Grinding wheel online state monitoring system and method based on multi-sensor fusion

Abstract

The invention discloses a grinding wheel online state monitoring system and method based on multi-sensor fusion. The monitoring system is composed of an acquisition module, a data preprocessing module and a man-machine interaction module. Firstly, multi-source information collection is carried out through a sensor, then time-frequency domain analysis and wavelet transformation are carried out on the collected information in a targeted mode, and data feature values are extracted; according to the method, image characteristic values such as a Hilbert spectrum are obtained through Hilbert-Huang transform and fast Fourier transform, high-dimensional data are subjected to importance evaluation through a reelief algorithm to obtain multi-modal data with high correlation, a ResNet-DenseNet-LSTM hybrid network model is built, accurate monitoring of the online state of the grinding wheel is achieved through data driving, and the accuracy of the online state of the grinding wheel is improved. And finally, integrating a monitoring algorithm in an upper computer interface. According to the invention, sufficient monitoring precision and accuracy can be ensured, online monitoring is realized, and a complicated process of offline data processing is avoided.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent monitoring of the state of precision grinding tools, and particularly relates to an on-line state monitoring system and method for grinding wheels based on multi-sensor fusion. Background Art

[0002] In recent years, precision grinding technology has been widely applied in fields such as aerospace, mechanical electronics, and optics. During the grinding process, especially when processing materials with low thermal conductivity and high hardness, the phenomenon of grinding wheel wear is relatively serious. For profile grinding, the contour of the workpiece is "copied" from the topography of the grinding wheel, and the wear of the grinding wheel will seriously affect the shape accuracy of the workpiece. For conventional surface grinding, when the grinding wheel wears, the abrasive grains become dull, the cutting ability of the abrasive grains decreases, resulting in an increase in the surface roughness of the workpiece, and the grinding force will increase accordingly, leading to an increase in vibration during the machining process, seriously affecting the surface quality of the workpiece and increasing the generation of defective products. Therefore, it is very necessary to on-line monitor the state of the grinding wheel during the precision grinding process. Since the precision grinding processing area is often non-visual, there are the following difficulties in monitoring the wear state of the grinding wheel:

[0003] (1) The traditional method for detecting the state of the grinding wheel is generally off-line in-situ detection. This method not only increases the downtime of the machine tool and reduces the processing efficiency, but also has poor accuracy in manually controlling the downtime. Premature replacement of the grinding wheel will waste the service life of the grinding wheel.

[0004] (2) With the development of sensor technology, many scholars use sensors for monitoring. Due to the small amount of material removed in precision grinding, the signals are extremely weak, and coupled with the complex and variable environmental noises such as the spindle, frequency converter, and grinding fluid, it is difficult to achieve accurate monitoring by using a single sensor to monitor and analyze the common time-frequency characteristic values.

[0005] (3) The existing monitoring technology measures data through sensors and obtains the wear state of the grinding wheel after off-line analysis and processing of the data, which is difficult to ensure the timeliness of the data.

[0006] In summary, at present, it has become an urgent problem to accurately on-line monitor the wear state of the grinding wheel during the precision grinding process. Summary of the Invention

[0007] Aiming at the problems of difficult monitoring of the state of the grinding wheel during the precision grinding process, low efficiency, poor accuracy, and difficulty in realizing on-line real-time monitoring of the existing monitoring methods, the present invention provides an on-line state monitoring system and method for grinding wheels based on multi-sensor fusion, which can ensure sufficient monitoring accuracy and accuracy, and at the same time realize on-line monitoring, avoiding the cumbersome process of off-line data processing.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] An on-line state monitoring system for grinding wheels based on multi-sensor fusion, the system comprising: a collection module, a data preprocessing module and a human-computer interaction module;

[0010] The collection module consists of an acoustic emission sensor, a vibration sensor, a current sensor and a high-shielding cable; the data preprocessing module consists of a constant current source, a signal preamplifier, a current-voltage conversion module, a power supply and a signal acquisition card; the human-computer interaction module consists of a high-speed USB transmission cable, a digital input / output signal interface and a magnetic block; the monitoring system is integrated in a control box, the collection module is placed in the control box through a wire hole, the human-computer interaction module is magnetically attached to the outer shell of the box body, and the box body is designed with a porous heat dissipation plate and a system alarm lamp;

[0011] The collection module is used to collect multi-source information of the grinding wheel through sensors;

[0012] The data preprocessing module is used to preprocess the collected information to realize the on-line state monitoring of the grinding wheel;

[0013] The human-computer interaction module is used to integrate the monitoring results into the upper computer interface and communicate with the numerical control system of the machine tool.

[0014] Preferably, the acoustic emission sensor is a dual-channel acoustic emission sensor, which is respectively adsorbed on the spindle end and the workpiece end of the machine tool. The sensor adsorbed on the workpiece end is used as a switch to trigger the deep neural network; the vibration sensor and the current sensor are respectively placed on the spindle end of the grinding process and the three-phase electric coil of the spindle.

[0015] Preferably, the data preprocessing module includes: an extraction unit, a fusion unit, an evaluation unit and a monitoring unit;

[0016] The extraction unit is used to perform time-frequency domain analysis and wavelet transform on the collected information to extract the feature vectors of the data;

[0017] The fusion unit is used to fuse the time-frequency information of the feature vectors through Hilbert-Huang transform and fast Fourier transform to obtain the image data of different wear states of the grinding wheel;

[0018] The evaluation unit is used to evaluate the importance of the feature vectors and image data of the data by using the relief algorithm to obtain multi-modal data with strong correlation;

[0019] The monitoring unit is used to input the multi-modal data into a ResNet-DenseNet-LSTM hybrid network model to monitor the on-line state of the grinding wheel.

[0020] The present invention also provides an on-line state monitoring method for a grinding wheel based on multi-sensor fusion, which is implemented by using the aforementioned on-line state monitoring system for a grinding wheel based on multi-sensor fusion. The method includes the following steps:

[0021] Collect multi-source information of the grinding wheel through sensors;

[0022] Preprocess the collected information to monitor the on-line state of the grinding wheel;

[0023] Integrate the monitoring results into the upper computer interface and communicate with the machine tool numerical control system.

[0024] Preferably, the method for preprocessing the collected information to monitor the on-line state of the grinding wheel includes:

[0025] Perform time-frequency domain analysis and wavelet transform on the collected information to extract the feature vectors of the data;

[0026] Fuse the time-frequency information of the feature vectors through Hilbert-Huang transform and fast Fourier transform to obtain the image data of different wear states of the grinding wheel;

[0027] Use the relief algorithm to evaluate the importance of the feature vectors and image data of the data to obtain multi-modal data with strong correlation;

[0028] Build a ResNet-DenseNet-LSTM hybrid network model;

[0029] Input the multi-modal data into the ResNet-DenseNet-LSTM hybrid network model to monitor the on-line state of the grinding wheel.

[0030] Preferably, the method for performing time-frequency domain analysis and wavelet transform on the collected information to extract the feature vectors of the data includes:

[0031] Calculate the kurtosis and margin factor of the AE signal as time domain eigenvalues. The calculation formulas for kurtosis and margin factor CrestFactor are:

[0032]

[0033] where x i is the signal sampling point, μ is the signal mean, and N is the signal sampling length;

[0034] Calculate the spectral eigenvalues of the acceleration sensor and the current sensor:

[0035]

[0036] where X [k] is the result after Fourier transform, and x iis the original time-domain signal, N is the signal sampling length, and i and k are the number of operations;

[0037] Calculate the center frequency SC and the mean square frequency MSF as frequency-domain eigenvalues, and the calculation formulas are as follows:

[0038]

[0039] where, f k is the frequency value of the k-th frequency point, f s is the sampling rate, P k is the power spectrum value of the k-th frequency point, P k = |X [k] | 2 ;

[0040] Perform multi-resolution wavelet analysis on the three-channel signal, select the Haar wavelet as the wavelet basis function ψ(t), and decompose the signal into 8 levels of wavelet decomposition.

[0041]

[0042] Perform the inner product operation on the signal f(t) and the wavelet basis function ψ(t):

[0043]

[0044] where, x(t) is the original signal, t is the time variable, a and b are the scale and translation parameters, and ψ* is the conjugate wavelet;

[0045] Obtain eight groups of detail coefficients D1 - D8 and one group of approximation coefficient A8, and calculate the time-frequency eigenvalues and energy ratios of the coefficients of the three sensors respectively as data eigenvalues.

[0046] Preferably, the method for fusing the time-frequency information of the eigenvectors through Hilbert-Huang transform and fast Fourier transform to obtain the image data of different wear states of the grinding wheel includes:

[0047] Decompose the original signal x(t) into several intrinsic mode functions;

[0048] Identify all the maximum points e max (t) and minimum points e min (t) of the signal, and calculate the mean envelope m k (t);

[0049]

[0050] Update IMF

[0051] h k (t) = r k-1 (t) - mk (t)

[0052] Among them, h k (t) is the modal function to be selected, and r k (t) is the residual function;

[0053] If h k (t) satisfies the conditions of the IMF, then let IMF k = h k (t), and then update the remaining signal r k (t), otherwise continue to repeat the update of h k (t);

[0054] r k (t) = r k-1 (t) - IMF k (t)

[0055] Perform Hilbert transform z k (t) on each IMF component, and calculate the instantaneous amplitude A k (t) and the instantaneous frequency f k (t) of the signal;

[0056]

[0057] Among them, j is the imaginary unit, is the Hilbert-Huang transform of h k (t), and the calculation formula is:

[0058]

[0059] Among them, PV is the principal value integral, and τ is the time variable;

[0060] Combine the instantaneous amplitudes and frequencies of all IMFs into a time-frequency energy distribution Hilbert spectrum;

[0061]

[0062] Among them, δ(f - f k (t) is the Dirac δ function, f is the frequency variable, and f k (t) is the k-th frequency variable;

[0063] Integrate the Hilbert spectrum on the time axis to obtain the energy distribution in the frequency dimension and plot the marginal spectrum;

[0064]

[0065] Take the Hilbert spectra, marginal spectra, and time-frequency diagrams of the short-time Fourier transform of different grinding states of the grinding wheel as the picture eigenvalue data.

[0066] Preferably, the method for evaluating the importance of the feature vectors of data and image data by using the Relief algorithm to obtain multimodal data with strong correlation includes:

[0067] Evaluating the discrimination ability of features in neighboring samples through the Relief algorithm, and transforming the color feature picture into a three-dimensional tensor X ∈ R h×w×3 , and then flattening the three-dimensional tensor into a one-dimensional vector:

[0068]

[0069] X flat =[R 11 ,R 12 ...R hw ,G 11 ,G 12 ...G hw ,B 11 ,B 12 ...B hw

[0070] Among them, R hw ,G hw ,B hw represent the red, green, and blue channel values of the h-th row and w-th column respectively;

[0071] Randomly select a sample R from the dataset, find a nearest neighbor sample NH from the same category as R through Euclidean distance calculation, and then find the nearest neighbor sample NM from different categories of R. Successively traverse the weights of each eigenvalue through the weight update formula and update them:

[0072]

[0073] For each feature F, calculate the difference between the values of the instance R and NH on this feature using Euclidean distance. If the difference between R and NH is less than the preset threshold, it indicates that the importance of the feature F is lower than the preset threshold, and the weight of the feature F is weakened. The situation of R and NM is just the opposite.

[0074] Preferably, the method for monitoring the online state of the grinding wheel by inputting multimodal data into the ResNet-DenseNet-LSTM hybrid network model includes:

[0075] ​Process the image features, perform a series of convolutional, pooling, and normalization operations through the stem module to obtain a feature map of 75*75*64, and then enter the residual block module. Add a dense block behind the residual block. Dense connections enhance feature propagation and cross-layer information fusion. The output data of the dense block is a feature map of 75*75*192. Perform convolution and average pooling on the output data of the dense block through the transition layer to output a feature map of 37*37*96. Use the Flat module to expand the multi-dimensional features into a one-dimensional vector 96×37×37 = 131328; Combine the eigenvalue data extracted by the convolutional neural network with the time-frequency features of the original data and input them into the LSTM network together, and finally perform prediction through the fully connected layer and the output layer.

[0076] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0077] 1. Through multi-sensor data fusion processing, the present invention obtains multi-source information data, and applies the relief algorithm for data correlation analysis, making the model data set highly correlated, and the monitoring accuracy reaches more than 95%. It can effectively monitor the grinding wheel state online and avoid complex operations such as off-line monitoring.

[0078] 2. The present invention is applicable to different processing conditions. The system supports self-learning settings. First, self-learn and measure the multi-sensor data of the grinding wheel in different grinding states, start the network model for training, and then can accurately identify the wear state of the grinding wheel in this self-learning working condition, avoiding interference from different grinding parameters, etc., and having strong versatility.

[0079] 3. The present invention is highly integrated and easy to operate. It independently designs the hardware instrument box of the data storage module II and is easy to carry. When in use, first arrange the sensors in different processing areas, and opening the upper computer system can realize automatic online monitoring of processing.

[0080] 4. The present invention uses low-cost materials and simple designs, combined with artificial intelligence technology, not only reduces the production cost, but also has the advantage of being easy to mass-produce, and is applicable to industrial practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions of the present invention, the following briefly introduces the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0082] Figure 1 It is the hardware system design diagram in the embodiment of the present invention;

[0083] Figure 2 It is the overall architecture diagram in the embodiment of the present invention;

[0084] Figure 3 This is the signal time-frequency analysis diagram in the embodiment of the present invention;

[0085] Figure 4 This is the dimensionality reduction data diagram of the relief algorithm in the embodiment of the present invention;

[0086] Figure 5 This is the hierarchical diagram of the deep neural network in the embodiment of the present invention;

[0087] Figure 6 This is the specific structural diagram of the convolutional neural network of the present invention;

[0088] Figure 7 This is the online recognition result diagram of the grinding wheel wear state in the embodiment of the present invention;

[0089] Figure 8 This is the schematic diagram of the human-computer interaction interface of the grinding wheel wear online monitoring system in the embodiment of the present invention.

[0090] Note: Ⅰ - Data acquisition module; Ⅱ - Data preprocessing module; Ⅲ - Human-computer interaction module; 1 - Acoustic emission (AE) sensor; 2 - Vibration sensor; 3 - Current sensor; 4 - Threading hole; 5 - Acquisition card junction box; 6 - Constant current source; 7 - Signal preamplifier; 8 - Current-voltage conversion module; 9 - Power supply; 10 - Signal acquisition card; 11 - Monitoring system alarm device; 12 - Transmission cable; 13 - Usb high-speed transmission cable; 14 - Digital input / output signal interface; 15 - Magnetic block. Detailed implementation manners

[0091] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0092] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0093] Embodiment 1

[0094] Combined with Figure 1-8 , the present invention proposes an online state monitoring system for grinding wheels based on multi-sensor fusion, and the monitoring system is composed of a data acquisition module Ⅰ, a data preprocessing module Ⅱ, and a human-computer interaction module Ⅲ.

[0095] The acquisition module Ⅰ is used to collect multi-source information of the grinding wheel through sensors;

[0096] The data preprocessing module II is used to preprocess the collected information to monitor the on-line state of the grinding wheel;

[0097] The human-computer interaction module III is used to integrate the monitoring results into the host computer interface and communicate with the machine tool numerical control system.

[0098] The data acquisition module I is specifically composed of an acoustic emission (AE) sensor 1, a vibration sensor 2, a current sensor 3 and a high-shielding cable; the data preprocessing module II is specifically composed of a constant current source 6, a signal preamplifier 7, a current-voltage conversion module 8, a power supply 9 and a signal acquisition card 10; the human-computer interaction module III is specifically composed of a usb high-speed transmission cable 13, a digital input / output signal interface 14 and a magnetic block 15; the data preprocessing module II collects multi-sensor information in real time at high speed and transmits it to the host computer through a transmission cable 12. The host computer is provided with a digital input / output signal interface 14 for communicating with the machine tool numerical control system. This monitoring system is highly integrated in a control box of 300*200*420mm. The data acquisition module 1 can be placed in the control box through a wire passing hole 4. The human-computer interaction module III can be magnetically adsorbed on the outer shell of the box body. The box body is designed with a porous heat dissipation plate partition and a monitoring system alarm device 11.

[0099] Furthermore, the sensor layout method is crucial for obtaining highly sensitive monitoring signals. The AE sensor 1 can collect high-frequency elastic wave signals generated by workpiece materials under different stress actions after the grinding wheel wears. Therefore, a dual-channel AE sensor is selected and adsorbed on the spindle end and the workpiece end of the machine tool respectively. The sensor adsorbed on the workpiece end responds quickly and can be used as a switch to trigger the deep neural network. When the grinding wheel wears, the spindle current and vibration will change due to the increase in grinding force. Therefore, the vibration sensor 2 and the current sensor 3 are respectively placed on the spindle end of the grinding process and the three-phase electric coil of the spindle. Through multi-channel signal acquisition, the grinding process information can be comprehensively covered.

[0100] Furthermore, this monitoring system first obtains multi-source information on the grinding process through acoustic emission, vibration, and current sensors. The data preprocessing module includes: an extraction unit, a fusion unit, an evaluation unit, and a monitoring unit;

[0101] The extraction unit is used to perform time-frequency domain analysis and wavelet transform on the collected information to extract the feature vectors of the data;

[0102] The fusion unit is used to fuse the time-frequency information of the feature vectors through Hilbert-Huang transform and fast Fourier transform to obtain the image data of different wear states of the grinding wheel;

[0103] An evaluation unit for using the relief algorithm to evaluate the importance of the feature vectors of data and image data to obtain multimodal data with strong correlation;

[0104] A monitoring unit for inputting multimodal data into a ResNet-DenseNet-LSTM hybrid network model to monitor the on-line state of the grinding wheel.

[0105] Specifically, since the grinding wheel wear signal is mainly reflected in the high-frequency components of acoustic emission, in order to obtain the high-frequency characteristics of the signal, the system sampling rate is set to 2M / s, and an adaptive adjustment program is written in the signal acquisition module to obtain the original acquisition data Among them, Is the total acquisition data of 6 channels, Is the AE signal at the workpiece end, Is the AE signal at the shaft end, Is the vibration signal, Is the current signal.

[0106] Through multimodal signal processing methods, according to the characteristics of different signals, corresponding signal processing methods are respectively used to obtain the feature vectors of the data Among them, Is the total eigenvalue of multi-dimensional data, a1, a2... a n Respectively represent the kurtosis, mean square frequency, wavelet energy ratio and other eigenvalue of acoustic emission and vibration signals.

[0107] Using Hilbert-Huang transform and short-time Fourier transform to fuse the time-frequency characteristics of the original signal to obtain the image matrix of different wear states That is, the image data, through the relief algorithm for And For screening, extracting information with strong correlation. Among them, r, g, and b refer to the red, green, and blue channels in the image respectively.

[0108] Process the image features, and perform a series of convolution, pooling and normalization processes through the stem module to obtain a feature map of 75*75*64. The calculation formula for the output feature map and its size of the convolutional layer is as follows:

[0109]

[0110] Among them, Y i,j,k Represents the output feature map of the convolutional layer, W m,n,k Is the weight of the kth convolutional kernel, b k Is the bias of the kth convolutional kernel, p represents the input channel being processed currently, and i, j, m, n, k are the index values of the feature map and the convolutional kernel respectively.

[0111]

[0112] Among them, H out and H in are the output and input feature map sizes respectively, q is the zero-padding size, f is the convolutional kernel and pooling kernel sizes, and s is the stride.

[0113] Next, activate the convolutional data Y i,j,k to obtain Z i,j,k . Select the ReLU activation function, and its expression is as follows:

[0114] Z i,j,k = ReLU(Y i,j,k ), ReLU(x) = max(0, x)

[0115] Perform average pooling on the activated function to obtain the pooled feature map data A i,j,k :

[0116]

[0117] Among them, P is the size of the pooling window, and S is the pooling stride.

[0118] Then enter the second-layer residual network module, add a dense block behind the residual block. The residual network adds skip connections in the convolution. The third-layer dense block uses the method of feature reuse. The l-th layer network performs convolution operations on the data of the previous l - 1 layers, and then splices the outputs together. After passing through the residual dense block, a feature map with a size of 75 * 75 * 192 is output. The output calculation formula of the convolutional layer of the residual network is as follows:

[0119]

[0120] Among them, Y i,j,k 2 represents the output feature map of the second convolutional layer, and the definitions of the remaining parameters are the same as above.

[0121] The pooling layer and activation layer are consistent with the above formula. Through the calculation of the second-layer residual network, A i,j,k 2 is obtained and enters the third-layer network for dense connection. The convolution formula of the dense layer is as follows:

[0122]

[0123] Among them, Y 3(l) i,j,k represents the output of the l-th layer of the dense network, including all the feature values of the previous l - 1 layers, and is used as the input of the l + 1 layer.

[0124] Pass the output data A 3 i,j,kPerform convolution and average pooling, and output a feature map of 37*37*96. Use the Flat module to expand the multi-dimensional features into a one-dimensional vector. The size is 96×37×37 = 131328; the eigenvalue data extracted by the convolutional neural network and the time-frequency eigenvalues of the original data are jointly input into the LSTM network, and finally the output layer is predicted through the fully connected layer.

[0125] Among them, is the one-dimensional vector after flattening the image data after passing through the convolutional neural network, is the summary of the data eigenvalues obtained by performing time-frequency analysis on the data.

[0126] The LSTM network architecture consists of three gating mechanisms, all of which output weights from 0 to 1 through the Sigmoid function. The forget gate determines which information to discard, the input gate determines which information to update, and the output gate determines which information to output. Their calculation formulas are as follows:

[0127] f t = σ(W f ·[h t-1 , x t +b t )

[0128] i t = σ(W i ·[h t-1 , x t +b i )

[0129] o t = σ(W o ·[h t-1 , x t +b o )

[0130] Among them, f t , i t , o t respectively refer to the outputs of the three gates. W and b represent the weight matrix and the bias matrix, and the rest are index values.

[0131] Furthermore, an adaptive adjustment program is written for the sensor. To solve the problem of signal distortion caused by the signal exceeding the range during the acquisition process, the signal gain is dynamically adjusted to avoid signal saturation or being too weak. Its gain coefficient k is calculated as follows:

[0132]

[0133] Among them, α is the attenuation coefficient, taking 0.5. When the sensor output value x is large / small, the gain G decreases / increases accordingly.

[0134] Furthermore, an adaptive calibration program is designed for the sensor. To solve problems such as the acquired signal exceeding the range, threshold settings for parameters such as signal amplitude and frequency are added to the acquisition program. The sensor parameters are automatically adjusted according to different machining environment changes. When the overall amplitude of the acquired signal is higher than the upper limit of the sensor or the signal is overall weak, the program will automatically lower or raise the amplification factor to prevent signal distortion and other situations during the multi-sensor acquisition process.

[0135] Furthermore, the data transmission module II is integrated in the control box, mainly responsible for preliminarily amplifying and filtering the multi-sensor signals and transmitting the signals to the upper computer through the acquisition card. To promptly detect system paralysis caused by acquisition card failures and power supply failures and prevent losses to machining monitoring, parallel alarm devices are set at the input end and power supply end of the acquisition card. When any channel acquisition signal in the multi-channel is abnormal or the supply current is abnormal, the alarm device 11 of the monitoring system lights up a red light and is accompanied by an alarm bell ringing to prompt the operator of the abnormality.

[0136] Furthermore, the human-computer interaction interface III incorporates data processing algorithms and a deep learning monitoring model. The upper computer interface communicates with the sensors of the lower computer to read the multi-sensor grinding signals in real time. At the same time, the interface communicates with the machine tool numerical control system through the RS-232 communication protocol. Once the program detects severe wear of the grinding wheel, a warning signal is promptly sent to trigger the machine tool to stop the machining program. Real-time monitoring of the online state of the grinding wheel is beneficial to improving the machining accuracy of the workpiece and reducing the generation of scrap rate.

[0137] Furthermore, to improve the response efficiency of the system and save calculation time, a network trigger switch is set before the deep learning monitoring network. Specifically, a signal threshold is set, which is slightly less than the severe wear threshold of the grinding wheel. The sensor signals are collected and intercepted in real time, and the threshold percentage of the signal amplitude deviation is calculated. When the threshold percentage exceeds 30%, the system automatically activates the deep learning network for further data processing and analysis for accurate grinding wheel state monitoring.

[0138] Specifically, to improve the response rate of the system and save calculation time as described above, a network trigger switch is set before the deep learning monitoring network, and the acoustic emission vector at the workpiece end Specifically includes (x 11 , x 12 ... x 1n ), the signal threshold is set as P n , and the total number of signal points is N.

[0139]

[0140] When the threshold percentage exceeds 30%, the trigger switch system automatically activates the deep learning network for further data processing and analysis to perform precise grinding wheel status monitoring.

[0141] Furthermore, the online monitoring system uses a combination of edge computing and cloud computing for prediction. First, the system processes and analyzes the real-time data through edge computing to monitor the grinding wheel status. Then, the information of the grinding wheel wear data is uploaded to the cloud platform database. The cloud platform has powerful computing and storage capabilities. The historical wear data in the cloud can update and optimize the machine learning model. By comparing and training the historical data with the new real-time data, the model will continuously adjust its parameters, thereby improving the recognition accuracy of the grinding wheel wear.

[0142] Specifically, the online monitoring system uses a combination of edge computing and cloud computing for prediction. First, the system processes and analyzes the real-time wear data to obtain the grinding wheel wear status

[0143]

[0144] Then, the data of each grinding wheel wear is uploaded to the database, and the database includes multiple sets of grinding wheel wear data The database data is used to perform secondary training on the model, continuously adjust the model parameters, and improve the prediction accuracy and generalization ability of the model.

[0145] Embodiment 2

[0146] The present invention proposes an on-line status monitoring method for a grinding wheel based on multi-sensor fusion, which is specifically carried out according to the following steps:

[0147] Step 1: The system is divided into the design of a self-learning module and the design of an on-line monitoring module. The present invention is applicable to the on-line monitoring of the grinding wheel status under different machining conditions. Before using a new device and after changing the machining parameters each time, a self-learning operation needs to be carried out first. First, supervised learning is required. The signals of different wear states of the grinding wheel are monitored through multiple sensors to obtain multi-sensor data for subsequent data processing and eigenvalue extraction.

[0148] Step 2: Analyze the internal relationship between the data eigenvalues of the self-learning in different grinding wheel wear stages and the grinding wheel wear. Since there are differences in the sensitive information captured by different sensor signals, it is necessary to analyze the time-domain and frequency-domain eigenvalues specifically. In addition, wavelet transform is introduced for multi-scale analysis to refine the frequency-domain interval, extract the subtle changes and long-term trends of the signal, and calculate the proportion of different wavelet energies as the eigenvalue vector.

[0149] Step 3: Analyze the internal relationship between the self - learned image feature values in different grinding wheel wear stages described in Step 1 and the grinding wheel wear. Single - data features are often easily affected by physical parameters such as spindle speed, feed rate, and grinding depth, and cannot deeply explore the complexity of wear information. Introducing image information such as Hilbert spectrum, marginal spectrum, and time - frequency spectrum can achieve multi - modal data fusion, making the model monitoring more accurate and more robust in dealing with complex problems.

[0150] Step 4: The time - frequency data analysis and processing above can form a data set. Set the sample space number of each group of data to 2000. The data information of three sensors contains 54 feature values, and the picture information contains 9 feature values. To reduce the size of the sample capacity and improve the speed and accuracy of the monitoring model training, the key features of the data are screened by the Relief algorithm for data dimensionality reduction.

[0151] Step 5: Build an online monitoring model for grinding wheel wear with a long - depth fusion neural network ResNet - DenseNet - LSTM classification. Introduce the residual network (ResNet), dense connection network (DenseNet), and long - short - term memory neural network. First, process the grinding wheel wear picture feature values through the first two networks, and then expand the data obtained from multiple convolutional pooling through the fully - connected layer and merge it with the original time - frequency data feature values and import them into the LSTM network for training. LSTM can effectively learn the pattern of wear changing with time, thereby achieving precise monitoring of grinding wheel wear.

[0152] Step 6: Use LABVIEW software to collect data in real - time. The processes described in the above Steps 1 to 5 complete operations such as data collection, processing, feature value extraction, and training of the deep neural network. Write this process into a set of data - driven processing algorithms using MATLAB, then integrate it into the LABVIEW software, and finally import the program into the upper computer. Next, collect grinding data in real - time through sensors, truncate the data stream every 2 seconds for threshold statistics. When the threshold exceeds the set range, the operation switch of the deep neural network is triggered to precisely monitor the online state of the grinding wheel.

[0153] Further, Step 1 is specifically as follows: First, artificially and supervisedly divide different wear stages of the grinding wheel according to processing requirements, and use a three - channel sensor to collect grinding signals in different stages. According to the Nyquist sampling theorem, the sampling rate should be higher than twice the highest frequency of the signal. Therefore, set the sampling rate of the acquisition card to 2M / s and save the data to the system through the acquisition card.

[0154] Further, Step 2 is specifically as follows: The time domain of the AE signal is sensitive to wear information. Therefore, calculate the kurtosis, crest factor, etc. as time - domain feature values. The calculation formulas for kurtosis and crest factor are:

[0155]

[0156] Where: x i is the signal sampling point, μ is the signal mean value, and N is the signal sampling length.

[0157] The time-domain signals of the acceleration sensor and the current sensor are not sensitive, and their frequency spectrum analysis is carried out:

[0158]

[0159] Where: X [k] is the result after Fourier transform, x i is the original time-domain signal, N is the signal sampling length, i and k are the number of operations.

[0160] Calculate the center frequency (SC) and the mean square frequency (MSF) as the frequency-domain eigenvalues, and their calculation formulas are as follows:

[0161]

[0162] Where: f i is the frequency value of the i-th frequency point, where f s is the sampling rate; P i is the power spectrum value of the i-th frequency point, P i = |X [i] | 2 . (Based on the FFT result)

[0163] Since the grinding wheel wear signal belongs to a non-linear and non-stationary signal, and the sensitive frequency band ranges of different sensors are different, so multi-resolution continuous wavelet analysis is carried out on the three-channel signals. The Haar wavelet is selected as the wavelet basis function ψ(t) to decompose the signal into 8 levels of wavelet decomposition.

[0164]

[0165] Convert the signal x i into a function x(t) with time t as the independent variable and perform an inner product operation with the wavelet basis function ψ(t):

[0166]

[0167] Among them, x(t) is the original signal, t is the time variable, a and b are the scale and translation parameters, and ψ* is the conjugate wavelet. Eight groups of detail coefficients D1-D8 and a group of approximation coefficients A8 are obtained through variation. The above time-frequency eigenvalue calculation and energy ratio calculation are respectively carried out on the detail coefficients of the three sensors as eigenvalues.

[0168] Further, the specific operation of Step 3 is to perform Hilbert-Huang transform on the signal in two-dimensional scale to obtain the Hilbert spectrum of different wear states of the grinding wheel. First, the original signal x(t) is decomposed into a number of intrinsic mode functions.

[0169] Identify all the maximum points e max (t) and minimum points e min (t) of the signal, and calculate the mean envelope m k (t) of the signal.

[0170]

[0171] Update IMF

[0172] h k (t) = r k-1 (t) - m k (t)

[0173] where h k (t) is the candidate mode function, and r k (t) is the residual function.

[0174] If h k (t) meets the conditions of IMF, then let IMF k = h k (t), and then update the remaining signal r k (t). Otherwise, continue to repeat the above steps to update h k (t).

[0175] r k (t) = r k-1 (t) - IMF k (t)

[0176] Perform Hilbert transform on each IMF component to obtain z k (t), and calculate the instantaneous amplitude A k (t) and instantaneous frequency f k (t) of the signal.

[0177]

[0178] where j is the imaginary unit, is the Hilbert-Huang transform of h k (t), and its calculation formula is as follows:

[0179]

[0180] where PV is the principal value integral and τ is the time variable.

[0181] Combine all the instantaneous amplitudes and frequencies of the IMF into the time-frequency energy distribution Hilbert spectrum H(f,t).

[0182]

[0183] where δ(f - f k (t) is the Dirac δ function, f is the frequency variable, and f k (t) is the k th frequency variable.

[0184] Integrate the Hilbert spectrum along the time axis to obtain the energy distribution in the frequency dimension, and plot the marginal spectrum M(f).

[0185]

[0186] Take the Hilbert spectrum, marginal spectrum, and time-frequency diagram of the short-time Fourier transform of different grinding states of the grinding wheel as image matrix data.

[0187] Furthermore, the specific content of Step 4 is to evaluate the discrimination ability of features in neighboring samples through the Relief algorithm. The main method is to convert the color feature image into a three-dimensional tensor X ∈ R h×w×3 , and then flatten it into a one-dimensional vector X flat .

[0188] where

[0189]

[0190] X flat = [R 11 , G 11 , B 11 , R 12 , G 12 , B 12 ,..., R hw , G hw , B hw

[0191] where: R hw , G hw , B hw represent the red, green, and blue channel values of the h-th row and w-th column respectively.

[0192] Randomly select a sample R from the dataset, find a nearest neighbor sample NH of the same category as R through the Euclidean distance calculation formula, and then find the nearest neighbor sample NM of a different category from R. Update the weights of each eigenvalue one by one through the weight update formula.

[0193]

[0194] ​For each feature F, the Euclidean distance is used to calculate the difference in the values of the instance R and NH on this feature. If the difference between R and NH is small, it indicates that the importance of feature F is low, and the weight of feature F is weakened. The situation of R and NM is just the opposite.

[0195] Further, the specific content of step five is as follows. The picture features are simplified to 5 through the relief algorithm, and the data features are simplified to 30. First, the picture features are processed. Through the stem module, a series of convolutional, pooling, and normalization processes are carried out to obtain a feature map of 75*75*64 for the subsequent model. Then, it enters the residual block module. Residual connections can solve the problem of gradient disappearance in deep networks. A dense block is added behind the residual block. Dense connections enhance feature propagation and cross-layer information fusion. The output data of the dense block is a feature map of 75*75*192. Through the transition layer, convolution and average pooling are performed on the data of the previous module to reduce the amount of calculation and prevent overfitting caused by the subsequent network being too deep, and the output is 37*37*96. The Flat module is used to expand the multi-dimensional features into a one-dimensional vector 96×37×37 = 131328. The eigenvalue data extracted by the convolutional neural network is combined with the time-frequency features of the original data and jointly input into the LSTM network. Finally, prediction is carried out through the fully connected layer and the output layer. The specific structural steps of the LSTM module are as follows:

[0196] (1) Input layer, integrating the and inputs into a one-dimensional tensor X' = {x'1, x'2... x' n}

[0197] (2) Through the basic unit of LSTM, the gating mechanism is used to retain important information and process the sequence data step by step. Specifically, it includes steps such as the forget gate, input gate, candidate state, state update, output gate, and hidden state output. The calculation formulas are as follows:

[0198] (3) Dropout prevents overfitting by randomly discarding a part of the neurons and maintains the generalization ability of the model. The Dropout of this system model is set to 0.3.

[0199] (4) The fully connected layer maps the output of LSTM to the target space, using the following formula:

[0200]

[0201] where: W k and W i belong to the weight matrix, b k and b i belong to the bias matrix, and h is the intermediate hidden vector.

[0202] The Adam optimizer is selected, and the learning rate is set to 0.001. The categorical cross-entropy function is used as the loss function to measure the difference between the probability distribution predicted by the model and the true distribution. In the multi-classification problem of the grinding wheel wear state, the categorical cross-entropy can effectively evaluate the prediction accuracy of the model for each wear state, thereby guiding the parameter adjustment and performance improvement of the model.

[0203] Further, the specific content of step six is that during the monitoring process, since there are obvious changes in the acoustic emission time-domain signal after the tool is successfully aligned with the workpiece, the time-domain threshold of the acoustic emission signal is selected as the trigger switch of the deep neural network. The system refreshes the data every 2 seconds, and sets the intensity threshold P of the acoustic emission signal of the grinding wheel wear n , P n should be less than the signal intensity of severe grinding wheel wear. Each time the data is refreshed, calculate the percentage of the current signal S exceeding the threshold. If the intensity of the current signal reaches more than 30% of the set threshold, the switch triggers the deep neural network to work, and accurately monitors the state of the grinding wheel. When it is monitored that the grinding wheel is not in a severely worn state, communicate with the machine tool numerical control system through RS232 to trigger the numerical control system to stop the machining program.

[0204] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An on-line state monitoring system for grinding wheels based on multi-sensor fusion, characterized in that, The system includes: a collection module, a data preprocessing module, and a human-machine interaction module; The collection module consists of an acoustic emission sensor, a vibration sensor, a current sensor, and a high-shielding cable; the data preprocessing module consists of a constant current source, a signal preamplifier, a current-voltage conversion module, a power supply, and a signal acquisition card; the human-machine interaction module consists of a high-speed USB transmission cable, a digital input / output signal interface, and a magnetic block; the monitoring system is integrated in a control box, the collection module is placed in the control box through a threading hole, the human-machine interaction module is magnetically attached to the outer shell of the box, and the box is designed with a porous heat dissipation plate and a system alarm light; The collection module is used to collect multi-source information of the grinding wheel through sensors; The data preprocessing module is used to preprocess the collected information to monitor the online state of the grinding wheel; The human-machine interaction module is used to integrate the monitoring results into the host computer interface and communicate with the machine tool numerical control system.

2. The system according to claim 1, wherein The acoustic emission sensor selects a dual-channel acoustic emission sensor, which is respectively adsorbed on the spindle end and the workpiece end of the machine tool. The sensor adsorbed on the workpiece end is used as a switch to trigger the deep neural network; the vibration sensor and the current sensor are respectively placed on the spindle end of the grinding process and the three-phase electric coil of the spindle.

3. The system according to claim 1, wherein The data preprocessing module includes: an extraction unit, a fusion unit, an evaluation unit, and a monitoring unit; The extraction unit is used to perform time-frequency domain analysis and wavelet transform on the collected information to extract the feature vectors of the data; The fusion unit is used to fuse the time-frequency information of the feature vectors through Hilbert-Huang transform and fast Fourier transform to obtain image data of different wear states of the grinding wheel; The evaluation unit is used to use the relief algorithm to evaluate the importance of the feature vectors and image data of the data to obtain strongly correlated multi-modal data; The monitoring unit is used to input the multi-modal data into a ResNet-DenseNet-LSTM hybrid network model to monitor the online state of the grinding wheel.

4. An on-line state monitoring method for grinding wheels based on multi-sensor fusion, characterized in that, Implemented by using the on-line state monitoring system of the grinding wheel based on multi-sensor fusion according to any one of claims 1-3, wherein the method includes the following steps: Collect multi-source information of the grinding wheel through sensors; Preprocess the collected information to monitor the online state of the grinding wheel; Integrate the monitoring results into the host computer interface and communicate with the machine tool numerical control system.

5. The method according to claim 4, wherein The method for preprocessing the collected information to monitor the online state of the grinding wheel includes: Perform time-frequency domain analysis and wavelet transform on the collected information to extract the feature vectors of the data; Fuse the time-frequency information of the feature vectors through Hilbert-Huang transform and fast Fourier transform to obtain image data of different wear states of the grinding wheel; Use the relief algorithm to evaluate the importance of the feature vectors and image data of the data to obtain strongly correlated multi-modal data; Build a ResNet-DenseNet-LSTM hybrid network model; Input the multi-modal data into a ResNet-DenseNet-LSTM hybrid network model to monitor the online state of the grinding wheel.

6. The method according to claim 5, wherein The method for performing time-frequency domain analysis and wavelet transform on the collected information and extracting the feature vectors of the data includes: Calculating the kurtosis and crest factor of the AE signal as time-domain eigenvalues. The calculation formulas for kurtosis and crest factor are: where x i is the signal sampling point, μ is the signal mean, and N is the signal sampling length; Calculating the spectral eigenvalues of the acceleration sensor and the current sensor: Where: X [k] is the result after Fourier transform, x i is the original time-domain signal, N is the signal sampling length, and i and k are the number of operations; Calculating the center frequency SC and the mean square frequency MSF as frequency-domain eigenvalues. The calculation formulas are: Among them, f k is the frequency value at the k-th frequency point, f s is the sampling rate, P k is the power spectrum value at the k-th frequency point, P k = |X [k] | 2 ; Performing multi-resolution wavelet analysis on the three-channel signal, selecting the Haar wavelet as the wavelet basis function ψ(t), and decomposing the signal into 8 levels of wavelet decomposition. Performing an inner product operation on the signal f(t) and the wavelet basis function ψ(t): where x(t) is the original signal, t is the time variable, a and b are the scale and translation parameters, and ψ* is the conjugate wavelet; Obtaining eight groups of detail coefficients D1 - D8 and one group of approximation coefficients A8, and respectively calculating the time-frequency eigenvalues and the energy proportion of the coefficients of the three sensors as the data eigenvalues.

7. The method according to claim 6, wherein The method for fusing the time-frequency information of the feature vectors through Hilbert-Huang transform and fast Fourier transform to obtain the image data of different wear states of the grinding wheel includes: Decomposing the original signal x(t) into several intrinsic mode functions; Identify all the maximum points e max (t) and minimum points e min (t) of the identification signal, and calculate the mean envelope m k (t); Updating the IMF h k r(t) = k-1 m(t) - k t where h k (t) is the candidate mode function, and r k (t) is the residual function; If h k (t) satisfies the conditions of the IMF, then let IMF k = h k (t), and then update the remaining signal r k (t), otherwise continue to repeatedly update h k (t); r k r(t) = k-1 r(t) - IMF k r(t) Perform the Hilbert transform \(z(t)\) on each IMF component and calculate the instantaneous amplitude \(A(t)\) and instantaneous frequency \(f(t)\) of the signal; k (t), and calculate the instantaneous amplitude A k (t) and the instantaneous frequency f k (t); where j is the imaginary unit, is the Hilbert-Huang transform of h k (t), and the calculation formula is: where PV is the principal value integral and τ is the time variable; Combining the instantaneous amplitudes and frequencies of all IMFs into the time-frequency energy distribution Hilbert spectrum; Among them, δ(ff k (t) is Dirac delta including tax, f is the frequency variable, f k (t) is the kth rating variable; Integrating the Hilbert spectrum on the time axis to obtain the energy distribution in the frequency dimension and drawing the marginal spectrum; Taking the Hilbert spectrum, marginal spectrum, and time-frequency diagram of the short-time Fourier transform of different grinding states of the grinding wheel as the picture eigenvalue data.

8. The method according to claim 7, wherein The method for evaluating the importance of the feature vectors and image data of the data using the relief algorithm to obtain strongly correlated multi-modal data includes: Evaluate the discrimination ability of features in neighboring samples through the Relief algorithm, and convert the color feature image into a three-dimensional tensor X ∈ R h ×w×3 , and then flatten the three-dimensional tensor into a one-dimensional vector: X flat = [R 11 , R 12 ... R hw , G 11 , G 12 ... G hw , B 11 , B 12 ... B hw ​ Among them, R hw , G hw , B hw respectively represent the red, green, and blue channel values of the h-th row and w-th column; Randomly selecting a sample R from the dataset, finding a nearest neighbor sample NH of the same category as R through Euclidean distance calculation, and then finding the nearest neighbor sample NM of a different category from R. Successively traverse the weights of each eigenvalue through the weight update formula and update them: For each feature F, calculate the difference between the values of the instance R and NH on this feature using Euclidean distance. If the difference between R and NH is less than the preset threshold, it indicates that the importance of feature F is lower than the preset threshold, and the weight of feature F is weakened. The situation of R and NM is just the opposite.

9. The method according to claim 8, wherein The method for monitoring the online state of the grinding wheel by inputting the multi-modal data into the ResNet-DenseNet-LSTM hybrid network model includes: Process the image features. Through the stem module, perform a series of convolutional, pooling, and normalization operations to obtain a feature map of 75*75*64. Then enter the residual block module, and add a dense block behind the residual block. Dense connections enhance feature propagation and cross-layer information fusion. The output data of the dense block is a feature map of 75*75*192. Through the transition layer, perform convolution and average pooling on the output data of the dense block, and output a feature map of 37*37*96. Use the Flat module to expand the multi-dimensional features into a one-dimensional vector 96×37×37 = 131328; combine the eigenvalue data extracted by the convolutional neural network with the time-frequency features of the original data and input them into the LSTM network together. Finally, perform prediction through the fully connected layer and the output layer.

Images