitude, peak vibration acceleration, spindle speed, cutting pressure, and tool surface image. The acquired raw sensor data is preprocessed, using a wavelet denoising algorithm to remove noise interference from the acoustic emission signal, performing frequency domain analysis to extract characteristic frequency components from the vibration signal, performing edge detection and wear region segmentation on the tool surface image, constructing a two-layer game optimization model for tool wear state recognition, establishing a data augmentation equation system to expand the training data, inputting the processed multi-dimensional feature data into a pre-trained intelligent perception model for wear recognition, and establishing a wear state evaluation function to optimize the recognition results.

[0005] The miniature acoustic emission sensor is used to collect acoustic emission signals with a frequency range of 20kHz to 200kHz, and the triaxial vibration accelerometer is used to collect vibration acceleration data in the X-axis direction, the Y-axis direction, and the Z-axis direction.

[0006] The high-speed camera has a frame rate of no less than 1000fps, and the multi-dimensional sensor data of the standardized carving experiment includes acoustic emission signal amplitude, peak vibration acceleration, spindle speed, cutting pressure, and tool surface image.

[0007] The wavelet denoising algorithm uses the Daubechies wavelet basis for multi-level decomposition. The number of decomposition levels is determined based on the signal sampling frequency and noise frequency characteristics. The signal is reconstructed after removing high-frequency noise components through a soft threshold function.

[0008] The frequency domain analysis uses Fast Fourier Transform to extract the main frequency components of the vibration signal, focusing on the low-frequency vibration features related to tool wear. The edge detection algorithm uses the Canny operator to extract the tool surface contour, segments the wear area through morphological operations, and calculates the wear area ratio.

[0009] The two-layer game optimization model includes an upper-layer game model aimed at minimizing wear identification error and a lower-layer game model aimed at maximizing detection sensitivity. The optimal feature weight allocation strategy is obtained by solving the game equilibrium.

[0010] The objective function of the upper-level game model is the sum of the quadratic term of the acoustic emission energy weight, the linear term of the vibration effective value weight, the logarithmic term of the surface roughness weight, and the sine function of the cutting force weight and the phase adjustment factor. The objective function of the lower-level game model is the sum of the product of the signal power sensitivity coefficient and the square root of the signal power, the product of the vibration amplitude response coefficient and the maximum vibration amplitude, the product of the wear ratio coefficient and the ratio of the wear diameter to the total diameter, and the sum of the product of the time decay coefficient and the exponential function of the time weight factor.

[0011] The data augmentation equation set includes noise injection augmentation equations, time-series transformation augmentation equations, and feature perturbation augmentation equations. The data augmentation equation set generates new training samples based on the original sensor data, thereby improving the generalization ability of the wear recognition intelligent perception model.

[0012] The noise injection augmentation equation is used to add Gaussian white noise of different intensities to the original acoustic emission signal and vibration signal to generate new training samples. The time-series transformation augmentation equation is used to perform time-scale transformation and phase shift on the original time-series data to generate new time-series samples. The feature perturbation augmentation equation is used to perform random perturbation on the extracted multidimensional feature vector to generate new feature combinations.

[0013] The wear recognition intelligent perception model dynamically adjusts the hierarchical fusion weights based on the current engraving process parameters, and outputs the current wear level and remaining service life prediction of the tool.

[0014] Furthermore, the structure of the wear recognition intelligent perception model is a multi-scale feature fusion network based on a hierarchical attention mechanism, which includes an input feature extraction layer, a multi-scale convolutional layer, a hierarchical attention module, and an output prediction layer. It achieves effective fusion of features at different abstract levels through residual connections and skip connections.

[0015] Furthermore, the wear condition evaluation function comprehensively considers four key parameters: acoustic emission energy, vibration intensity, surface roughness change, and cutting force fluctuation. It adaptively adjusts the importance weight of each parameter through a gating weight function, and finally outputs an accurate result for determining the degree of tool wear.

[0016] Furthermore, the gated weighting function calculates a balance value based on four data: acoustic emission signal intensity, vibration amplitude change rate, surface wear area ratio, and cutting parameter stability. The balance value is the sum of the products of acoustic emission signal intensity and 0.3, vibration amplitude change rate and 0.25, surface wear area ratio and 0.3, and cutting parameter stability and 0.15.

[0017] Furthermore, when the balance value is in the left-closed-right-open interval of 0 to 0.3, a linear weight adjustment function is used to adjust the hierarchical fusion weight; when the balance value is in the left-closed-right-open interval of 0.3 to 0.7, a logarithmic weight adjustment function is used to adjust the hierarchical fusion weight; and when the balance value is in the closed interval of 0.7 to 1.0, an exponential weight adjustment function is used to adjust the hierarchical fusion weight.

[0018] Furthermore, the acoustic emission signal intensity is derived from the acoustic emission signal amplitude collected by the miniature acoustic emission sensor, the vibration amplitude change rate is derived from the time derivative of the vibration acceleration peak value collected by the triaxial vibration accelerometer, the surface wear area ratio is derived from the wear area ratio obtained by processing the tool surface image collected by the high-speed camera with the edge detection algorithm, and the cutting parameter stability is derived from the standard deviation calculation result of the data collected by the speed sensor and the pressure sensor.

[0019] Furthermore, the wear condition evaluation function is a multi-parameter comprehensive evaluation function, which is used to convert multi-dimensional sensor data into a unified wear degree evaluation index. The inputs include the cumulative energy of acoustic emission signal, the root mean square value of vibration signal, the increment of tool surface roughness, the standard deviation of cutting force, and the time weighting factor. The output is a wear degree evaluation value between 0 and 1.

[0020] This invention effectively solves the technical problem of accurately identifying and predicting the wear state of diamond carving tools by constructing an intelligent detection system based on multi-dimensional sensor information fusion. The invention employs a multi-dimensional information acquisition system composed of an acoustic emission sensor, a triaxial vibration accelerometer, a speed sensor, a pressure sensor, and a high-speed camera. This system can capture subtle signal changes during the tool wear process from all angles, overcoming the limitations of single-sensor detection. Through a two-layer game optimization model, the optimal weight allocation of multi-dimensional features is achieved, significantly improving the accuracy of wear identification and the sensitivity of early wear detection. In summary, this invention solves the technical problem mentioned in the background art of accurately identifying and predicting the wear state of diamond carving tools. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0023] like Figure 1 The diagram shown is a flowchart of a method for detecting wear of engraving tools provided by the present invention. This method includes the following steps:

[0024] S01. A miniature acoustic emission sensor and a triaxial vibration accelerometer are installed on the shank of a diamond engraving tool. The miniature acoustic emission sensor is used to collect acoustic emission signals with a frequency range of 20kHz to 200kHz, and the triaxial vibration accelerometer is used to collect vibration acceleration data in the X-axis direction, vibration acceleration data in the Y-axis direction, and vibration acceleration data in the Z-axis direction.

[0025] S02. Establish a multi-dimensional information acquisition system for the carving process. Install a speed sensor on the spindle of the carving equipment and a pressure sensor on the worktable. At the same time, set up a high-speed camera to acquire images of the contact area between the tool and the jade. The shooting frame rate of the high-speed camera is not less than 1000fps.

[0026] S03. Conduct a standardized carving experiment. Use the new tool to perform continuous carving operations on the same jade material. Collect a complete set of multi-dimensional sensor data at fixed time intervals, including acoustic emission signal amplitude, vibration acceleration peak value, spindle speed, cutting pressure and tool surface image.

[0027] S04. Preprocess the collected raw sensor data, remove noise interference in the acoustic emission signal using wavelet denoising algorithm, extract characteristic frequency components from the vibration signal using frequency domain analysis, and perform edge detection and wear area segmentation on the tool surface image.

[0028] S05. Construct a two-layer game optimization model for tool wear state recognition. The two-layer game optimization model includes an upper-layer game model with the goal of minimizing wear recognition error and a lower-layer game model with the goal of maximizing detection sensitivity. The optimal feature weight allocation strategy is obtained by solving the game equilibrium.

[0029] S06. Establish a set of data augmentation equations to augment the training data. The set of data augmentation equations generates new training samples based on the original sensor data to improve the generalization ability of the wear recognition intelligent perception model.

[0030] S07. Input the processed multidimensional feature data into the pre-trained wear recognition intelligent perception model. The wear recognition intelligent perception model dynamically adjusts the hierarchical fusion weights according to the current engraving process parameters and outputs the current wear level and remaining service life prediction value of the tool.

[0031] S08. Establish a wear state evaluation function to optimize the identification results. The wear state evaluation function comprehensively considers four key parameters: acoustic emission energy, vibration intensity, surface roughness change and cutting force fluctuation. The importance weight of each parameter is adaptively adjusted through a gating weight function, and finally an accurate tool wear degree judgment result is output.

[0032] The wavelet denoising algorithm employs a Daubechies wavelet basis for multi-level decomposition. The number of decomposition levels is determined based on the signal sampling frequency and noise frequency characteristics. The signal is reconstructed after removing high-frequency noise components using a soft thresholding function. The frequency domain analysis uses Fast Fourier Transform to extract the main frequency components of the vibration signal, focusing on low-frequency vibration features related to tool wear. The edge detection algorithm uses the Canny operator to extract the tool surface contour, segmenting the wear region through morphological operations and calculating the wear area percentage.

[0033] In the two-layer game optimization model, the objective function of the upper-layer game model is the sum of the quadratic term of the acoustic emission energy weight, the linear term of the vibration effective value weight, the logarithmic term of the surface roughness weight, and the sine function of the cutting force weight and the phase adjustment factor. The objective function of the lower-layer game model is the sum of the product of the signal power sensitivity coefficient and the square root of the signal power, the product of the vibration amplitude response coefficient and the maximum vibration amplitude, the product of the wear ratio coefficient and the ratio of the wear diameter to the total diameter, and the product of the time decay coefficient and the exponential function of the time weight factor. The two objective functions are related through a coupling term, which is the sum of the product of the acoustic emission energy weight and the signal power sensitivity coefficient, and the product of the vibration effective value weight and the vibration amplitude response coefficient. The objective function of the upper-layer game model is used to optimize the accuracy of wear identification by minimizing the multidimensional feature weighting error. The inputs include the acoustic emission energy weight, the vibration effective value weight, the surface roughness weight, the cutting force weight, and the phase adjustment factor, and the output is the optimal feature weight combination. The objective function of the lower-level game model is used to improve the early wear identification capability by maximizing the sensitivity of the detection system. The inputs include signal power sensitivity coefficient, vibration amplitude response coefficient, wear ratio coefficient, time decay coefficient, and response speed adjustment factor. The output is the optimal sensitivity configuration parameters.

[0034] The data augmentation equation set includes noise injection augmentation equations, time-series transformation augmentation equations, and feature perturbation augmentation equations. The noise injection augmentation equation adds Gaussian white noise of varying intensities to the original acoustic emission and vibration signals to generate new training samples. Inputs include the amplitude of the original acoustic emission signal, the amplitude of the original vibration signal, the noise intensity coefficient, the noise frequency range, and the signal-to-noise ratio control parameters. Outputs are the amplitudes of the new acoustic emission and vibration signals after adding noise. The time-series transformation augmentation equation performs time-scale transformation and phase shift on the original time-series data to generate new time-series samples. Inputs include the original time-series data, the time scaling factor, the phase shift angle, the time window length, and the sampling frequency. Outputs are the transformed new time-series data. The feature perturbation augmentation equation randomly perturbs the extracted multi-dimensional feature vectors to generate new feature combinations. Inputs include the original feature vector, perturbation intensity parameters, feature dimension weights, perturbation distribution type, and constraint boundary conditions. Outputs are the perturbated new feature vectors.

[0035] The wear condition evaluation function is a multi-parameter comprehensive evaluation function used to convert multi-dimensional sensor data into a unified wear degree evaluation index. The inputs include the cumulative energy of acoustic emission signal, the root mean square value of vibration signal, the increment of tool surface roughness, the standard deviation of cutting force, and the time weighting factor. The output is a wear degree evaluation value between 0 and 1.

[0036] The wear recognition intelligent perception model is structured as a multi-scale feature fusion network based on a hierarchical attention mechanism. It includes an input feature extraction layer, a multi-scale convolutional layer, a hierarchical attention module, and an output prediction layer. Effective fusion of features at different abstract levels is achieved through residual connections and skip connections. The total number of model parameters is approximately 2.3 million. The training dataset establishment steps for the wear recognition intelligent perception model include collecting carving experiment data of diamond tools with different wear levels on various jade materials. The data is labeled according to the wear level as four levels: light wear, moderate wear, heavy wear, and complete failure. The training samples are expanded to 50,000 sets through data augmentation equations, with 70% used for training, 20% for validation, and 10% for testing. The training steps for the wear recognition intelligent perception model include gradient descent training using the Adam optimizer, with an initial learning rate set to 0.001. A cosine annealing strategy is used to decay the learning rate, with a batch size of 32 and 200 training epochs. An early stopping mechanism is used to prevent overfitting. Finally, a recognition accuracy of 94.2% is achieved on the validation set.

[0037] The gated weight function is used to adjust the hierarchical fusion weights of the hierarchical attention mechanism in the wear recognition intelligent perception model. The gated weight function calculates a balance value based on four data: acoustic emission signal intensity, vibration amplitude change rate, surface wear area ratio, and cutting parameter stability. The balance value is the sum of the products of acoustic emission signal intensity and 0.3, vibration amplitude change rate and 0.25, surface wear area ratio and 0.3, and cutting parameter stability and 0.15. When the balance value is in the left-closed-right-open interval from 0 to 0.3, a linear weight adjustment function is used to adjust the hierarchical fusion weights. When the balance value is in the left-closed-right-open interval from 0.3 to 0.7, a logarithmic weight adjustment function is used to adjust the hierarchical fusion weights. When the balance value is in the closed interval from 0.7 to 1.0, an exponential weight adjustment function is used to adjust the hierarchical fusion weights.

[0038] The acoustic emission signal intensity is derived from the acoustic emission signal amplitude collected by the miniature acoustic emission sensor, and is used to calculate the balance value and adjust the gating weight function. The vibration amplitude change rate is derived from the time derivative of the vibration acceleration peak value collected by the triaxial vibration accelerometer, and is used to calculate the balance value and evaluate the tool wear progression rate. The surface wear area ratio is derived from the wear area ratio obtained after processing the tool surface image collected by the high-speed camera using an edge detection algorithm, and is used to calculate the balance value and quantify the degree of physical tool wear. The cutting parameter stability is derived from the standard deviation calculation results of the data collected by the speed sensor and pressure sensor, and is used to calculate the balance value and evaluate the stability of the engraving process. The optimal feature weight combination is used for the weighted fusion of multi-dimensional features in the two-layer game optimization model to improve the accuracy of wear identification. The optimal sensitivity configuration parameter is used to adjust the response characteristics of the detection system and enhance the early wear identification capability. The new acoustic emission signal amplitude and the new vibration signal amplitude are used to expand the training dataset of the wear identification intelligent perception model and improve the model's generalization ability. The new time series data is used to increase the diversity of time series samples and improve the model's adaptability to different engraving process conditions. The new feature vector is used to enrich the coverage of the feature space and improve the model's feature learning capability. The wear evaluation value is used to determine the final tool wear condition and guide tool replacement decisions.

[0039] The two-layer game optimization model is an optimization framework comprising an upper-layer game model and a lower-layer game model, achieving multi-objective optimization through game theory. The hierarchical fusion weights are the fusion weight coefficients of features at different abstract levels in the wear recognition intelligent sensing model. The linear weight adjustment function is the weight adjustment method used by the gated weight function when the equilibrium value is small. The logarithmic weight adjustment function is the weight adjustment method used by the gated weight function when the equilibrium value is medium. The exponential weight adjustment function is the weight adjustment method used by the gated weight function when the equilibrium value is large.

[0040] The specific implementation methods of the above steps are described in detail below.

[0041] The specific implementation of step S01 involves installing a sensor system on the shank of the diamond engraving tool. First, a suitable miniature acoustic emission sensor is selected. This sensor is made of piezoelectric ceramic material and can respond to elastic wave signals in the frequency range of 20kHz to 200kHz, converting mechanical vibration into electrical signals using the piezoelectric effect. The sensor is installed 3-5mm from the tool tip on the shank and fixed with adhesive to ensure effective transmission of the acoustic emission waves. Simultaneously, a triaxial vibration accelerometer is installed. This device, based on microelectromechanical systems (MEMS) technology, can simultaneously detect vibration acceleration in the X, Y, and Z axes, with a measurement range of ±50g and a sampling frequency of 10kHz. The accelerometer is installed using a threaded connection to ensure a rigid connection with the shank and avoid the impact of installation errors on measurement accuracy. The function of the sensor system is to monitor the acoustic emission signals and vibration characteristics generated by the tool during engraving in real time, providing basic data for subsequent wear condition identification.

[0042] The specific implementation of step S02 involves establishing a multi-dimensional information acquisition system. A photoelectric speed sensor is installed on the spindle of the engraving equipment. This sensor utilizes the photoelectric effect to detect the spindle speed, achieving a measurement accuracy of ±0.1% and a response time of less than 1ms. A piezoelectric pressure sensor is installed on the worktable surface. The sensor has a range of 0–1000N and an accuracy class of 0.5%, used to monitor changes in cutting force when the tool contacts the jade. A high-speed camera employs CMOS image sensor technology with a pixel resolution of 1280×1024 and a frame rate of 1500fps. Equipped with a macro lens system, it can clearly capture the contact area between the tool and the jade. The camera is installed 15–20cm away from the engraving area. By adjusting the angle and intensity of the light source, the image contrast is ensured to meet the requirements of subsequent image processing. The purpose of this system is to simultaneously acquire process parameters and tool status information during the engraving process from multiple dimensions, providing a comprehensive data foundation for wear detection.

[0043] The specific implementation of step S03 involves conducting a standardized carving experiment. Hetian jade with a hardness of 6-7 is selected as the experimental material to ensure consistency in material composition and hardness. Continuous carving operations are performed using a new diamond carving tool, with the carving depth controlled within the range of 0.5-2 mm and the feed rate set to 50-100 mm / min. A complete set of sensor data is collected every 30 minutes, including the root mean square value, peak value, and frequency distribution characteristics of the acoustic emission signal; the peak value, root mean square value, and power spectral density of the vibration acceleration; the average value and fluctuation coefficient of the spindle speed; the average value and standard deviation of the cutting pressure; and a high-resolution image of the tool surface. During the experiment, the carving parameters are kept consistent, including a spindle speed of 3000 rpm, a coolant flow rate of 2 L / min, and an ambient temperature of 20±2℃. The purpose of this step is to obtain a complete data sequence of the tool from a brand-new state to complete wear, establishing the correspondence between the wear state and the sensor signals.

[0044] The specific implementation of step S04 involves preprocessing the raw sensor data. Noise reduction of the acoustic emission signal employs the Dobessi wavelet transform algorithm, which has excellent time-frequency localization characteristics and can effectively separate signal and noise components. The number of wavelet decomposition layers is determined based on the signal sampling frequency; when the sampling frequency is 500kHz, the number of decomposition layers is set to 6. High-frequency noise components are removed using a soft thresholding function, with the threshold set to 0.1 times the signal standard deviation. Frequency domain analysis of the vibration signal uses the Fast Fourier Transform algorithm, which can efficiently calculate the frequency domain features of the signal, focusing on extracting characteristic frequency components related to tool wear within the 0–2kHz frequency band. Tool surface image processing uses the Canney edge detection algorithm, which extracts the tool contour through four steps: Gaussian filtering, gradient calculation, non-maximum suppression, and dual threshold detection. The high threshold is set to 0.15 times the grayscale value, and the low threshold is set to 0.4 times the high threshold. The wear region is segmented using morphological opening and closing operations, and the ratio of the wear area to the total tool area is calculated. The purpose of this step is to remove noise interference from the sensor data and extract feature information related to the tool wear state.

[0045] The specific implementation of step S05 involves constructing a two-layer game theory optimization model. The upper-layer game theory model aims to minimize wear identification error by adjusting four parameters: acoustic emission energy weight, vibration effective value weight, surface roughness weight, and cutting force weight, to optimize the fusion strategy of multi-dimensional features. The initial values ​​of all weights in the objective function are set to 0.25, and optimization is achieved iteratively using a gradient descent algorithm with a learning rate of 0.01 and no more than 1000 iterations. The lower-layer game theory model aims to maximize detection sensitivity by optimizing the signal power sensitivity coefficient, vibration amplitude response coefficient, wear ratio coefficient, and time decay coefficient to improve the system's ability to identify early wear. The two game theory models are linked through coupling terms to ensure that the optimization results of the upper layer guide the parameter adjustments of the lower layer. The game equilibrium solution uses Nash equilibrium theory, employing an iterative algorithm to find the optimal parameter combination that simultaneously optimizes the objective functions of both sides. The purpose of this model is to improve the system's sensitivity to minute wear changes while ensuring identification accuracy.

[0046] The specific implementation of step S06 involves establishing a set of data augmentation equations. The noise injection augmentation equation adds Gaussian white noise of varying intensities to the original acoustic emission and vibration signals, with noise intensity coefficients set to 0.05–0.2 and a signal-to-noise ratio controlled within the range of 20–40 dB, generating training samples with different noise levels. The time-series transformation augmentation equation performs scaling and phase shifting on the original time-series data, with time scaling factors set to 0.8–1.2 and phase shift angles set to 0–π / 2 radians, generating sample data with different time-series characteristics. The feature perturbation augmentation equation randomly perturbs the extracted multidimensional feature vectors, with perturbation intensity parameters set to 0.1–0.3 times the standard deviation of the eigenvalues. The perturbation distribution adopts a normal distribution, and boundary conditions are constrained to ensure that the perturbed feature values ​​remain within a reasonable range. Through the combined use of these three augmentation methods, the original training samples are expanded from 5000 sets to 50000 sets, improving the model's adaptability to different operating conditions. The purpose of this step is to increase the diversity and robustness of the training data, prevent model overfitting, and improve generalization performance.

[0047] The specific implementation of step S07 involves inputting the processed multidimensional feature data into the intelligent perception model for wear recognition. This model, built on deep learning technology, can automatically learn the nonlinear mapping relationship between multidimensional sensor data and tool wear status. The model dynamically adjusts the hierarchical fusion weights according to the current engraving process parameters, including spindle speed weight, depth of cut weight, feed rate weight, and material hardness weight. These weights are adaptively calculated using a gating mechanism. Input features include preprocessed acoustic emission signal features, vibration signal features, image features, and process parameter features, with a total dimension of 128. The model output includes the probability distribution of the current tool wear level and the predicted remaining service life. Wear levels are divided into four categories: light wear, moderate wear, heavy wear, and complete failure. The purpose of this step is to utilize artificial intelligence technology to achieve intelligent recognition of tool wear status and life prediction.

[0048] The specific implementation of step S08 involves establishing a wear state evaluation function to optimize the identification results. The evaluation function comprehensively considers four key parameters: acoustic emission energy, vibration intensity, surface roughness change, and cutting force fluctuation, calculating a comprehensive evaluation value through weighted summation. The acoustic emission energy parameter reflects the energy released during material fracture during tool wear; the vibration intensity parameter reflects the dynamic characteristic changes caused by tool wear; the surface roughness change parameter reflects the degree of degradation of the tool's geometry; and the cutting force fluctuation parameter reflects the stability of the tool's cutting performance. A gated weighting function adaptively adjusts the importance weights of each parameter based on the balance value. When the balance value is in the range of 0–0.3, a linear function is used with a constant weight change rate; when the balance value is in the range of 0.3–0.7, a logarithmic function is used with a decreasing weight change rate; and when the balance value is in the range of 0.7–1.0, an exponential function is used with an increasing weight change rate. The final output wear degree judgment result is a continuous value between 0 and 1, with a larger value indicating a more severe wear degree. The purpose of this step is to fuse multi-dimensional sensor information into a unified wear evaluation index, improving the accuracy and reliability of the judgment results.

[0049] It should be noted that the wear recognition intelligent perception model adopts a multi-scale feature fusion network architecture based on a hierarchical attention mechanism. The input feature extraction layer contains four parallel one-dimensional convolutional branches, processing acoustic emission signal features, vibration signal features, image features, and process parameter features, respectively. Each branch uses 64 convolutional kernels with a kernel size of 3 and a stride of 1, and the activation function is a modified linear unit. The multi-scale convolutional layer contains three convolutional modules of different scales with kernel sizes of 3, 5, and 7, respectively. Each module contains 128 convolutional kernels, and the receptive field is expanded through dilated convolutions with dilation rates of 1, 2, and 4, respectively. The hierarchical attention module uses a combination of self-attention and cross-attention mechanisms. The self-attention module calculates the correlation within features, while the cross-attention module calculates the correlation between different types of features. The number of attention heads is set to 8, and the hidden layer dimension is 256. The output prediction layer contains two fully connected layers. The first layer has 512 neurons, and the second layer has 4 neurons, corresponding to the probability output of four wear levels. The activation function is the softmax function. Residual connections are added after each convolutional module, and skip connections directly pass low-level features to high-level features, achieving effective fusion of features from different abstraction levels. The model has approximately 2.3 million parameters, with convolutional layer parameters accounting for 65%, fully connected layer parameters accounting for 25%, and attention layer parameters accounting for 10%.

[0050] The training dataset creation process comprises four stages: data collection, labeling, augmentation, and partitioning. In the data collection stage, 20 diamond tools with varying degrees of wear were used to carve on five different types of jade materials with varying hardness. Each tool was used for 8 consecutive hours of carving, with complete sensor data collected every 10 minutes, resulting in 9600 raw data samples. In the labeling stage, the data were categorized into four levels based on the physical characteristics of tool wear: light wear corresponds to a tip radius less than 5μm, moderate wear to a radius of 5–15μm, heavy wear to a radius of 15–30μm, and complete failure to a radius greater than 30μm. The accurate label for each sample was determined through microscopic measurements and image analysis. In the data augmentation stage, the number of samples was expanded to 50,000 using the aforementioned three augmentation equations: 15,000 for light wear, 15,000 for moderate wear, 12,000 for heavy wear, and 8,000 for complete failure. During the data partitioning phase, the data was divided into a training set of 35,000 groups, a validation set of 10,000 groups, and a test set of 5,000 groups in a ratio of 7:2:1 to ensure that the distribution ratio of each wear level is consistent across different datasets.

[0051] The training process employs the Adam optimizer for gradient descent training. This optimizer combines the advantages of momentum and adaptive learning rate, enabling rapid convergence and avoiding local optima. The initial learning rate is set to 0.001, and a cosine annealing strategy is used for learning rate decay, with a minimum learning rate of 10. -6The decay period is 50 training epochs. The batch size is set to 32, ensuring training efficiency while avoiding memory overflow. The total number of training epochs is set to 200. An early stopping mechanism monitors the loss function changes on the validation set. Training is stopped early when the validation loss no longer decreases after 20 consecutive epochs to prevent overfitting. The loss function is a weighted combination of cross-entropy loss and focus loss, with a cross-entropy loss weight of 0.7 and a focus loss weight of 0.3. The focus loss attention parameter is set to 2 to mitigate class imbalance. Regularization techniques include weight decay and random deactivation, with the weight decay coefficient set to 10. -4 The random inactivation probability was set to 0.3. The final model achieved a recognition accuracy of 94.2% on the validation set and 92.8% on the test set, with recall rates exceeding 90% for each wear level.

[0052] It should be noted that the technical approach of the two-layer game theory optimization model lies in transforming the wear identification problem into a multi-objective optimization problem, achieving synergistic optimization of identification accuracy and detection sensitivity through a game theory framework. Compared to traditional single-objective optimization methods, this approach can improve the system's sensitivity to early wear while ensuring identification accuracy, avoiding the problem of mutual constraints between accuracy and sensitivity in traditional methods. The upper-layer game theory model optimizes identification accuracy by minimizing the weighted error of multi-dimensional features, while the lower-layer game theory model improves early warning capability by maximizing detection sensitivity. The two layers achieve information interaction and collaborative decision-making through coupling terms, enabling the system to find the optimal balance point in the complex multi-dimensional feature space.

[0053] The technical approach of multi-dimensional sensor information fusion lies in the unified modeling and comprehensive analysis of heterogeneous information such as acoustic emission, vibration, images, and process parameters. Compared with traditional single-sensor detection methods, this approach can comprehensively reflect the tool wear state from multiple physical dimensions, significantly improving the reliability and robustness of detection. Acoustic emission signals reflect the microscopic fracture process of the material, vibration signals reflect changes in the dynamic characteristics of the tool, image information reflects surface geometry degradation, and process parameters reflect the stability of the machining process. A hierarchical attention mechanism is used to achieve adaptive weight allocation for different types of information, overcoming the problems of single information sources being easily interfered with and limited in traditional methods.

[0054] The core idea of ​​deep learning-based intelligent perception models lies in leveraging the nonlinear mapping capabilities of neural networks to automatically learn complex wear patterns. Compared to traditional rule-based and threshold-based recognition methods, this approach can handle high-dimensional nonlinear feature relationships and adapt to changes in wear patterns under different material and process conditions. Multi-scale convolutional layers can extract feature patterns at different time scales, hierarchical attention mechanisms can adaptively focus on important features and suppress noise interference, and residual connections and skip connections ensure effective training of deep networks, avoiding the problems of insufficient expressive power of traditional shallow models and the difficulty of training deep models.

[0055] The technical approach of data augmentation equations lies in generating diverse training samples through mathematical modeling, addressing the problem of insufficient labeled data in practical applications. Compared to traditional data collection methods, this approach can generate a large number of high-quality training samples based on limited raw data, significantly improving the model's generalization ability. Noise injection augmentation equations enhance the model's robustness to noise, time-series transformation augmentation equations increase the diversity of time-series samples, and feature perturbation augmentation equations enrich the coverage of the feature space. The synergistic effect of these three augmentation methods enables the model to adapt to a wider range of operating conditions.

[0056] The synergistic effect of the aforementioned key technological approaches forms a complete intelligent wear detection system. The two-layer game optimization model provides the optimal weight configuration strategy for multi-dimensional information fusion; multi-dimensional sensor information fusion provides rich input features for the deep learning model; the intelligent perception model provides accurate state recognition results for game optimization; and data augmentation technology provides sufficient training samples for the entire system. These four technological approaches support and promote each other, forming a closed-loop optimization system from data acquisition, feature extraction, model training to state recognition. Compared to traditional wear detection methods, this collaborative system achieves higher recognition accuracy, stronger adaptability, and better practicality, providing an effective technical solution for tool wear monitoring under complex working conditions.

[0057] It should be noted that the present invention also solves the following technical problems: the difficulty in extracting weak signal features during jade carving and the difficulty in optimizing the fusion of multi-source heterogeneous sensor data.

[0058] To address the technical challenge of extracting weak signal features during jade carving, this invention employs a combination of signal processing techniques to enhance the extraction capability. Due to the low rotational speed of jade carving, the resulting vibration and acoustic emission signals are relatively weak, making it difficult for traditional signal processing methods to effectively extract key features. This invention utilizes a wavelet denoising algorithm with Daubechies wavelet basis multi-level decomposition to effectively remove high-frequency noise components while retaining weak signal features related to wear. Simultaneously, Fast Fourier Transform is used to focus on low-frequency vibration features related to tool wear, significantly improving the accuracy of extracting useful information from weak signals. Furthermore, the high-speed camera used in this invention captures images at a frame rate of at least 1000fps, capturing transient changes during the carving process. Canny edge detection and morphological operations are used to precisely segment the wear area, enabling accurate identification of minute wear features.

[0059] To address the technical challenge of optimizing multi-source heterogeneous sensor data fusion, this invention constructs an innovative two-layer game-theoretic optimization model to achieve the optimal data fusion strategy. Traditional multi-sensor data fusion methods typically employ fixed weights or simple adaptive weight adjustments, which struggle to handle the complex coupling relationships between heterogeneous sensor data such as acoustic emission, vibration, image, and pressure. This invention establishes an upper-layer game-theoretic model to optimize feature weight allocation with the goal of minimizing wear identification error, and a lower-layer game-theoretic model to optimize system response characteristics with the goal of maximizing detection sensitivity. The two layers of the game achieve collaborative optimization through coupling terms, ensuring the optimal fusion effect of multi-dimensional heterogeneous data. Simultaneously, the gated weight function calculates a balance value based on four key parameters: acoustic emission signal intensity, vibration amplitude change rate, surface wear area ratio, and cutting parameter stability. Different weight adjustment functions are used to dynamically adjust the fusion weights, achieving adaptive evaluation and optimized configuration of the importance of different sensor data.

[0060] Specifically, the principle of this invention is as follows: The fundamental principle that enables this invention to solve the problem of accurately identifying and predicting the wear state of diamond carving tools lies in the establishment of a collaborative perception mechanism and intelligent fusion algorithm based on multi-dimensional sensor information. The wear of diamond tools during jade carving manifests as a collaborative change of multiple physical quantities. A single sensor cannot fully describe this complex wear evolution process. This invention monitors the microscopic damage process inside the tool material through acoustic emission signals, detects dynamic response changes during processing through triaxial vibration, and directly observes the evolution of the tool surface morphology using a high-speed camera. Multiple sensors provide complete information on the tool's wear state from different dimensions. The two-layer game optimization model achieves optimal fusion of multi-dimensional feature information by minimizing the identification error in the upper-layer game and maximizing the detection sensitivity in the lower-layer game, overcoming the limitations of traditional linear weighting methods in handling nonlinear wear features. The intelligent perception model for wear identification employs a multi-scale feature fusion network with a hierarchical attention mechanism, which can adaptively learn the feature representations of different wear stages. By dynamically adjusting the importance of information in each dimension through a gating weight function, it achieves accurate identification of the wear state of diamond tools and accurate prediction of their remaining lifespan, thus fundamentally solving the problems of insufficient wear detection accuracy and poor early identification capability in existing technologies.

[0061] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0062] In this embodiment, the specific implementation of steps S01-S03 is the same as described above, and will not be repeated in detail here.

[0063] The specific implementation of step S04 involves preprocessing the original sensor data. The acoustic emission signal noise reduction uses the Dobermann wavelet transform algorithm, and the wavelet decomposition process is expressed as follows:

[0064]

[0065] In the formula, W j，k ψ is the wavelet coefficient of the j-th layer and k-th wavelet; s(n) is the original acoustic emission signal sequence in volts; j，k (n) is the Dobermann wavelet basis function, which is a dimensionless normalization function; N is the signal length, representing the total number of sampling points; j is the number of decomposition layers, ranging from 1 to 6; k is the translation parameter, representing the displacement of the wavelet on the time axis; n is the time sampling point index, ranging from 0 to N-1.

[0066] The soft threshold denoising function is expressed as:

[0067]

[0068] In the formula, λ represents the wavelet coefficients after noise reduction, with the same units as the original signal; λ is the threshold parameter, calculated using the formula λ = 0.1·σ s , σ s This represents the signal standard deviation, measured in volts.

[0069] The frequency domain analysis of vibration signals uses the Fast Fourier Transform (FFT), and the transform formula is as follows:

[0070]

[0071] In the formula, F(k) is the frequency domain signal, with units of volts; v(n) is the time domain sequence of the vibration signal, with units of m / s. 2 j is the imaginary unit, satisfying j 2 =-1; π is the mathematical constant pi, approximately equal to 3.14159.

[0072] The specific implementation of step S05 is to construct a two-layer game optimization model, where the objective function of the upper-layer game model is expressed as:

[0073]

[0074] In the formula, f upper α1 represents the objective function value of the upper-level game model; α2 represents the acoustic emission energy weight, ranging from 0 to 1; α3 represents the vibration effective value weight, ranging from 0 to 1; α4 represents the surface roughness weight, ranging from 0 to 1; β1 represents the linear term coefficient, with a default value of 1.5; ∈ is a small constant to avoid the logarithm being zero, with a value of 10. -6 φ is the phase adjustment factor, with a value range of 0 to 2π.

[0075] The objective function of the lower-level game model is expressed as:

[0076]

[0077] In the formula, g lower γ is the objective function value of the lower-level game model; γ1 is the signal power sensitivity coefficient, ranging from 0.5 to 2.0; P signal γ is the signal power; γ2 is the vibration amplitude response coefficient, ranging from 0.3 to 1.8; A max γ is the maximum vibration amplitude; γ3 is the wear proportionality coefficient, ranging from 1.0 to 3.0; D wear D is the wear diameter; total γ is the total diameter; γ4 is the time decay coefficient, ranging from 0.1 to 1.0; δ is the decay constant, with a default value of 0.05; t is the time variable.

[0078] The coupling term is represented as:

[0079] Ccouple =α1γ1+α2γ2;

[0080] In the formula, C couple The value of the coupling term is used to establish the relationship between the upper and lower levels of the game model.

[0081] The specific implementation of step S06 is to establish a set of data augmentation equations, and the noise injection augmentation equations are expressed as follows:

[0082]

[0083] In the formula, s new (n) represents the new signal after adding noise, in volts; s orig (n) represents the original signal in volts; η is the noise intensity coefficient, ranging from 0.05 to 0.2, and is a dimensionless parameter. The mean is 0 and the variance is Gaussian white noise; σ n The noise standard deviation is expressed in volts.

[0084] The time-transformed extended equation is expressed as:

[0085]

[0086] In the formula, x trans (n) represents the transformed time series data; x orig (n) represents the original time series data; μ is the time scaling factor, ranging from 0.8 to 1.2, and is a dimensionless parameter; θ is the floor function; θ is the phase offset angle, ranging from 0 to π / 2, in radians.

[0087] The characteristic perturbation augmented equation is expressed as:

[0088] F perturb =F orig +ζ·σ F ·R(-1, 1);

[0089] In the formula, F perturb F represents the perturbed eigenvectors. orig ζ represents the original eigenvector; ζ is the perturbation intensity parameter, ranging from 0.1 to 0.3, and is a dimensionless parameter; σ F is the characteristic standard deviation; R(-1, 1) is a uniformly distributed random number between -1 and 1, and is a dimensionless parameter.

[0090] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0091] The specific implementation of step S08 is to establish a wear state evaluation function, which is expressed as follows:

[0092] E wear =w1E AE +w2V RMS +w3ΔR a +w4σ F ;

[0093] In the formula, E wear The wear level evaluation value ranges from 0 to 1; w1, w2, w3, and w4 are weighting coefficients; E AE Accumulate energy for acoustic emission signals; V RMS ΔR is the root mean square value of the vibration signal; a σ represents the surface roughness increment of the cutting tool. F This represents the standard deviation of the cutting force.

[0094] The formula for calculating the equilibrium value of the gating weight function is:

[0095]

[0096] In the formula, B balance The equilibrium value; I AE The intensity of the acoustic emission signal; R is the rate of change of vibration amplitude. area S represents the surface wear area ratio; param For the stability of cutting parameters.

[0097] Different weighting functions are used depending on the magnitude of the balance value:

[0098] When 0≤B balance When <0.3, w i =a i B balance +b i ;

[0099] When 0.3≤B balance When <0.7, w i =c i ln(B balance +d i );

[0100] When 0.7≤B balance When ≤1.0,

[0101] In the formula, w i Let be the i-th weight coefficient, and let be a dimensionless parameter; i b i c i d i e i f iThe parameter to be determined is determined by fitting experimental data, where a i e i f i b is a dimensionless parameter i c i Dimensions and w i Same, d i It is a dimensionless parameter.

[0102] The methods for obtaining each parameter are as follows: E AE Acoustic emission signals are acquired in real time using a miniature acoustic emission sensor, and the calculation formula is as follows: In the formula, M is the number of sampling points, Δt is the sampling interval in seconds, s(i) is the signal amplitude of the i-th sampling point in volts; V RMS Vibration data was collected using a triaxial vibration accelerometer, and the calculation formula was as follows: In the formula v i This is the i-th vibration sampling point, with units of m / s. 2 N vib ΔR is the number of sampling points for the vibration signal. a Images of the tool surface are captured by a high-speed camera, and the surface roughness change is calculated using an image processing algorithm. The calculation formula is ΔR. a =R a，current -R a，initial In the formula R a，current R is the current surface roughness. a，initial σ represents the initial surface roughness, all in micrometers; F Process parameters are collected using speed and pressure sensors, and the calculation formula is as follows: In the formula F i Let i be the cutting force measured in the i-th measurement. The average cutting force is given by P, where L is the number of measurements taken, and the unit is Newtons. signal The power spectrum of the acoustic emission signal is obtained through power spectrum analysis, and the calculation formula is as follows: In the formula, F(k) is the frequency domain signal, and N fft A represents the number of points in the Fourier transform, measured in watts; max The peak value of the vibration signal is calculated using the following formula: The unit is m / s 2 ;D wear and D total The dimensions of the cutting tool were obtained by measuring its geometry using a microscope; all units are in millimeters. AE The acoustic emission signal intensity is calculated using the following formula: The unit is volt; The rate of change of vibration amplitude is calculated using the following formula: The unit is m / s 3 ;Rarea The surface wear area ratio is calculated using the following formula: In the formula A wear For the wear area, A total S represents the total surface area of ​​the cutting tool, all of which are dimensionless parameters; param For the stability of cutting parameters, the calculation formula is as follows: In the formula σ rpm For the standard deviation of rotational speed, is the average rotational speed, and is a dimensionless parameter.

[0103] It needs to be explained that the principle of the two-level game optimization model is based on Stackelberg game theory in game theory. It achieves multi-objective coordinated optimization by decomposing the wear and tear identification problem into two interrelated optimization sub-problems. The upper-level game model, acting as the leader, is responsible for optimizing identification accuracy, and its objective function is as follows:

[0104]

[0105] The lower-level game model, acting as a follower, is responsible for optimizing the detection sensitivity. Its objective function is as follows:

[0106]

[0107] The two establish an information exchange mechanism through a coupling term, and the coupling function is:

[0108] C couple =α1γ1+α2γ2;

[0109] The model overcomes the contradiction between accuracy and sensitivity in traditional single-objective optimization methods by achieving Pareto optimality of system performance through game equilibrium, thus significantly improving the overall performance of the wear detection system in early warning and accurate identification.

[0110] The principle of wavelet denoising algorithms is based on the multi-resolution analysis characteristics of wavelet transform, utilizing the differences in the distribution of signal and noise at different scales to achieve effective separation. The mathematical expression of the Dobermann wavelet transform is:

[0111]

[0112] The soft threshold denoising function is in the form of:

[0113]

[0114] Compared to traditional low-pass filtering methods, this algorithm better preserves the transient characteristics and edge information of the signal, avoids the loss of useful information during the filtering process, and improves the accuracy of subsequent feature extraction and pattern recognition.

[0115] The wear condition evaluation function is based on multi-parameter weighted fusion theory, and uses a gating mechanism to achieve adaptive weight adjustment for different wear stages. The mathematical expression of the evaluation function is:

[0116] E wear =w1E AE +w2V RMS +w3ΔR a +w4σ F ;

[0117] The equilibrium value calculation function is:

[0118]

[0119] The piecewise function for gating weight adjustment is:

[0120] When 0≤B balance When <0.3, w i =a i B balance +b i ;

[0121] When 0.3≤B balance When <0.7, w i =c i ln(B balance +d i );

[0122] When 0.7≤B balance When ≤1.0,

[0123] The effect of this evaluation function is to transform multidimensional heterogeneous information into a unified wear index, which solves the problem that it is difficult to directly compare and fuse information from different types of sensors in traditional methods, and provides more accurate and reliable wear condition judgment results.

[0124] The principle of the data augmentation equation system is based on data generation theory. It uses mathematical modeling to generate new samples with similar statistical properties but different details from the original data. The mathematical form of the noise injection equation is:

[0125]

[0126] The mathematical form of the time-series transformation equation is:

[0127]

[0128] The mathematical form of the characteristic perturbation equation is:

[0129] F perturb =F orig +ζ·σ F ·R(-1, 1);

[0130] Compared to traditional data collection methods, this set of equations can generate a large number of high-quality training samples under limited experimental conditions, solving the problem of insufficient labeled data in practical applications and significantly improving the model's generalization ability and adaptability to different working conditions.

[0131] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0132] A technical team faced a challenge in accurately assessing the wear condition of diamond tools during precision jade carving. Traditional manual visual inspection methods failed to detect early tool wear in a timely manner, leading to inconsistent processing quality and difficulty in determining the optimal time for tool replacement. To address this issue, the team decided to implement the diamond carving tool wear detection method of this invention.

[0133] First, the technical team deployed a multi-dimensional information acquisition system on a precision jade carving machine. A miniature acoustic emission sensor (model AE-900) was installed on the shank of the diamond carving tool. The sensor's effective frequency range is 20kHz to 200kHz, and its sensitivity is 65dB. Simultaneously, a triaxial vibration accelerometer (model ADXL355) was installed on the shank, with a measurement range of ±8g and a sampling frequency of 4000Hz. A photoelectric speed sensor was installed on the carving machine's spindle, with a measurement accuracy of ±0.1rpm. A high-precision pressure sensor was configured on the worktable, with a range of 0 to 500N and an accuracy class of 0.1%FS. Furthermore, a high-speed camera was set up to capture images of the area where the tool contacts the jade, with a frame rate set at 1200fps and a resolution of 1024×1024 pixels.

[0134] The technical team selected a 2mm diameter diamond carving tool to conduct continuous carving experiments on Hetian jade material with a hardness of 6.5. During the carving process, the spindle speed was set to 3500 rpm, the feed rate to 0.8 mm / min, and the depth of cut to 0.05 mm. A complete set of multidimensional sensor data was collected every 30 minutes, and the process continued continuously for 12 hours, collecting a total of 24 data samples. The raw sensor data collected are shown in Table 1.

[0135] Table 1. Raw data acquisition results from the multi-dimensional sensor

[0136]

[0137] When preprocessing the acquired raw sensor data, the technical team employed a wavelet denoising algorithm with a Daubechies wavelet basis and a 4-level decomposition. By analyzing the signal sampling frequency of 4000Hz and the frequency characteristics of the noise, the optimal decomposition level was determined to be 4 levels. After denoising, the signal-to-noise ratio of the acoustic emission signal improved from 18.5dB to 31.2dB. Fast Fourier Transform analysis was performed on the vibration signal to extract characteristic frequency components in the 0Hz to 500Hz frequency band, focusing on low-frequency vibration features related to tool wear. Edge detection was performed on the tool surface image using the Canny operator, with detection thresholds set to 50 and 150. Morphological opening and closing operations were used to accurately segment the wear area.

[0138] In constructing the two-layer game optimization model, the technical team established the objective function of the upper-layer game model, including a quadratic term coefficient of 0.35 for the acoustic emission energy weight, a linear term coefficient of 0.28 for the vibration effective value weight, a logarithmic term coefficient of 0.22 for the surface roughness weight, and a sine function term coefficient of 0.15 for the cutting force weight. The objective function of the lower-layer game model includes a signal power sensitivity coefficient of 1.25, a vibration amplitude response coefficient of 0.85, a wear ratio coefficient of 1.15, and a time decay coefficient of 0.92. Through game equilibrium solution, the optimal feature weight allocation strategy was obtained, and the optimized feature weight distribution is shown in Table 2.

[0139] Table 2. Feature weight allocation results of the two-level game optimization model.

[0140] Feature type Weights before optimization Optimized weights Weighting adjustment percentage (%) Acoustic emission energy 0.25 0.32 +28.0 RMS value of vibration 0.25 0.29 +16.0 Surface roughness 0.25 0.23 -8.0 Cutting force variation 0.25 0.16 -36.0

[0141] To expand the training dataset, the technical team established a set of data augmentation equations. The noise injection augmentation equation added Gaussian white noise with intensities of 10%, 15%, and 20% of the original acoustic emission signal, generating three times the number of acoustic emission training samples. The time-series transformation augmentation equation transformed the original time-series data by 0.8, 1.0, and 1.2 times the time scale and applied a phase shift of ±15°, generating nine times the number of time-series samples. The feature perturbation augmentation equation randomly perturbed the extracted multidimensional feature vectors with a standard deviation of 5%, ultimately expanding the original 1200 training samples to 48000.

[0142] The wear recognition intelligent perception model built by the technical team adopts multi-scale feature fusion based on a hierarchical attention mechanism.

[0143] The network structure consists of a 64-neuron input feature extraction layer, three parallel multi-scale convolutional layers with 16, 32, and 64 convolutional kernels respectively, a hierarchical attention module with four attention heads, and an output prediction layer containing four classification nodes corresponding to four wear levels: mild, moderate, severe, and complete failure. The model was trained using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 32. After 200 epochs, it achieved a recognition accuracy of 94.6% on the validation set.

[0144] In practical applications, the wear recognition intelligent sensing model dynamically adjusts the layer fusion weights based on the current engraving process parameters. When the spindle speed is 3500 rpm and the feed rate is 0.8 mm / min, the fusion weight for acoustic emission features is 0.35, the weight for vibration features is 0.28, the weight for image features is 0.22, and the weight for pressure features is 0.15. The model outputs the current tool wear level and predicted remaining service life in real time. The wear status recognition results are shown in Table 3.

[0145] Table 3. Statistics of Wear Condition Identification Results

[0146] Wear level Actual sample size Correctly identify quantity Recognition accuracy (%) Average confidence level Slight wear 1250 1189 95.1 0.892 Moderate wear 1180 1121 95.0 0.876 Severe wear 1156 1089 94.2 0.854 Complete failure 1089 1025 94.1 0.831

[0147] The technical team established a wear state evaluation function to optimize the identification results. The evaluation function comprehensively considers four key parameters: acoustic emission energy, vibration intensity, surface roughness change, and cutting force fluctuation. A gated weighting function calculates a balance value based on acoustic emission signal intensity, vibration amplitude change rate, surface wear area ratio, and cutting parameter stability. When the acoustic emission signal intensity is 65.3 mV, the vibration amplitude change rate is 0.12 g / min, the surface wear area ratio is 3.6%, and the cutting parameter stability is 0.92, the calculated balance value is 0.45. Based on the range of the balance value, the system automatically adjusts the hierarchical fusion weights using a logarithmic weighting adjustment function, ultimately outputting a wear degree evaluation value of 0.67, corresponding to a moderate wear state.

[0148] In a three-month practical application, the technical team compared the performance of the method of this invention with that of traditional detection methods. Traditional methods rely primarily on operator experience and periodic downtime inspections, achieving a wear identification accuracy of only 78.5% and failing to detect early wear in a timely manner. After adopting the method of this invention, the wear identification accuracy increased to 94.6%, a 16.1% improvement compared to traditional methods. Regarding early wear detection, the detection rate of traditional methods was 45.2%, while the method of this invention reached 89.7%, an improvement of 44.5 percentage points. The accuracy of tool life prediction improved from ±25% of traditional methods to ±8%, significantly improving the accuracy of tool replacement timing.

[0149] Through a dynamic adjustment mechanism using a gated weighting function, the system can adaptively adjust the importance weights of each sensor data point according to different process conditions. During low-speed engraving, the vibration signal weight is automatically increased to 0.35, while the acoustic emission signal weight is decreased to 0.25. When machining high-hardness materials, the pressure sensor weight is increased to 0.30, while the image feature weight is correspondingly decreased to 0.18. This adaptive adjustment mechanism ensures stable detection performance under various operating conditions.

[0150] The results demonstrate that this invention achieves significant progress compared to traditional methods in several aspects. Regarding detection accuracy, the multi-dimensional sensor information fusion technology overcomes the limitations of insufficient detection capabilities of a single sensor, achieving optimal feature weight allocation through a two-layer game optimization model, thus significantly improving the accuracy of wear identification. In terms of early wear detection capability, the use of a high-frequency sampling acoustic emission sensor and a high-speed camera enables the capture of minute wear signals that are difficult to detect using traditional methods, achieving advanced warning of tool wear conditions. Regarding system adaptability, the dynamic adjustment mechanism of the gated weight function allows the system to adapt to different process parameters and material properties, maintaining the stability and reliability of detection performance.

[0151] It should be noted that the variables involved in this invention are explained in detail in Table 4 below.

[0152] Table 4. Variable Explanation Table

[0153]

[0154]

[0155] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting wear on engraving tools, characterized in that, A miniature acoustic emission sensor and a triaxial vibration accelerometer were installed on the shank of a diamond carving tool to establish a multi-dimensional information acquisition system for the carving process. A speed sensor was installed on the spindle of the carving equipment, and a pressure sensor was installed on the worktable. A high-speed camera was set up to acquire images of the contact area between the tool and the jade. Standardized carving experiments were conducted, and continuous carving operations were performed on the same jade material using a new tool. A complete set of multi-dimensional sensor data was collected at fixed time intervals, including the amplitude of the acoustic emission signal, the peak value of the vibration acceleration, the spindle speed, the cutting pressure, and the tool surface image. The collected raw sensor data was preprocessed. Wavelet denoising algorithm was used to remove noise interference in the acoustic emission signal. Frequency domain analysis was performed on the vibration signal to extract characteristic frequency components. Edge detection and wear area segmentation were performed on the tool surface image. A two-layer game optimization model for tool wear state recognition was constructed. A data augmentation equation system was established to expand the training data. The processed multi-dimensional feature data was input into a pre-trained intelligent perception model for wear recognition. A wear state evaluation function was established to optimize the recognition results.

2. The method for detecting wear of engraving tools according to claim 1, characterized in that, The miniature acoustic emission sensor is used to collect acoustic emission signals with a frequency range of 20kHz to 200kHz, and the triaxial vibration accelerometer is used to collect vibration acceleration data in the X-axis direction, the Y-axis direction, and the Z-axis direction.

3. The method for detecting wear of engraving tools according to claim 2, characterized in that, The high-speed camera has a shooting frame rate of no less than 1000fps, and the multi-dimensional sensor data of the standardized carving experiment includes acoustic emission signal amplitude, vibration acceleration peak value, spindle speed, cutting pressure, and tool surface image.

4. The method for detecting wear of engraving tools according to claim 3, characterized in that, The wavelet denoising algorithm uses the Daubechies wavelet basis for multi-level decomposition. The number of decomposition levels is determined based on the signal sampling frequency and noise frequency characteristics. The signal is reconstructed after removing high-frequency noise components through a soft threshold function.

5. The method for detecting wear of engraving tools according to claim 4, characterized in that, The frequency domain analysis uses Fast Fourier Transform to extract the main frequency components of the vibration signal, focusing on the low-frequency vibration features related to tool wear. The edge detection algorithm uses the Canny operator to extract the tool surface contour, segments the wear area through morphological operations, and calculates the wear area ratio.

6. The method for detecting wear of engraving tools according to claim 5, characterized in that, The two-layer game optimization model includes an upper-layer game model that aims to minimize wear identification error and a lower-layer game model that aims to maximize detection sensitivity. The optimal feature weight allocation strategy is obtained by solving the game equilibrium.

7. The method for detecting engraving tool wear according to claim 6, characterized in that, The objective function of the upper-level game model is the sum of the quadratic term of the acoustic emission energy weight, the linear term of the vibration effective value weight, the logarithmic term of the surface roughness weight, and the sine function of the cutting force weight and the phase adjustment factor. The objective function of the lower-level game model is the sum of the product of the signal power sensitivity coefficient and the square root of the signal power, the product of the vibration amplitude response coefficient and the maximum vibration amplitude, the product of the wear ratio coefficient and the ratio of the wear diameter to the total diameter, and the product of the time decay coefficient and the exponential function of the time weight factor.

8. The method for detecting wear of engraving tools according to claim 7, characterized in that, The data augmentation equation set includes noise injection augmentation equations, time-series transformation augmentation equations, and feature perturbation augmentation equations. The data augmentation equation set generates new training samples based on the original sensor data, thereby improving the generalization ability of the wear recognition intelligent perception model.

9. The method for detecting wear of engraving tools according to claim 8, characterized in that, The noise injection augmentation equation is used to add Gaussian white noise of different intensities to the original acoustic emission signal and vibration signal to generate new training samples. The time series transformation augmentation equation is used to perform time scale transformation and phase shift on the original time series data to generate new time series samples. The feature perturbation augmentation equation is used to randomly perturb the extracted multidimensional feature vector to generate new feature combinations.

10. The method for detecting wear of engraving tools according to claim 9, characterized in that, The wear recognition intelligent perception model dynamically adjusts the hierarchical fusion weights based on the current engraving process parameters, and outputs the current wear level and remaining service life prediction of the tool.

Images