Industrial production and manufacturing intelligent supervision system based on Internet of Things

Abstract

The invention relates to the field of production and manufacturing supervision, in particular to an industrial production and manufacturing intelligent supervision system based on the Internet of Things, which comprises a data collection module, a feature extraction module, a finite element simulation module and a generative adversarial network model, the feature extraction module is used for obtaining a track of dominant frequency changing along with time, the finite element simulation module is used for simulating a stress nephogram according to input energy distribution and the track of dominant frequency changing along with time, and the generative adversarial network model is used for generating a crack distribution diagram. And comparing the crack distribution map with a simulated stress nephogram to judge the authenticity of the image. By collecting the vibration information of the equipment and carrying out frequency band analysis on the vibration information, preparation can be carried out in advance before the equipment is damaged, and economic losses caused by the fact that faults cannot be processed in time when occurring are avoided.

Description

Technical Field

[0001] The present invention relates to the field of production and manufacturing supervision, and particularly to an intelligent supervision system for industrial production and manufacturing based on the Internet of Things. Background Art

[0002] With the continuous development of the Internet of Things, the association between devices has become increasingly complex, and systematic and automated production lines have gradually replaced traditional manual production and manufacturing. With the continuous popularization of systematization and automation, more and more enterprises have begun to adopt intelligent supervision deployments for the supervision of production lines. Different from manual supervision, the intelligent supervision system not only has higher advantages in supervision efficiency and supervision cost, but also has faster feedback efficiency in response to faulty devices.

[0003] In the existing supervision systems adopted in industrial production and manufacturing, it is usually judged whether the devices in the Internet of Things are in an abnormal working state according to the working conditions such as the service life, current, and voltage of the devices. Although this supervision method can accurately know the location of the abnormal state, it cannot prevent the faults of the devices in advance. This is because the aging of the devices does not cause the devices to be unusable, but when the load of the aging devices is too large, direct fault problems may occur, and these fault problems cannot be handled by detecting the current, voltage, etc. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides the following technical solutions: An intelligent supervision system for industrial production and manufacturing based on the Internet of Things, comprising: a data collection module, a feature extraction module, a finite element simulation module, and a generative adversarial network model.

[0005] The data collection module is used to collect the vibration information of the devices. The feature extraction module is used to perform wavelet packet decomposition on the vibration information to obtain the energy distribution of each frequency band, and obtain the trajectory of the dominant frequency changing with time from the abnormal frequency band. The finite element simulation module is used to simulate the stress nephogram according to the input energy distribution and the trajectory of the dominant frequency changing with time. The generative adversarial network model is used to generate a crack distribution map according to the trajectory of the dominant frequency changing with time and random noise, compare the crack distribution map with the simulated stress nephogram to judge the authenticity of the image, and analyze according to the stress nephogram and the crack distribution map to judge the fault condition of the devices.

[0006] As an improvement of the above technical solution, the data collection module at least includes: a laser Doppler vibrometer module and an acoustic emission signal processing module.

[0007] The vibration information at least includes the vibration signal and stress wave information on the surface of the device. The laser Doppler vibrometer module is used to measure the vibration signal on the surface of the device, and the acoustic emission signal processing module is used to capture the stress wave information released by internal cracks in the material.

[0008] As an improvement of the above technical solution, the feature extraction module at least includes: a time-domain convolutional network module.

[0009] The method for obtaining the trajectory of the dominant frequency changing with time from the abnormal frequency band includes the following steps: Decompose the input vibration information into several sub-bands through the time-domain convolutional module, calculate the energy entropy of each band, and locate the abnormal frequency band according to the energy entropy.

[0010] Extract the trajectory of the dominant frequency changing with time from the abnormal frequency band and capture the instantaneous impact characteristics.

[0011] As an improvement of the above technical solution, the wavelet packet decomposition adopts the multi-Bessel wavelet decomposition method to decompose the vibration information into fourth-order wavelets, generate 16 sub-bands, and calculate the energy entropy of each sub-band.

[0012] As an improvement of the above technical solution, the extraction of the trajectory of the dominant frequency changing with time from the abnormal frequency band includes the following steps: Perform sub-band transformation on the vibration information by using the multi-Bessel wavelet decomposition method and calculate the time-frequency distribution to generate a time-frequency matrix.

[0013] Traverse the time points, find the maximum energy frequency at each time point, record the time-frequency coordinates of this energy frequency, and output the time-frequency ridge line of the energy frequency as the extracted trajectory of the dominant frequency changing with time.

[0014] As an improvement of the above technical solution, the generation of the time-frequency matrix depends on the following formula:

[0015] where represents the input original vibration information, is the scale parameter, is the translation parameter, is the wavelet basis function, is the scaling factor.

[0016] As an improvement of the above technical solution, the method for the finite element simulation module to perform stress cloud map simulation includes the following steps: Input the trajectory of the dominant frequency changing with time into the finite element simulation module to obtain the corresponding dynamic load, extract the frequency and energy characteristics of the dynamic load, and obtain the transient load boundary conditions according to the frequency and energy characteristics.

[0017] Obtain the slope of the time-frequency ridge line according to the trajectory of the dominant frequency changing with time, and estimate the crack growth rate according to the slope of the ridge line.

[0018] Perform a simulation of the stress contour map according to the transient load boundary conditions and the crack growth rate. If the matching degree between the simulated stress contour map and the measured time-frequency characteristics is higher than the preset value, it is determined that the characteristics are valid. If it is lower than the preset value, feature extraction is re-performed through the time-domain convolution module.

[0019] As an improvement of the above technical solution, the generative adversarial network model includes a generator module and a discriminator module; The generator module is used to generate a crack distribution map according to the trajectory of the dominant frequency changing with time and random noise, and the discriminator module is used to compare the input crack distribution map with the simulated stress contour map to judge the authenticity probability of the simulated stress contour map image.

[0020] As an improvement of the above technical solution, the judgment of the authenticity of the image includes the following steps: If the authenticity probability of the image is judged to be greater than or equal to the preset threshold after comparing the crack distribution map with the simulated stress contour map, it is directly output. If the authenticity probability of the image is judged to be lower than the preset threshold after comparing the crack distribution map with the simulated stress contour map, it is considered that there is an error in the simulation result.

[0021] Advantages of the present invention: By collecting the vibration information of the device and performing frequency band analysis on the vibration information, the simulation of the device material can be carried out, the aging of cracks and other conditions can be judged, and the usage situation of the device can be analyzed. According to this information, preparations can be made in advance before the device is damaged to avoid economic losses caused by the inability to handle in time when a fault occurs. Brief Description of the Drawings

[0022] Figure 1 It is the principle block diagram of the present invention. Detailed Embodiments

[0023] The following specific examples illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0024] In the existing supervision systems adopted in industrial production and manufacturing, the device in the Internet of Things is usually judged whether it is in an abnormal working state according to the service life, current, voltage and other working conditions of the device. Although this supervision method can accurately know the location of the abnormal state, it cannot prevent the device failure in advance. This is because the aging of the device does not cause the device to be unusable, but when the load of the aging device is too large, direct failure problems may occur, and such failure problems cannot be handled by detecting the current, voltage and other conditions.

[0025] To solve the above problems, please refer to the figure. There is provided an intelligent supervision system for industrial production and manufacturing based on the Internet of Things, including: a data collection module, a feature extraction module, a finite element simulation module, and a generative adversarial network model.

[0026] The data collection module is used to collect the vibration information of the device. The feature extraction module is used to perform wavelet packet decomposition on the vibration information to obtain the energy distribution of each frequency band, and obtain the trajectory of the dominant frequency changing with time from the abnormal frequency band. The finite element simulation module is used to simulate the stress nephogram according to the input energy distribution and the trajectory of the dominant frequency changing with time. The generative adversarial network model is used to generate a crack distribution map according to the trajectory of the dominant frequency changing with time and random noise, judge the authenticity of the image by comparing the crack distribution map with the simulated stress nephogram, and analyze according to the stress nephogram and the crack distribution map to judge the fault condition of the device.

[0027] First, the data collection module collects the vibration information of the device. This vibration information at least includes the vibration signal and stress wave information on the surface of the device. The vibration signal on the surface of the device refers to the vibration state generated during the operation of the device. The relevant information of these vibration states is usually collected by the laser Doppler vibrometer module under the data collection module. The vibration speed on the surface of the device is measured by the method of laser interference. The stress wave information refers to the internal crack condition of the material of the manufacturing device. These internal crack conditions cannot be measured by the method of laser interference. Therefore, the method of acoustic emission signal processing is used for measurement, that is, the AE detection device. It judges the degree of internal structure damage by detecting the stress wave emitted inside the stressed material. Through this method, the internal crack condition of the material can be detected.

[0028] After the vibration information is collected, feature extraction needs to be performed on the vibration information. In this step, we usually use the feature extraction module. In this embodiment, the feature extraction module uses a time-domain convolutional network module. Based on the TCN model, it can expand the receptive field of the convolutional kernel by adjusting the dilation coefficient, so as to more effectively process the features of long-distance dependence. Therefore, we use the TCN model to perform feature extraction on the vibration information, which can effectively process long time series data and maintain the time order.

[0029] When considering that the generated output image of the crack situation may be a grayscale image, wavelet packet decomposition can be used to process the vibration information during feature extraction. Among them, the wavelet packet decomposition adopts the multi-Bezier wavelet decomposition method to decompose the vibration information into fourth-order wavelets, generating 16 sub-bands, and calculating the energy entropy of each sub-band.

[0030] Based on the above analysis, it can be known that the method for obtaining the trajectory of the dominant frequency changing with time from the abnormal frequency band can adopt the following steps: Decompose the input vibration information into 16 sub-bands through the time-domain convolution module, calculate the energy entropy of each band, and locate the abnormal frequency band according to the energy entropy.

[0031] Extract the trajectory of the dominant frequency changing with time from the abnormal frequency band and capture the instantaneous impact characteristics.

[0032] For the TCN model, first, a neural network module needs to be established, and this neural network module is used to extract features from the input vibration signal. Here, after the input vibration information is decomposed into 16 sub-bands and the energy entropy is calculated, it is input into the neural network module. The structure of its neural network module is as follows: class DilatedConv1d(nn.Module): def __init__(self, in_channels=64, out_channels=64, dilation=1): super().__init__() self.conv = nn.Conv1d( in_channels, out_channels, kernel_size=3, dilation=dilation,# Coefficient of dilation (2^i) padding=dilation# Keep the time series length unchanged ) self.relu = nn.ReLU() def forward(self, x): return self.relu(self.conv(x)) Among them, the line "class DilatedConv1d(nn.Module)" defines a new class DilatedConv1d that inherits from nn.Module, indicating that this is a custom PyTorch neural network module. nn.Module is the base class for all neural network layers and models in PyTorch.

[0033] def __init__(self, in_channels=64, out_channels=64, dilation=1) represents the constructor of the class and initializes the parameters of this convolutional module.

[0034] super().__init__() calls the constructor of the parent class nn.Module to ensure that the class inherited from nn.Module is correctly initialized.

[0035] self.conv = nn.Conv1d(...) This line defines a 1D convolutional layer, namely nn.Conv1d.

[0036] self.relu = nn.ReLU() Here, a ReLU activation function (nn.ReLU()) is defined. ReLU is a commonly used activation function, and its role is to set the negative value part to zero and keep the positive value part unchanged. ReLU is often used to improve the non-linear expression ability of neural networks.

[0037] def forward(self, x): This is the forward propagation function of the model. The forward function in PyTorch defines how data flows through this network module.

[0038] return self.relu(self.conv(x)) This line indicates that the data is processed in the following two steps: 1. First, the input x is passed into the convolutional layer self.conv(x), which performs a 1D dilated convolution operation and returns the output after convolution.

[0039] 2. Then, the output of the convolution passes through the ReLU activation function self.relu() to perform non-linear processing on the convolution result. Finally, the result after convolution and activation is returned.

[0040] The data output by the above architecture is usually the important features of the vibration signal in the time domain. In this embodiment, after performing sub-band transformation on the vibration information using the multi-Bezier wavelet decomposition method, it is necessary to calculate the time-frequency distribution information in the transformation process, and integrate these time-frequency distribution information to generate a time-frequency matrix. The generation of the time-frequency matrix depends on the following formula:

[0041] Among them, represents the original vibration information of the input, is the scale parameter, is the translation parameter, is the wavelet basis function, is the scaling factor.

[0042] After the generation of the time-frequency matrix is completed, the time-domain convolutional neural network module traverses the time points to find the maximum energy frequency at each time point, records the time-frequency coordinates of this energy frequency, and the connection of these time-frequency coordinates is the time-frequency ridge line, and this time-frequency ridge line is the trajectory of the dominant frequency changing with time.

[0043] The extraction of the time-frequency ridge line usually depends on the following code: def extract_ridge(tf_matrix): ridges = [] for t in range(tf_matrix.shape[1]): # Traverse the time points freq_idx = np.argmax(tf_matrix[:, t]) # Find the maximum energy frequency at the current moment ridges.append((t, freq_idx)) # Record the time-frequency coordinates return ridges ridges = [] indicates the initialization of the ridge line list. The time points are traversed through the code "for t in range(tf_matrix.shape[1]:", the maximum energy frequency at this time point is found through "freq_idx = np.argmax(tf_matrix[:, t])", and the time-frequency coordinates of the current time point are recorded through "ridges.append((t, freq_idx))". The recording position is the ridges ridge line list. After all time points are recorded, the connection of the obtained time-frequency coordinates is the time-frequency ridge line.

[0044] After obtaining the time-frequency ridge line, the simulation generation of the stress nephogram can be carried out through finite element simulation analysis. Specifically, the simulation method of the stress nephogram by the finite element simulation module includes the following steps: Input the time-frequency ridge line into the finite element simulation module to obtain the corresponding dynamic load, extract the frequency and energy characteristics of the dynamic load, and obtain the transient load boundary conditions based on the frequency and energy characteristics.

[0045] The above steps rely on the following code: def prepare_fem_input(ridges): freq = [f for (t, f) in ridges] energy = compute_energy(ridges) load_params = { 'frequency': np.mean(freq), 'amplitude': energy_to_force(energy), 'duration': len(ridges) * time_step } return load_params Extract the corresponding data parameters from the ridges ridge line list. This data parameter usually contains frequency and energy characteristics. Specifically, freq = [f for (t, f) in ridges] means extracting the frequency characteristics of each ridge line and generating a frequency list freq. Each element in the list represents the maximum energy frequency corresponding to a certain time point. energy = compute_energy(ridges) means calculating the energy characteristics of the vibration signal by calling the compute_energy function and calculating the total energy according to the distribution of the ridge line in time and frequency. 'frequency': np.mean(freq) means calculating the average value of all frequencies in the frequency list to represent the frequency of the applied load. 'amplitude': energy_to_force(energy) means converting the calculated energy into the amplitude of the applied force. 'duration': len(ridges) * time_step means estimating the duration of the applied load by multiplying the number of ridge lines by each time step. Finally, these three parameters are integrated and returned as the transient load boundary conditions.

[0046] On the basis of the above steps, further estimate the crack growth length. Since the crack growth length needs to consider the relationship between frequency and dynamic load, it is also necessary to obtain the time-frequency ridge slope according to the trajectory of the dominant frequency changing with time, and then estimate the crack growth rate according to the ridge slope.

[0047] Specifically, the relationship between frequency and dynamic load is as follows:

[0048] Among them, represents the value of the dynamic load at time moment, represents the amplitude of the dynamic load, which is obtained by back-calculating the frequency band energy entropy, represents the dominant frequency extracted from the time-frequency ridge, which represents the most significant frequency component in the signal, is the time variable, represents a sine function, indicating the dynamic load changing with time.

[0049] The calculation of the crack propagation rate depends on the following formula:

[0050] Among them, represents the crack propagation rate, C is a material constant, determined by experiments, and is usually related to the fatigue characteristics of the material, is the stress intensity factor amplitude, which is the stress level at the crack tip and is related to the external load and crack size. When the load increases, increases accordingly, thus promoting crack propagation, is a material parameter, usually an exponent, indicating the sensitivity of the crack propagation rate to the stress intensity factor.

[0051] After obtaining the transient load boundary conditions and the crack growth rate, stress cloud diagram simulation can be carried out according to these parameters. If the matching degree between the simulated stress cloud diagram and the measured time-frequency characteristics is higher than the preset value, it is judged that the characteristics are effective. If it is lower than the preset value, feature extraction is re-performed through the time domain convolution module.

[0052] The preset value can usually be set to 5%-10%. When the error is less than the preset value, it is considered that the simulation result generated according to the characteristics is accurate, that is, the characteristics are effective. If the error is higher than the preset value, it is considered that the error of the simulation result generated according to the characteristics is large, that is, the effectiveness of the characteristics is in doubt. At this time, feature extraction can be re-performed through the time domain convolution module. When re-extracting, the number of layers of wavelet packet decomposition needs to be increased, that is, the decomposition wavelet order can be set to 6, and then the calculations in the above steps are re-performed.

[0053] Of course, after the generation of the simulation image, stress and crack visualization is still required. In this embodiment, it is processed through a generative adversarial network model, where the generative adversarial network model includes a generator module and a discriminator module. The generator module is used to generate a crack distribution map based on the trajectory of the dominant frequency changing with time and random noise, and the discriminator module is used to compare the input crack distribution map with the simulation stress nephogram to judge the authenticity probability of the simulation stress nephogram image.

[0054] Specifically, the judgment of the authenticity of the image includes the following steps: If the authenticity probability of the image is judged to be greater than or equal to the preset threshold after comparing the crack distribution map with the simulated stress nephogram, it is directly output. If the authenticity probability of the image is judged to be lower than the preset threshold after comparing the crack distribution map with the simulated stress nephogram, it is considered that there is an error in the simulation result.

[0055] For the further implementation of the generator module and the discriminator module in the above solution, the details are as follows: In order for the generator to receive the generated features, the extracted feature information also needs to be converted into a 128-dimensional time series feature vector through the following code. The details are as follows: class TCN(nn.Module): def forward(self, x): # Input x: (batch_size, seq_len, input_dim) # Output: (batch_size, 128) return self.tcn_layers(x)[:, -1, :] # Take the last state of the sequence The structure of the generator module is set up as follows: class Generator(nn.Module): def __init__(self): super().__init__() self.fc = nn.Linear(160, 256*8*8) self.upconv = nn.Sequential( nn.ConvTranspose2d(256, 128, 4, stride=2, padding=1), nn.BatchNorm2d(128), nn.ReLU(), nn.ConvTranspose2d(128, 64, 4, stride=2, padding=1), nn.BatchNorm2d(64), nn.ReLU(), nn.ConvTranspose2d(64, 1, 4, stride=2, padding=1), nn.Tanh() ) def forward(self, z, features): x = torch.cat([z, features], dim=1) x = self.fc(x).view(-1, 256, 8, 8) return self.upconv(x) The functional principle of the above code is as follows: Define a class named Generator, which inherits from nn.Module. This is the basic class for all neural network modules in PyTorch.

[0056] In the __init__ method, call the constructor of the parent class super().__init__().

[0057] self.fc = nn.Linear(160, 256 * 8 * 8) means using a fully connected layer Linear. The input is a combination of the feature vector and the random noise (128 - dimensional feature vector + 32 - dimensional noise), with a total of 160 dimensions, and the output size is 256×8×8. The role of this layer is to expand the input into a higher - dimensional feature map for subsequent transposed convolution processing.

[0058] self.upconv = nn.Sequential(……) means using nn.Sequential to define a series of transposed convolution layers to upsample a smaller feature map to generate a larger image. The role of these layers is to gradually expand the feature map from 8x8 to 64x64.

[0059] The forward method defines the forward propagation process of the generator: Input: Receive a random noise vector z and a feature vector features.

[0060] Concatenation: Use `torch.cat` to concatenate the noise `z` and the feature vector along the channel dimension (dim = 1) to form a new input tensor.

[0061] Fully connected layer processing: Process the concatenated tensor through the fully connected layer `self.fc`, and then reshape it to 256 feature maps with a size of 8x8 using `.view(-1, 256, 8, 8)` in preparation for the subsequent transposed convolution process.

[0062] Output the crack distribution image: Input the feature maps into `self.upconv` to generate the final crack distribution image.

[0063] The structure of the discriminator module is set up as follows: class Generator(nn.Module): def __init__(self): super().__init__() self.fc = nn.Linear(160, 256 * 8 * 8) self.upconv = nn.Sequential( nn.ConvTranspose2d(256, 128, 4, stride = 2, padding = 1), nn.BatchNorm2d(128), nn.ReLU(), nn.ConvTranspose2d(128, 64, 4, stride = 2, padding = 1), nn.BatchNorm2d(64), nn.ReLU(), nn.ConvTranspose2d(64, 1, 4, stride = 2, padding = 1), nn.Tanh() ) def forward(self, z, features): x = torch.cat([z, features], dim = 1) x = self.fc(x).view(-1, 256, 8, 8) return self.upconv(x) The functional principle of the above code is as follows: Define a class named Discriminator, which inherits from nn.Module, the base class for all neural network modules in PyTorch. Initialize it in the __init__ method by calling the constructor of the parent class super().__init__().

[0064] Use nn.Sequential to define a series of convolutional layers, activation functions, batch normalization, and fully connected layers to form a sequential network structure. Specifically, convolutional layer (Conv2d), activation function (LeakyReLU), batch normalization (BatchNorm2d), flattening (Flatten), and fully connected layer (Linear).

[0065] The forward method defines how to perform forward propagation: Accept two inputs: img (crack distribution map) and stress_map (stress nephogram).

[0066] Use torch.cat to concatenate the two images along the channel dimension to form a tensor of shape (B, 2, H, W), where B is the batch size, and H and W are the height and width of the images respectively.

[0067] Input the concatenated tensor into the defined convolutional network downconv, and limit the output between 0 and 1 through the torch.sigmoid function, representing the authenticity probability of the input image. Usually, a threshold a is set. When the output result is between a - 1, it is considered that the crack distribution image is close to the real situation. If the output result is between 0 - a, it is considered that there is a large error between the image and the real situation.

[0068] The finally generated crack distribution map is an image with strong authenticity after identification. Combine the simulation results of the stress nephogram to judge the fault situation of the device. When the device is at the fault critical point, an alarm can be issued in advance to remind the construction personnel to turn off the corresponding device and replace it. This supervision scheme can achieve early warning actions before the device is damaged. Compared with the traditional monitoring of current, voltage, heat, etc., it can play a better preventive role. Of course, in this system, the traditional detection system can still be integrated for unified coordination and scheduling to achieve a more perfect intelligent supervision system for production and manufacturing.

[0069] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes completed by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. An intelligent supervision system for industrial production and manufacturing based on the Internet of Things, characterized in that, Including: A data collection module for collecting vibration information of the device; A feature extraction module for performing wavelet packet decomposition on the vibration information to obtain the energy distribution of each frequency band, and obtaining the trajectory of the dominant frequency changing with time from the abnormal frequency band; A finite element simulation module for simulating the stress nephogram according to the input energy distribution and the trajectory of the dominant frequency changing with time; A generative adversarial network model for generating a crack distribution map according to the trajectory of the dominant frequency changing with time and random noise, comparing the crack distribution map with the simulated stress nephogram to judge the authenticity of the image, and analyzing according to the stress nephogram and the crack distribution map to judge the fault condition of the device.

2. The intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 1, characterized in that: The data collection module at least includes: a laser Doppler vibrometer module and an acoustic emission signal processing module; The vibration information at least includes the vibration signal on the surface of the device and the stress wave information. The laser Doppler vibrometer module is used to measure the vibration signal on the surface of the device, and the acoustic emission signal processing module is used to capture the stress wave information released by cracks inside the material.

3. An intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 1, characterized in that: The feature extraction module at least includes: a time-domain convolutional network module; The method for obtaining the trajectory of the dominant frequency changing with time from the abnormal frequency band includes the following steps: Decompose the input vibration information into several sub-frequency bands through the time-domain convolution module, calculate the energy entropy of each frequency band, and locate the abnormal frequency band according to the energy entropy; Extract the trajectory of the dominant frequency changing with time from the abnormal frequency band, and capture the instantaneous impact characteristics.

4. An intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 3, characterized in that: The wavelet packet decomposition adopts the multi-Bezier wavelet decomposition method to decompose the vibration information into fourth-order wavelets, generate 16 sub-frequency bands, and calculate the energy entropy of each sub-frequency band.

5. An intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 4, characterized in that: The method for extracting the trajectory of the dominant frequency changing with time from the abnormal frequency band includes the following steps: Perform sub-frequency band transformation on the vibration information by using the multi-Bezier wavelet decomposition method and calculate the time-frequency distribution to generate a time-frequency matrix; Traverse the time points, find the maximum energy frequency at each time point, record the time-frequency coordinates of the energy frequency, and output the time-frequency ridge line of the energy frequency as the extracted trajectory of the dominant frequency changing with time.

6. The intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 5, wherein: The generation of the time-frequency matrix depends on the following formula: Among them, \(x(t)\) represents the original vibration information of the input, \(a\) is the scale parameter, \(b\) is the translation parameter, and \(\psi\) is the wavelet basis function. is the scaling factor.

7. An intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 3, characterized in that: The method for the finite element simulation module to simulate the stress nephogram includes the following steps: Input the trajectory of the dominant frequency changing with time into the finite element simulation module to obtain the corresponding dynamic load, extract the frequency and energy characteristics of the dynamic load, and obtain the transient load boundary conditions according to the frequency and energy characteristics; Obtain the slope of the time-frequency ridge line according to the trajectory of the dominant frequency changing with time, and estimate the crack growth rate according to the ridge line slope; Perform the simulation of the stress nephogram according to the transient load boundary conditions and the crack growth rate. If the matching degree between the simulated stress nephogram and the measured time-frequency characteristics is higher than the preset value, it is judged that the characteristics are valid. If it is lower than the preset value, the feature extraction is re-performed through the time-domain convolution module.

8. An intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 7, characterized in that: The generative adversarial network model includes a generator module and a discriminator module; The generator module is used to generate a crack distribution map according to the trajectory of the dominant frequency changing with time and random noise, and the discriminator module is used to compare the input crack distribution map with the simulated stress nephogram to judge the authenticity probability of the simulated stress nephogram image.

9. An intelligent supervision system for industrial production and manufacturing based on the Internet of Things according to claim 8, characterized in that: The judgment of the authenticity of the image includes the following steps: If the probability of judging the authenticity of the image is greater than or equal to the preset threshold after comparing the crack distribution map with the simulated stress nephogram, it is directly output. If the probability of judging the authenticity of the image is lower than the preset threshold after comparing the crack distribution map with the simulated stress nephogram, it is considered that there is an error in the simulation result.

Images