A method for monitoring the wear state of a grinding wheel and related products

Abstract

The application provides a grinding wheel wear state monitoring method and related products, the method comprising: obtaining a grinding force signal and a vibration signal of a grinding wheel grinding process, and obtaining a historical grinding wheel wear state training sample in a database, splicing the grinding force signal and the vibration signal to obtain a feature component matrix; using a whale optimization algorithm, based on the historical grinding wheel wear state training sample, constructing an improved CNN-GRU model to obtain an optimal performance monitoring model; inputting the feature component matrix into the optimal performance monitoring model to calculate the grinding wheel wear state information. A set of methods for detecting the grinding wheel wear state using the optimal performance monitoring model are formed, and the problem that existing grinding wheel state monitoring methods based on deep learning models cannot comprehensively extract hidden features in data is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of grinding wheel state monitoring, and particularly relates to a grinding wheel wear state monitoring method and related products. BACKGROUND

[0002] In the prior art, a robot weld seam grinding uses a large feed depth, so that a large grinding force and a high grinding zone temperature appear, and so on. These cause the grinding wheel to wear quickly and be consumed seriously. Once the grinding wheel is seriously worn, the cutting ability of the abrasive grains will be significantly reduced, so that the grinding quality and precision are not easy to control. Therefore, the grinding wheel wear state is a key factor affecting the performance of the robot grinding. However, in actual production, workers often determine the grinding wheel dressing time according to experience. If dressing is performed in advance, not only will the grinding efficiency be reduced, but also the grinding wheel will be wasted. If dressing is delayed, the grinding wheel is easy to cause serious wear, deteriorate the grinding quality, and even cause grinding burn and chatter, and damage the base material. The workers are difficult to determine the best time for grinding wheel dressing and replacement, which causes great difficulty in further improving the efficiency and quality of the weld seam grinding. Therefore, it is crucial to monitor the wear degree of the grinding wheel in time and accurately. The existing grinding wheel wear state monitoring has problems such as excessive dependence on manual feature extraction and selection, poor robustness in actual production environment, and difficulty in comprehensively extracting data implicit feature information by a single deep learning model.

[0003] The existing grinding wheel wear monitoring methods can be mainly divided into direct monitoring method and indirect monitoring method. The direct monitoring method is to judge the grinding wheel wear state by directly measuring the change of the grinding wheel surface morphology by using an optical or machine vision system, which has high monitoring precision, but is easily affected by the machining process and is difficult to measure on line, and is rarely applied in actual production. The indirect monitoring method is to realize the grinding wheel wear state monitoring by establishing the complex mapping relationship between the state parameters in the machining process and the grinding wheel wear. At present, the signals such as force, vibration, acoustic emission and current in the grinding process are collected, and then the acquired signals are accurately associated with the grinding wheel wear state by using machine learning, so as to realize the grinding wheel wear state monitoring. This indirect monitoring method can realize on-line real-time monitoring, and is more suitable for actual production. It not only avoids the difficult mechanism modeling process, but also has strong universality. However, the traditional machine learning model has two disadvantages. On the one hand, it needs to excessively rely on experience for manual feature extraction and selection, which is time-consuming and laborious and difficult to guarantee the consistency of extraction quality; on the other hand, the data in the machining process increases sharply, and the traditional machine learning method cannot mine the deep features of large-scale multi-source heterogeneous data. The most commonly used deep learning models in the research on the grinding wheel wear state monitoring are deep belief network (DBN), stacked auto-encoding network (SAE), convolutional neural network (CNN) and recurrent neural network (RNN), and the monitoring method based on a single deep learning model is difficult to comprehensively extract the hidden features in the data. The combined deep learning model has limited effect on improving the grinding wheel state monitoring precision, and increases the calculation cost.

[0004] Therefore, it is of great significance to study a grinding wheel wear state monitoring method based on an optimal performance monitoring model to quickly and accurately detect the grinding wheel wear state, ensure the safety of the grinding wheel, and prevent safety accidents of the grinding wheel. SUMMARY

[0005] The present application provides a grinding wheel wear state monitoring method and related products, which solves the problem that the existing grinding wheel state monitoring method based on a deep learning model is difficult to comprehensively extract hidden features in data, and lays an important foundation for the safe use of the grinding wheel.

[0006] In a first aspect, the present application provides a grinding wheel wear state monitoring method, comprising:

[0007] obtaining a grinding force signal and a vibration signal of a grinding wheel grinding process, and obtaining historical grinding wheel wear state training samples in a database; the historical grinding wheel wear state training samples include historical grinding force signals and historical vibration signals of historical grinding wheel grinding processes and corresponding historical grinding wheel wear state information;

[0008] splicing the grinding force signal and the vibration signal to obtain a feature component matrix;

[0009] An improved CNN-GRU model is constructed based on the historical grinding force signals and the historical vibration signals of the historical grinding process of the grinding wheel and corresponding historical grinding wheel wear state information by using a whale optimization algorithm to obtain an optimal performance monitoring model.

[0010] The feature component matrix is input into the optimal performance monitoring model to obtain grinding wheel wear state information through calculation.

[0011] Optionally, the grinding force signals and the vibration signals are spliced to obtain a feature component matrix, including:

[0012] The grinding force signals are denoised to obtain denoised grinding force signals.

[0013] The vibration signals are subjected to symplectic geometric modal decomposition to obtain symplectic geometric modal components.

[0014] The denoised grinding force signals and the symplectic geometric modal components are spliced to obtain a feature component matrix.

[0015] Optionally, an improved CNN-GRU model is constructed based on the historical grinding force signals and the historical vibration signals of the historical grinding process of the grinding wheel and corresponding historical grinding wheel wear state information by using a whale optimization algorithm to obtain an optimal performance monitoring model, including:

[0016] The historical grinding force signals and the historical vibration signals of the historical grinding process of the grinding wheel are spliced to obtain a historical feature component matrix.

[0017] An improved CNN-GRU model is constructed based on the historical feature component matrix and corresponding historical grinding wheel wear state information by using a whale optimization algorithm to obtain an optimal performance monitoring model.

[0018] Optionally, an improved CNN-GRU model is constructed based on the historical feature component matrix and corresponding historical grinding wheel wear state information by using a whale optimization algorithm to obtain an optimal performance monitoring model, including:

[0019] An improved CNN-GRU model corresponding to the historical feature component matrix is constructed to obtain a preliminary detection model.

[0020] The preliminary detection model is trained according to data for training in the historical feature component matrix and corresponding historical grinding wheel wear state information by using a whale optimization algorithm to obtain a trained preliminary detection model.

[0021] The trained preliminary detection model is verified according to data not used for training in the historical feature component matrix to obtain an optimal performance monitoring model.

[0022] Optionally, an improved CNN-GRU model corresponding to the historical feature component matrix is constructed to obtain a preliminary detection model, comprising:

[0023] A batch normalization layer and a BiGRU unit composed of a double-layer bidirectional GRU structure are added in the preset CNN-GRU model to obtain a first CNN-GRU model;

[0024] A Dropout layer is arranged behind each BiGRU unit in the first CNN-GRU model to obtain a second CNN-GRU model;

[0025] A Self-Attention mechanism is introduced in the second CNN-GRU model to construct an improved CNN-GRU model corresponding to the historical feature component matrix to obtain a preliminary detection model

[0026] Optionally, the whale optimization algorithm is used to train the preliminary detection model according to the data for training in the historical feature component matrix and the corresponding historical grinding wheel wear state information to obtain a trained preliminary detection model, comprising:

[0027] The data for training in the historical feature component matrix is input into the preliminary detection model to obtain a corresponding grinding wheel wear state output;

[0028] The training error is determined according to the data for training in the historical feature component matrix, the corresponding historical grinding wheel wear state information and the corresponding grinding wheel wear state output;

[0029] The parameters and hyperparameters of the preliminary detection model are adjusted based on the training error through the whale optimization algorithm to obtain optimal parameters and optimal hyperparameters, and the preliminary detection model is optimized by using the optimal parameters and the optimal hyperparameters to obtain the trained preliminary detection model.

[0030] Optionally, before the data for training in the historical feature component matrix is input into the preliminary detection model to obtain a corresponding grinding wheel wear state output, the method further comprises:

[0031] The parameters of the preliminary detection model are initialized.

[0032] Optionally, the whale optimization algorithm is used to train the preliminary detection model according to the data for training in the historical feature component matrix and the corresponding historical grinding wheel wear state information to obtain a trained preliminary detection model, comprising:

[0033] Based on the training error, the parameters of the preliminary detection model are adjusted to obtain optimal parameters, and the preliminary detection model is optimized by using the optimal parameters to obtain a preliminary trained preliminary detection model;

[0034] The hyperparameters of the preliminary trained preliminary detection model are optimized by a whale optimization algorithm to obtain optimal hyperparameters.

[0035] The preliminary trained preliminary detection model is optimized by using the optimal hyperparameters to obtain the trained preliminary detection model.

[0036] Optionally, the hyperparameters of the preliminary trained preliminary detection model are optimized by a whale optimization algorithm to obtain optimal hyperparameters, comprising:

[0037] Setting a to-be-optimized hyperparameter and initializing the to-be-optimized hyperparameter;

[0038] The to-be-optimized hyperparameter is iterated and the error rate after each iteration is calculated by a preset iteration method and a preset error rate calculation formula;

[0039] The minimum value is extracted from the error rate after each iteration, and the to-be-optimized hyperparameter after iteration corresponding to the minimum value is determined as the optimal hyperparameter.

[0040] In a second aspect, the present application provides a grinding wheel wear state monitoring device, comprising:

[0041] An acquisition module is configured to acquire a grinding force signal and a vibration signal of a grinding wheel grinding process, and acquire historical grinding wheel wear state training samples in a database; the historical grinding wheel wear state training samples include historical grinding force signals and historical vibration signals of historical grinding wheel grinding processes and corresponding historical grinding wheel wear state information;

[0042] A splicing module is configured to splice the grinding force signal and the vibration signal to obtain a feature component matrix;

[0043] A construction module is configured to use a whale optimization algorithm to construct an improved CNN-GRU model based on the historical grinding force signals and the historical vibration signals of the historical grinding wheel grinding processes and the corresponding historical grinding wheel wear state information, to obtain a best performance monitoring model;

[0044] A calculation module is configured to input the feature component matrix into the best performance monitoring model to calculate and obtain grinding wheel wear state information.

[0045] The application provides a grinding wheel wear state monitoring method and related products, and the method comprises the following steps: acquiring a grinding force signal and a vibration signal of a grinding wheel grinding process, and acquiring a historical grinding wheel wear state training sample in a database; the historical grinding wheel wear state training sample comprises a historical grinding force signal and a historical vibration signal of a historical grinding wheel grinding process and corresponding historical grinding wheel wear state information; the grinding force signal and the vibration signal are spliced to obtain a feature component matrix; an improved CNN-GRU model is constructed based on the historical grinding force signal and the historical vibration signal of the historical grinding wheel grinding process and the corresponding historical grinding wheel wear state information by using a whale optimization algorithm, an optimal performance monitoring model is obtained, the feature component matrix is input into the optimal performance monitoring model, and grinding wheel wear state information is calculated. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0047] Figure 1 The flow step diagram of the first embodiment of the grinding wheel wear state monitoring method of the present application;

[0048] Figure 2 The flow step diagram of the second embodiment of the grinding wheel wear state monitoring method of the present application;

[0049] Figure 3 The structure diagram of the improved CNN-GRU model of the present application;

[0050] Figure 4 The flow step diagram of the preliminary detection model training and optimization of the present application;

[0051] Figure 5 The optimal hyperparameter change curve diagram when the whale optimization algorithm is used to optimize the preliminary detection model of the present application;

[0052] Figure 6 The structure block diagram of the grinding wheel wear state monitoring device embodiment of the present application. DETAILED DESCRIPTION

[0053] The embodiment of the present application provides a grinding wheel wear state monitoring method and related products, which are used for solving the problem that the existing grinding wheel state monitoring method based on a deep learning model is difficult to comprehensively extract hidden features in data, and laying an important foundation for safe use of the grinding wheel.

[0054] In order to make the application purpose, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the following described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application. Embodiment one

[0055] Please refer to Figure 1 , Figure 1 The flow step diagram of the first embodiment of the grinding wheel wear state monitoring method of the present application is shown in the figure, and the method comprises the following steps:

[0056] In step S101, the grinding force signal and the vibration signal of the grinding wheel grinding process are obtained, and the historical grinding wheel wear state training sample in the database is obtained; the historical grinding wheel wear state training sample comprises historical grinding force signal and historical vibration signal of the historical grinding wheel grinding process and corresponding historical grinding wheel wear state information;

[0057] The historical grinding force signal is the historical data of the grinding force signal of the grinding wheel grinding process, the historical vibration signal is the historical data of the vibration signal of the grinding wheel grinding process, and the historical grinding wheel wear state information is the historical data of the grinding wheel wear state information of the grinding wheel grinding process.

[0058] In step S102, the grinding force signal and the vibration signal are spliced to obtain a feature component matrix;

[0059] In the embodiment of the present application, the grinding force signal is denoised to obtain the denoised grinding force signal, the vibration signal is subjected to symplectic geometric modal decomposition to obtain a symplectic geometric modal component, and the denoised grinding force signal and the symplectic geometric modal component are spliced to obtain the feature component matrix.

[0060] In step S103, an improved CNN-GRU model (Convolutional Neural Network-Gate Recurrent Unit) is constructed based on the historical grinding force signal and the historical vibration signal of the historical grinding wheel grinding process and the corresponding historical grinding wheel wear state information by using the whale optimization algorithm, so as to obtain an optimal performance monitoring model;

[0061] In an optional embodiment, based on the historical grinding force signals and the historical vibration signals of the historical grinding process of the grinding wheel and the corresponding historical grinding wheel wear state information, an improved CNN-GRU model is constructed to obtain an optimal performance monitoring model, which comprises:

[0062] The historical grinding force signals and the historical vibration signals of the historical grinding process of the grinding wheel are spliced to obtain a historical feature component matrix;

[0063] An improved CNN-GRU model is constructed based on the historical feature component matrix and the corresponding historical grinding wheel wear state information by using a whale optimization algorithm to obtain an optimal performance monitoring model.

[0064] In the embodiment of the application, the historical grinding force signals and the historical vibration signals of the historical grinding process of the grinding wheel are spliced to obtain a historical feature component matrix, an improved CNN-GRU model corresponding to the historical feature component matrix is constructed to obtain a preliminary detection model, the preliminary detection model is trained according to the data for training in the historical feature component matrix and the corresponding historical grinding wheel wear state information by using a whale optimization algorithm, a trained preliminary detection model is obtained, and the optimal performance monitoring model is obtained by training the trained preliminary detection model according to the data not used for training in the historical feature component matrix.

[0065] In step S104, the feature component matrix is input into the optimal performance monitoring model to calculate the grinding wheel wear state information.

[0066] In the embodiment of the application, the feature component matrix is input into the optimal performance monitoring model to calculate the grinding wheel wear state information, and the grinding wheel wear state information includes information such as wear position, wear amount, and abrasive state for describing the grinding wheel wear state.

[0067] The grinding wheel wear state monitoring method provided in the embodiment of the application forms a method for detecting the grinding wheel wear state by using the optimal performance monitoring model, solves the problem that the existing grinding wheel state monitoring method based on a deep learning model is difficult to comprehensively extract hidden features in data, and lays an important foundation for the safe use of the grinding wheel.

[0068] Embodiment two

[0069] Please refer to Figure 2 , Figure 2 The flow step diagram of embodiment two of the grinding wheel wear state monitoring method of the application is shown in the following figure.

[0070] S201, the grinding force signals and the vibration signals of the grinding process of the grinding wheel are obtained, and the historical grinding wheel wear state training samples in the database are obtained; the historical grinding wheel wear state training samples include the historical grinding force signals and the historical vibration signals of the historical grinding process of the grinding wheel and the corresponding historical grinding wheel wear state information;

[0071] The embodiment of the application acquires the grinding force signal and the vibration signal of the grinding wheel grinding process through a sensor, and acquires a historical grinding wheel wear state training sample in a database, the historical grinding wheel wear state training sample including a historical grinding force signal and a historical vibration signal of a historical grinding wheel grinding process and corresponding historical grinding wheel wear state information, wherein the historical grinding force signal is historical data of the grinding force signal of the grinding wheel grinding process, the historical vibration signal is historical data of the vibration signal of the grinding wheel grinding process, and the historical grinding wheel wear state information is historical data of the grinding wheel wear state information of the grinding wheel grinding process.

[0072] S202, performing denoising processing on the grinding force signal to obtain a denoised grinding force signal;

[0073] The embodiment of the application performs denoising processing on the grinding force signal through an existing filtering algorithm to obtain a denoised grinding force signal. Due to the complexity of the grinding process and the influence of the machining environment, the grinding force often presents low-quality high-noise and non-stationary characteristics, and contains a large amount of noise. Preferably, a wavelet threshold denoising method is used to perform denoising processing on the collected grinding force signal. The wavelet threshold denoising method has the advantages of simple algorithm and efficient operation, so the wavelet denoising processing is performed on the collected grinding force signal to improve the signal-to-noise ratio. The wavelet threshold denoising first selects a wavelet base function to perform multi-scale wavelet decomposition on the signal, then extracts useful signal wavelet coefficients under each scale, and finally reconstructs the useful signal by inverse wavelet transform to complete the signal denoising.

[0074] S203, performing symplectic geometric modal decomposition on the vibration signal to obtain a symplectic geometric modal component;

[0075] The embodiment of the application decomposes the vibration signal into a series of symplectic geometry components by a symplectic geometry mode decomposition method, extracts data about the grinding wheel wear characteristic from the symplectic geometry mode component, and obtains the symplectic geometry mode component. The weld joint grinding is a nonlinear and non-stationary machining process. Due to the weak stiffness of the robot, the grinding wheel spindle is more susceptible to interference from the machining process. Compared with the force signal, the vibration signal is more complex, which not only contains a large amount of noise and other irrelevant interference information, but also contains vibration modal components of different frequencies related to the grinding wheel wear state. Therefore, the vibration signal is not ideal only by noise reduction processing, and it is necessary to accurately and effectively decompose and extract the vibration characteristic component related to the grinding wheel wear state from the complex signal by a signal decomposition algorithm, so as to help the subsequent monitoring model to more accurately identify the wear degree of the grinding wheel. The symplectic geometry mode decomposition (SGMD) is a new non-stationary signal decomposition method, which can effectively decompose the original signal into a series of symplectic geometry components (SGCs) without pre-setting parameters. The SGMD method is used to decompose the vibration signal, and the SGCs obtained are subjected to frequency spectrum analysis to extract data about the grinding wheel wear characteristic, and the symplectic geometry mode component is obtained.

[0076] S204, splicing the denoised grinding force signal and the symplectic geometry mode component to obtain a characteristic component matrix;

[0077] The embodiment of the application needs to synchronize the time of the denoised grinding force signal and the symplectic geometry mode component, and then splice the time-synchronized denoised grinding force signal and the symplectic geometry mode component to obtain a characteristic component matrix. Time synchronization can ensure that the denoised grinding force signal and the symplectic geometry mode component at the same time are spliced, thereby avoiding data conflicts caused by time deviation.

[0078] S205, splicing the historical grinding force signal and the historical vibration signal of the historical grinding process of the grinding wheel to obtain a historical characteristic component matrix;

[0079] The embodiment of the application splices the historical grinding force signal and the historical vibration signal of the historical grinding process of the grinding wheel according to the splicing method to obtain a historical characteristic component matrix.

[0080] S206, constructing an improved CNN-GRU model based on the historical characteristic component matrix and the corresponding historical grinding wheel wear state information by using a whale optimization algorithm to obtain an optimal performance monitoring model;

[0081] In an optional embodiment, an improved CNN-GRU model is constructed based on the historical feature component matrix and corresponding historical grinding wheel wear state information by using a whale optimization algorithm, and a best performance monitoring model is obtained, including:

[0082] An improved CNN-GRU model corresponding to the historical feature component matrix is constructed, and a preliminary detection model is obtained.

[0083] The whale optimization algorithm is used to train the preliminary detection model according to the data for training in the historical feature component matrix and corresponding historical grinding wheel wear state information, and a trained preliminary detection model is obtained.

[0084] The trained preliminary detection model is verified according to the data not used for training in the historical feature component matrix, and a best performance monitoring model is obtained.

[0085] In an optional embodiment, an improved CNN-GRU model corresponding to the historical feature component matrix is constructed, and a preliminary detection model is obtained, including:

[0086] A batch normalization layer and a BiGRU unit composed of a double-layer bidirectional GRU structure are added to a preset CNN-GRU model, and a first CNN-GRU model is obtained.

[0087] A Dropout layer is arranged behind each BiGRU unit in each layer of the first CNN-GRU model, and a second CNN-GRU model is obtained.

[0088] A Self-Attention mechanism is introduced into the second CNN-GRU model to construct an improved CNN-GRU model corresponding to the historical feature component matrix, and a preliminary detection model is obtained.

[0089] The embodiment of the application is to comprehensively and effectively extract the hidden grinding wheel wear related features in the grinding process data, combine the advantages of CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network), and construct a grinding wheel wear state recognition model based on the hybrid of CNN and RNN. Based on the basic structure of the CNN-GRU (GRU is an improved recurrent neural network (RNN) variant) model, first, a batch normalization (BN) layer is added to the CNN to standardize the data distribution and enhance the model training efficiency and robustness; then, the existing GRU is improved to construct a BiGRU unit composed of a bidirectional GRU structure, and a BiGRU layer is added on the basis of a single BiGRU layer to mine deep-level time sequence features; in addition, in order to prevent overfitting, a Dropout layer is set after each BiGRU layer, which enhances the generalization ability of the model by randomly discarding the output of a part of neural units; finally, the idea of Self-Attention mechanism is introduced, so that the model can selectively learn the key information in the signal features, further improving the prediction accuracy of the model. The structure of the improved CNN-GRU model is specifically shown in Figure 3 . Figure 3 The architecture of the improved CNN-GRU model (ICNN-GRU model) is shown, which not only makes up for the shortcomings of CNN in representing time sequence information and long-term sequence dependence, but also solves the problem of gradient disappearance in the training of traditional RNN network.

[0090] The input of the improved CNN-GRU model is the fused vibration signal and grinding force signal. In order to eliminate the differences in different scales and dimensions caused by multi-sensor signals and improve the generalization ability of the model, the signal needs to be processed by Min-Max normalization before being input into the improved CNN-GRU model after being sliced. The expression is as follows:

[0091] ;

[0092] In the formula, x n is the normalized data; x is the input data; x max and x minmax and min of the input data set, respectively. After processing, different component data are normalized to the same scale range, eliminating the dimensional influence between different features and making the weight between different features more balanced.

[0093] The input sample is first processed by the CNN layer, and the local spatial features in the processing data are extracted by two convolution, max pooling and batch normalization operations. In the convolution and pooling process, the step is 2, and the zero padding method is used. In order to improve the convergence speed of the network and reduce the overfitting problem, the linear rectified function (Rectified Linear Unit, ReLU) with faster convergence speed is used as the nonlinear activation function in this embodiment. By adaptively extracting the local spatial features of the input data through CNN, not only the input parameters of the subsequent network are reduced, the calculation speed is improved, but also the dimension of the extracted feature matrix is reduced, which is also helpful to highlight the feature information related to the grinding wheel wear. After the Flatten layer processing, the extracted features are converted into one-dimensional sequence data to be input into the subsequent neural network for extraction of time-related features.

[0094] In order to fully capture the long-term dependence in the grinding process signal and learn the time correlation information in the grinding wheel wear process, a bidirectional GRU structure is constructed by improving the traditional GRU, so that the model can extract the time features of the signal from the forward and reverse directions at the same time. In addition, two layers of BiGRU are built in the model to further improve the learning ability of the model to the time features of the signal and ultimately improve the accuracy of the model. In order to prevent the model from overfitting, the Dropout rate after the two BiGRU layers is set to 0.3. The BiGRU structure can obtain the time step dependence information. Since the features of the previous time step contribute differently to the current grinding wheel wear state, the Self-Attention mechanism is added subsequently to optimize the time step weight, in order to improve the BiGRU structure. The Self-Attention mechanism can dynamically distribute the weights of the features contained in different time steps, determine the importance of different time steps to the grinding wheel wear state, so that the model can optimize the feature extraction effect, retain important information related to the grinding wheel wear, and improve the long-term memory ability of the model. Finally, after the full connection layer, it is input into the Softmax classification layer, and the output is the grinding wheel wear category. The number of units in the full connection layer is set to 3.

[0095] In summary, the improved CNN-GRU model corresponding to the historical feature component matrix is constructed, and the preliminary detection model is obtained.

[0096] Before the input data is trained on the model, the parameters of the preliminary detection model are initialized to ensure that the model is in a normal use state.

[0097] Preferably, in an optional embodiment, the parameters and hyperparameters of the preliminary detection model are adjusted based on the training error by a whale optimization algorithm to obtain optimal parameters and optimal hyperparameters, and the preliminary detection model is optimized using the optimal parameters and the optimal hyperparameters to obtain the trained preliminary detection model, comprising:

[0098] The parameters of the preliminary detection model are adjusted based on the training error to obtain optimal parameters, and the preliminary detection model is optimized using the optimal parameters to obtain a preliminary trained preliminary detection model.

[0099] The hyperparameters of the preliminary trained preliminary detection model are optimized by a whale optimization algorithm to obtain optimal hyperparameters.

[0100] The preliminary trained preliminary detection model is optimized using the optimal hyperparameters to obtain the trained preliminary detection model.

[0101] In the embodiment of the application, the data for training in the historical feature component matrix is input into the preliminary detection model to obtain the corresponding grinding wheel wear state output. By comparing the data for training in the historical feature component matrix in the database and the corresponding historical grinding wheel wear state information with the grinding wheel wear state output obtained by inputting the data for training in the historical feature component matrix into the preliminary detection model, the training error is obtained, such as the position deviation and wear amount deviation of the wear position in the historical grinding wheel wear state information and the grinding wheel wear state output. The training error is used to adjust the parameters of the preliminary detection model by combining the whale optimization algorithm. For example, if the training error is the position deviation of the wear position in the historical grinding wheel wear state information and the grinding wheel wear state output, the corresponding model parameters are modified to make the wear position of the grinding wheel wear state output more accurate, and the optimal parameters are obtained. The preliminary detection model is optimized by the optimal parameters to obtain the trained preliminary detection model.

[0102] Preferably, although the preliminary model structure is determined when the preliminary detection model is constructed, the value of the hyperparameters also has a significant impact on the performance of the model. Unlike the parameters of the model, the hyperparameters need to be manually set according to experience during the training process. The hyperparameters are usually closely related to the characteristics of the data, and appropriate hyperparameters can improve the accuracy and generalization ability of the model, while improper selection may lead to overfitting or underfitting. However, manually adjusting the hyperparameters requires a lot of manpower and time, and it is difficult to guarantee the best effect. Therefore, the whale optimization algorithm (WOA) is used to automatically adjust the hyperparameters of the preliminary detection model, so that the network model better matches the grinding wheel wear data characteristics, thereby improving the accuracy.

[0103] The principle of the whale optimization algorithm is as follows:

[0104] WOA simulates the behavior of searching for the optimal solution in nature. The algorithm mainly contains three stages: encircling prey, bubble-net attacking and searching for prey. Before using WOA to optimize the model parameters, the above three types of predatory behavior need to be mathematically modeled.

[0105] Encircling prey: humpback whales can identify the location of prey and encircle it. In the process of encircling, humpback whales will select the search agent position closest to the prey as the optimal position, and other humpback whales will gradually approach the optimal position, and finally form an encirclement. This approach behavior is represented by the following mathematical formula:

[0106] ; (1)

[0107] ; (2)

[0108] In the formula, D is the distance between the current whale individual and the optimal individual position; represents the position of the optimal whale individual in the current population; X(t) is the position vector of the current whale individual; is the current optimal solution position vector; t is the current iteration number; A is the first coefficient vector; C is the second coefficient vector. Wherein the first coefficient vector A and the second coefficient vector C The calculation formula is as follows:

[0109] ; (3)

[0110] ; (4)

[0111] In the formula, , t max is the maximum iteration number, a decreases linearly from 2 to 0 with the increase of iteration number; r is a random vector in [0, 1].

[0112] Bubble-net attacking method: humpback whales use spiral motion to advance when hunting prey, and gradually reduce the encirclement. Among them, humpback whales choose to shrink the encirclement with a probability of P i , and 1- Pi The probability of choosing spiral progression, the mathematical model of this synchronous selection can be expressed as follows:

[0113] ; (5)

[0114] wherein, P is a random number in [0, 1]; l is a random number in [-1, 1]; is the distance vector between the current whale individual and the optimal individual; b is a constant used to describe the spiral shape; P i Generally taken as 0.5. As the number of iterations increases, a will linearly decrease, while A fluctuates between -a , a ];When A belongs to [-1, 1], the whale attacks the prey, and the next position of the whale individual is any position between the current position and the prey.

[0115] Search for prey: In the search for prey, the search mechanism of the whale individual depends on the size of |A| . When |A| <1, the whale individual will update its position according to the optimal position, performing local search. When |A| >1, the whale individual will move away from the current optimal position and update its position according to a randomly selected whale position, performing global search. |A | >1, the mathematical model of global search is as follows:

[0116] ; (6)

[0117] ; (7)

[0118] wherein, X rand (t ) is the position vector of the randomly selected whale individual.

[0119] For the preliminary detection model, adjusting the learning rate, CNN convolution kernel size, and the number of GRU neurons, etc. three key hyperparameters, can have a significant impact on the performance of the model. When training a deep learning model, the learning rate controls the step size of each parameter update. The size of the convolution kernel determines the receptive field of the CNN layer. The number of GRU neurons determines the complexity and memory capacity of the model. In addition to the above three key hyperparameters, according to experience, the batch size during training is set to 64, and since the preliminary detection model has a small capacity, a total of 50 rounds of iteration are performed during training.

[0120] In an optional embodiment, the hyperparameters of the preliminary trained preliminary detection model are optimized by a whale optimization algorithm to obtain optimal hyperparameters, comprising:

[0121] Set the hyperparameters to be optimized and initialize the hyperparameters to be optimized;

[0122] The hyperparameters to be optimized are iterated and the error rate after each iteration is calculated by a preset iteration method and a preset error rate calculation formula;

[0123] The minimum value is extracted from the error rate after each iteration, and the hyperparameters to be optimized after iteration corresponding to the minimum value are determined as the optimal hyperparameters.

[0124] In the embodiment of the application, the convolution kernel size of the CNN m , the number of neurons of the GRU n and the initial learning rate η are set as the hyperparameters to be optimized, and the upper and lower bounds of the search range are set as the hyperparameters to be optimized. The hyperparameters to be optimized are initialized to determine the initial whale population position X ( m , n , η ). The hyperparameters to be optimized are iterated and the error rate after each iteration is calculated by a preset iteration method (as described in detail in the following step 4) and a preset error rate calculation formula (as described in detail in the following step 3). When the iteration number reaches the maximum iteration number T max , the iteration is stopped, the minimum value is extracted from the error rate after each iteration, and the hyperparameters to be optimized after iteration corresponding to the minimum value are determined as the optimal hyperparameters.

[0125] In the WOA optimization process, the prey position is assumed to be the optimal solution, and the position of each whale is regarded as a potential solution. In each iteration, the position update strategy of each whale is determined according to the value of a random number P and the modulus of the coefficient vector A . As the iteration proceeds, the whale population gradually approaches the optimal solution. The preliminary detection model training and optimization process is shown in Figure 4 , and the specific steps are as follows:

[0126] Step 1: Data set making and division. The data used for training in the historical feature component matrix is labeled according to different wear stages, and is divided into a training set (used for optimizing model parameters) and a test set (used for optimizing model hyperparameters) according to a certain proportion. Finally, the data samples are normalized and input into the preliminary detection model.

[0127] Step 2: WOA parameter initialization. The convolution kernel size of the CNN m , the number of GRU neurons n , and the initial learning rate η are taken as the parameters to be optimized (hyperparameters to be optimized), and the upper and lower bounds of the search range are set. The number of whale populations N , the maximum number of iterations T max , and the spatial dimension of the whale individual position are set, and the whale population position X ( m , n , η ) is initialized.

[0128] Step 3: Fitness evaluation. The network model is trained and tested, and then the fitness (objective function value) of each whale individual in the population is calculated. The position of the whale individual with the smallest fitness value is taken as the current optimal position and is recorded, i.e., the global optimal solution is determined. The error rate of the network model during testing is taken as the objective function of WOA, as shown in the following formula:

[0129] ; (8)

[0130] wherein Ture Num is the number of correctly classified samples in the test set; Total Num is the total number of samples in the test set; during the iteration process, WOA obtains the optimal set of hyperparameters by minimizing the objective function.

[0131] Step 4: Update individual position. A random number P is generated, and when P ≥ 0.5, the whale individual position is updated in a spiral advancing manner according to formula (5); when P < 0.5 and |A| ≥ 1, the whale individual position is updated in a shrinking surrounding manner according to formula (2); when P < 0.5 and |A| < 1, the whale individual position is updated in a global search manner according to formula (7).

[0132] Step 5: Iteration process judgment. Determine whether the termination condition (whether the number of iterations reaches the maximum number of iterations T max ) is met, if the termination iteration condition is met, output the optimal hyperparameters, otherwise return to step 3 for continuous execution until the termination condition is met.

[0133] Step 6: Save the model. Save the optimal hyperparameters, and use the optimal hyperparameters to optimize the preliminary detection model to obtain the trained preliminary detection model.

[0134] For example, as Figure 5 shown, Figure 5 is the optimal hyperparameter change curve when using whale optimization algorithm to optimize the preliminary detection model, where the abscissa is the iteration number, and the ordinate is the fitness value of the optimal hyperparameter. From Figure 5 it can be seen that in the first 4 iterations, the optimal hyperparameter change curve rapidly decreases, and after the 5th iteration, the optimal hyperparameter change curve slows down, and finally converges to the optimal solution when iterating to the 13th time, the optimal hyperparameter change curve reaches a steady state, at this time the target function value is 0.222, and the hyperparameter combination searched by WOA is the optimal hyperparameter combination. The three optimal hyperparameters of the preliminary detection model are m = 3 ，n = 128 and η = 0.005. The optimal hyperparameter combination found by the WOA algorithm helps the preliminary detection model to achieve better performance, improve the generalization ability and accuracy of the model.

[0135] The data not used for training in the historical feature component matrix is input into the trained preliminary detection model to obtain the corresponding grinding wheel wear state output, which is used as verification data. The error of the trained preliminary detection model is compared with the historical grinding wheel wear state information corresponding to the data not used for training to determine whether the error is within an acceptable range (the acceptable range is generally 0 to 3%, which can be modified according to actual needs), verify the accuracy of the trained preliminary detection model, and obtain the best performance monitoring model.

[0136] S207, input the feature component matrix into the best performance monitoring model to calculate the grinding wheel wear state information;

[0137] The feature component matrix is input into the best performance monitoring model in the embodiment of the application, and the grinding wheel wear state information is calculated. The grinding wheel wear state information includes wear position, wear size, and abrasive state, and other information for describing the grinding wheel wear state.

[0138] The grinding wheel wear state monitoring method disclosed in the embodiment of the application can solve the problem that the existing deep learning model-based grinding wheel state monitoring method is difficult to comprehensively extract hidden features in data, and lays an important foundation for the safe use of grinding wheels. At the same time, the method has strong operability, and the process of grinding wheel wear state monitoring is intuitive and clear in actual use, and can be easily applied to the grinding wheel state detection process.

[0139] Embodiment three

[0140] Please refer to Figure 6 , Figure 6A structural block diagram of an embodiment of a grinding wheel wear state monitoring device of the present application, the device comprising:

[0141] The acquisition module 301 is configured to acquire a grinding force signal and a vibration signal of a grinding wheel grinding process, and acquire a historical grinding wheel wear state training sample in a database; the historical grinding wheel wear state training sample comprises a historical grinding force signal and a historical vibration signal of a historical grinding wheel grinding process and corresponding historical grinding wheel wear state information;

[0142] The splicing module 302 is configured to splice the grinding force signal and the vibration signal to obtain a feature component matrix;

[0143] The construction module 303 is configured to construct an improved CNN-GRU model based on the historical grinding force signal and the historical vibration signal of the historical grinding wheel grinding process and the corresponding historical grinding wheel wear state information by using a whale optimization algorithm, to obtain an optimal performance monitoring model;

[0144] The calculation module 304 is configured to input the feature component matrix into the optimal performance monitoring model to calculate and obtain grinding wheel wear state information.

[0145] In an optional embodiment, the splicing module 302 comprises:

[0146] The denoising submodule is configured to perform denoising processing on the grinding force signal to obtain a denoised grinding force signal;

[0147] The decomposition submodule is configured to perform symplectic geometric modal decomposition on the vibration signal to obtain a symplectic geometric modal component;

[0148] The first splicing submodule is configured to splice the denoised grinding force signal and the symplectic geometric modal component to obtain a feature component matrix.

[0149] In an optional embodiment, the construction module 303 comprises:

[0150] The second splicing submodule is configured to splice the historical grinding force signal and the historical vibration signal of the historical grinding wheel grinding process to obtain a historical feature component matrix;

[0151] The construction submodule is configured to construct an improved CNN-GRU model based on the historical feature component matrix and the corresponding historical grinding wheel wear state information by using a whale optimization algorithm, to obtain an optimal performance monitoring model.

[0152] In an optional embodiment, the construction submodule comprises:

[0153] The construction unit is configured to construct an improved CNN-GRU model corresponding to the historical feature component matrix to obtain a preliminary detection model;

[0154] a training unit, configured to train the preliminary detection model by using a whale optimization algorithm based on data for training in the historical feature component matrix and corresponding historical grinding wheel wear state information, to obtain a trained preliminary detection model;

[0155] a verification unit, configured to verify the trained preliminary detection model based on data not for training in the historical feature component matrix, to obtain an optimal performance monitoring model.

[0156] In an optional embodiment, the constructing unit comprises:

[0157] an adding subunit, configured to add a batch normalization layer and a BiGRU unit composed of a double-layer bidirectional GRU structure in a preset CNN-GRU model, to obtain a first CNN-GRU model;

[0158] a setting subunit, configured to set a Dropout layer behind each BiGRU unit in the first CNN-GRU model, to obtain a second CNN-GRU model;

[0159] an introducing subunit, configured to introduce a Self-Attention mechanism in the second CNN-GRU model, to construct an improved CNN-GRU model corresponding to the historical feature component matrix, to obtain a preliminary detection model.

[0160] In an optional embodiment, the training unit comprises:

[0161] an inputting subunit, configured to input data for training in the historical feature component matrix into the preliminary detection model, to obtain corresponding grinding wheel wear state output;

[0162] a determining subunit, configured to determine a training error based on the data for training in the historical feature component matrix, corresponding historical grinding wheel wear state information and the corresponding grinding wheel wear state output;

[0163] an optimizing subunit, configured to adjust parameters and hyperparameters of the preliminary detection model based on the training error by using a whale optimization algorithm, to obtain optimal parameters and optimal hyperparameters, and to optimize the preliminary detection model by using the optimal parameters and the optimal hyperparameters, to obtain the trained preliminary detection model.

[0164] In an optional embodiment, the training unit further comprises:

[0165] an initializing subunit, configured to initialize parameters of the preliminary detection model.

[0166] In an optional embodiment, the optimizing subunit further comprises:

[0167] a first optimization component configured to adjust parameters of the preliminary detection model based on the training error to obtain optimal parameters, and optimize the preliminary detection model by using the optimal parameters to obtain a preliminary trained preliminary detection model;

[0168] a second optimization component configured to optimize hyperparameters of the preliminary trained preliminary detection model by using a whale optimization algorithm to obtain optimal hyperparameters;

[0169] a third optimization component configured to optimize the preliminary trained preliminary detection model by using the optimal hyperparameters to obtain the trained preliminary detection model.

[0170] In an optional embodiment, the optimization component further includes:

[0171] a setting sub-component configured to set to-be-optimized hyperparameters and initialize the to-be-optimized hyperparameters;

[0172] an iteration sub-component configured to perform iteration on the to-be-optimized hyperparameters and calculate an error rate after each iteration by using a preset iteration mode and a preset error rate calculation formula;

[0173] a determination sub-component configured to extract a minimum value from the error rate after each iteration, and determine the to-be-optimized hyperparameters after the iteration corresponding to the minimum value as optimal hyperparameters.

[0174] Those skilled in the art can clearly understand the specific working process of the system, device and unit described above for the convenience and brevity of description, which can refer to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0175] In several embodiments provided in the present application, it should be understood that the disclosed method and related products can be implemented by other ways. For example, the above-described device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0176] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment scheme.

[0177] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0178] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or the entire or part of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a readable storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned readable storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0179] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A method for monitoring the wear state of a grinding wheel, characterized in that: include: Acquire a grinding force signal and a vibration signal of a grinding wheel grinding process, and acquire a historical grinding wheel wear state training sample from a database; the historical grinding wheel wear state training sample includes a historical grinding force signal and a historical vibration signal of a historical grinding wheel grinding process and corresponding historical grinding wheel wear state information; splicing the grinding force signal and the vibration signal to obtain a characteristic component matrix; Using the whale optimization algorithm, based on the historical grinding force signal and historical vibration signal of the historical grinding wheel grinding process and the corresponding historical grinding wheel wear state information, an improved CNN-GRU model is constructed to obtain the optimal performance monitoring model; Inputting the characteristic component matrix into the optimal performance monitoring model to calculate the grinding wheel wear state information; The grinding force signal and the vibration signal are spliced ​​to obtain a characteristic component matrix, including: performing denoising processing on the grinding force signal to obtain a denoised grinding force signal; performing symplectic geometric modal decomposition on the vibration signal to obtain symplectic geometric modal components; splicing the denoised grinding force signal and the symplectic geometric modal component to obtain a characteristic component matrix; Using the whale optimization algorithm, based on the historical grinding force signal and historical vibration signal of the historical grinding wheel grinding process and the corresponding historical grinding wheel wear state information, an improved CNN-GRU model is constructed to obtain the optimal performance monitoring model, including: splicing the historical grinding force signal and the historical vibration signal of the historical grinding wheel grinding process to obtain a historical characteristic component matrix; Constructing an improved CNN-GRU model corresponding to the historical feature component matrix to obtain a preliminary detection model; Using a whale optimization algorithm, the preliminary detection model is trained according to the training data in the historical feature component matrix and the corresponding historical grinding wheel wear state information to obtain a trained preliminary detection model; Testing the trained preliminary detection model based on data not used for training in the historical feature component matrix to obtain an optimal performance monitoring model; Among them, an improved CNN-GRU model corresponding to the historical feature component matrix is ​​constructed to obtain a preliminary detection model, including: Add a batch normalization layer and a two-layer BiGRU unit consisting of a bidirectional GRU structure to the preset CNN-GRU model to obtain the first CNN-GRU model; A Dropout layer is respectively set after the BiGRU unit in each layer of the first CNN-GRU model to obtain a second CNN-GRU model; A Self-Attention mechanism is introduced into the second CNN-GRU model to construct an improved CNN-GRU model corresponding to the historical feature component matrix to obtain a preliminary detection model.

2. The method for monitoring the wear state of a grinding wheel according to claim 1, wherein: The preliminary detection model is trained using the whale optimization algorithm according to the training data in the historical feature component matrix and the corresponding historical grinding wheel wear state information to obtain a trained preliminary detection model, including: Inputting the data used for training in the historical feature component matrix into the preliminary detection model to obtain a corresponding grinding wheel wear state output; determining a training error based on the data used for training in the historical feature component matrix and the corresponding historical grinding wheel wear state information and the corresponding wheel wear state output; Through the whale optimization algorithm, based on the training error, the parameters and hyperparameters of the preliminary detection model are adjusted to obtain the optimal parameters and optimal hyperparameters, and the preliminary detection model is optimized using the optimal parameters and the optimal hyperparameters to obtain the trained preliminary detection model.

3. The method for monitoring the wear state of a grinding wheel according to claim 2, wherein: Before inputting the training data in the historical feature component matrix into the preliminary detection model to obtain the corresponding grinding wheel wear state output, the method further includes: Initialize the parameters of the preliminary detection model.

4. The method for monitoring the wear state of a grinding wheel according to claim 1, wherein: Adjusting the parameters and hyperparameters of the preliminary detection model based on the training error using a whale optimization algorithm to obtain optimal parameters and optimal hyperparameters, and optimizing the preliminary detection model using the optimal parameters and the optimal hyperparameters to obtain the trained preliminary detection model, including: Based on the training error, adjusting the parameters of the preliminary detection model to obtain optimal parameters, and optimizing the preliminary detection model using the optimal parameters to obtain a preliminary detection model after preliminary training; Optimizing the hyperparameters of the preliminary detection model after the preliminary training by using a whale optimization algorithm to obtain optimal hyperparameters; The preliminary detection model after the preliminary training is optimized using the optimal hyperparameters to obtain the trained preliminary detection model.

5. The method for monitoring the wear state of a grinding wheel according to claim 4, wherein: The hyperparameters of the preliminary detection model after the preliminary training are optimized by the whale optimization algorithm to obtain the optimal hyperparameters, including: Set the hyperparameters to be optimized and initialize the hyperparameters to be optimized; By using a preset iteration method and a preset error rate calculation formula, the hyperparameter to be optimized is iterated while calculating the error rate after each iteration; The minimum value is extracted from the error rate after each iteration, and the hyperparameter to be optimized after the iteration corresponding to the minimum value is determined as the optimal hyperparameter.

6. A grinding wheel wear state monitoring device, characterized in that: include: An acquisition module is used to acquire a grinding force signal and a vibration signal of a grinding wheel grinding process, and to acquire a historical grinding wheel wear state training sample from a database; the historical grinding wheel wear state training sample includes a historical grinding force signal and a historical vibration signal of a historical grinding wheel grinding process and corresponding historical grinding wheel wear state information; a splicing module, configured to splice the grinding force signal and the vibration signal to obtain a characteristic component matrix; A construction module is used to construct an improved CNN-GRU model based on the historical grinding force signal and the historical vibration signal of the historical grinding wheel grinding process and the corresponding historical grinding wheel wear state information using a whale optimization algorithm to obtain an optimal performance monitoring model; a calculation module, configured to input the characteristic component matrix into the optimal performance monitoring model to calculate the grinding wheel wear state information; The splicing module includes: a denoising submodule, configured to perform denoising processing on the grinding force signal to obtain a denoised grinding force signal; a decomposition submodule, configured to perform symplectic geometric mode decomposition on the vibration signal to obtain symplectic geometric mode components; A first splicing submodule is used to splice the denoised grinding force signal and the symplectic geometric modal component to obtain a characteristic component matrix; The building blocks include: A second splicing sub-submodule is used to splice the historical grinding force signal and the historical vibration signal of the historical grinding wheel grinding process to obtain a historical characteristic component matrix; A submodule is constructed for constructing an improved CNN-GRU model based on the historical feature component matrix and the corresponding historical grinding wheel wear state information using a whale optimization algorithm to obtain an optimal performance monitoring model; The building block includes: A construction unit, configured to construct an improved CNN-GRU model corresponding to the historical feature component matrix to obtain a preliminary detection model; A training unit is used to train the preliminary detection model using a whale optimization algorithm according to the training data in the historical feature component matrix and the corresponding historical grinding wheel wear state information to obtain a trained preliminary detection model; a testing unit, configured to test the trained preliminary detection model based on data not used for training in the historical feature component matrix to obtain an optimal performance monitoring model; The building blocks include: Add a subunit to add a batch normalization layer and a two-layer BiGRU unit consisting of a bidirectional GRU structure to the preset CNN-GRU model to obtain the first CNN-GRU model; Setting a subunit, for setting a Dropout layer after each BiGRU unit in the first CNN-GRU model, to obtain a second CNN-GRU model; A subunit is introduced to introduce a Self-Attention mechanism into the second CNN-GRU model to construct an improved CNN-GRU model corresponding to the historical feature component matrix to obtain a preliminary detection model.

Images