Lightweight characterization and intelligent diagnosis methods for big data of rotating machinery condition monitoring

Abstract

A lightweight characterization and intelligent diagnosis method for big data of rotating machinery equipment condition monitoring aligns feature distribution with the original data set, reversely updates the data prototype, and then constructs a very small number of data prototypes with similar distribution but rich information. The constructed lightweight characterization data set replaces the original data set, and from the data volume level, replaces the massive original data with a very small amount of data. The present invention reduces the problems of data storage resource occupation, inability to transmit data in a timely manner, difficulty in data modeling, high computational cost, etc. caused by massive original data, and can provide data support for subsequent downstream tasks such as intelligent diagnosis, collaborative modeling, and privacy protection.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance and health management of rotating machinery equipment, and in particular to a lightweight characterization and intelligent diagnosis method of big data for status monitoring of rotating machinery equipment. Background Art

[0002] Rotating machinery plays an indispensable role in modern industry. Efficiently operating rotating machinery can accelerate production processes, improve production efficiency, ensure timely product delivery, and meet market demand. However, rotating components of rotating machinery operating under complex conditions cannot be maintained promptly due to difficulties in assembly and disassembly and high downtime costs. This can lead to fatigue damage and subsequent performance degradation. Early failures of rotating machinery can affect the normal operation of the equipment, while the continued deterioration of damage can cause production line downtime and even casualties. Therefore, it is necessary to implement condition monitoring, fault diagnosis, and remaining life prediction measures for rotating machinery to ensure stable operation.

[0003] The current historical condition monitoring data for rotating machinery is increasing, and related technologies in the fields of data-driven intelligent operation and maintenance and health management are developing rapidly. In actual engineering scenarios, to monitor the health status of a specific rotating machinery, real-time multi-point monitoring is often performed on its key components. This involves simultaneously deploying sensors of various signal types, such as vibration acceleration, acoustic emission, current, and temperature, at different measurement points on a single component to collect data from these different measurement points in real time. This massive amount of data provides data support for intelligent operation and maintenance methods, but the challenges brought about by big data are also becoming increasingly apparent.

[0004] The massive amount of historical monitoring data requires a large amount of storage resources for a long time, resulting in the inability to use resources in a timely manner. At the information transmission level, a large amount of data is limited by network transmission bandwidth and cannot be transmitted in time. The presence of redundant information in big data affects data analysis. The large amount of data consumes computing resources and increases the time cost when training models. Existing methods (Y.Fu, H.Cao, X.Chen, and J.Ding, "Improved broad learning system for machinery intelligent fault diagnosis with increasing fault samples, fault modes, and running conditions," ISA transactions, vol. 136, pp. 400-416, 2023.) filter the original data, that is, select a representative portion of data to form a sub-dataset with similar distribution, and then use this sub-dataset to replace the original data set. This method uses the sub-dataset to replace the original data, which reduces the data volume to a certain extent. However, it still retains a relatively large amount of data during the filtering process. Moreover, since the sub-dataset is real data filtered from the original data, it may have data privacy issues for data with high privacy requirements. On the other hand, due to the limited information in the original data, the performance of subsequent downstream tasks may be limited. Summary of the Invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to propose a lightweight representation and intelligent diagnosis method for big data of rotating machinery equipment condition monitoring, by aligning the feature distribution with the original data, reversely updating the data prototype, and then constructing a very small number of data prototypes with similar distribution but rich information, and replacing the original data set with the constructed lightweight representation data set. From the data volume level, a very small amount of data is used to replace the massive original data, reducing the problems of data storage resource occupation, inability to transmit data in a timely manner, difficulty in data modeling, high computing cost, etc. caused by massive original data, and can provide data support for subsequent downstream tasks such as intelligent diagnosis, collaborative modeling, and privacy protection.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for lightweight characterization and intelligent diagnosis of rotating machinery equipment condition monitoring big data includes the following steps:

[0008] Step 1: Given a time domain vibration signal dataset with Class C health status in represents the data point of the i-th sample, N represents the number of data points of each sample, y i={1,2,3,...,C} represents the sample data point x i The corresponding label, superscript N real Represents the total number of samples in the data set D1; Short Time Fourier Transform (STFT) is used to perform signal conversion on the time domain vibration signal data set D1, converting the one-dimensional vibration time domain signal into a two-dimensional time-frequency domain signal; the converted data set is a time-frequency domain dataset, and the time-frequency domain dataset is called the original dataset, and the corresponding label y after conversion i And the number of data set samples N real Remain unchanged, the time-frequency domain signal X i The calculation formula is as follows:

[0009]

[0010] Among them, x i (t) is the time domain signal, Represents the time-frequency domain signal, L and M are the dimensions of the signal in the time domain and frequency domain respectively, h(t-τ)e -j2πft is the basis function, h(t-τ) is the window function, f is the frequency in Fourier transform, j is the imaginary unit, τ is the time, and e is the natural base;

[0011] Step 2: Use the randomly selected data prototype X for each type of health status in the time-frequency domain dataset D2 j Initialize and get the initialized lightweight representation dataset in is the data prototype, and X j With X i The dimensions of y are consistent, j ={1,2,3,...,C} represents the sample point X j Corresponding health status label, superscript N prototype Represents the total number of samples in the lightweight representation dataset, and N prototype <N real ;

[0012] Step 3: Randomly select batch size N for each health state from the original dataset D2 and the lightweight representation dataset S batch The total number of samples selected from the original dataset D2 and the lightweight representation dataset S is C×N batch ;

[0013] Step 4: Perform data enhancement on the samples selected from the original dataset D2 in step 3. First, the time-frequency domain signal X selected from the original dataset D2 is i The corresponding time domain signal x before conversion iBy calculating the specified signal-noise ratio (SNR), we can obtain the signal-noise ratio of the time domain signal x. i One-dimensional noise signal x of the same dimension n ; Then, the one-dimensional noise signal x n Perform short-time Fourier transform to obtain the time-frequency domain signal X i Two-dimensional noise signal X with the same dimension n ; Finally, the time-frequency domain signal X i With the two-dimensional noise signal X n Add together to obtain the enhanced time-frequency domain signal Aug(X i ), the specific calculation formula is as follows:

[0014]

[0015]

[0016] Aug(X i )=X i +X n (4)

[0017] Where p represents the signal x n (t), N represents the length of the signal, SNR represents the signal-to-noise ratio, P s With P n Represent the power of signal and noise respectively, X i represents the time-frequency domain signal, X n Represents a two-dimensional noise signal, Aug(X i ) represents the enhanced time-frequency domain signal;

[0018] Step 5: Randomly initialize the data lightweight representation method model and fix the model weights, that is, the network cannot be trained; in each optimization process, the C×N obtained in step 3 is batch The number of time-frequency domain signals X i , data prototype X j The number obtained in step 4 is C×N batch The enhanced time-frequency domain signal Aug(X i ) are simultaneously input into the constructed data lightweight representation method model to obtain high-level feature representations f corresponding to different levels of the model for each health state. l (X i ),f l (X j ) and f l (Aug(X i )), where f l (·) represents the mapping relationship of the model, the subscript l represents the lth layer of the model, i,j={1,2,3...,N batch};

[0019] Step 6: Use the high-level features f of various mini-batch data obtained in step 5 l (X i ),f l (Aug(X i )) and f l (X j ) represent the original dataset and the lightweight representation dataset, respectively, with each class f l (X i ),f l (Aug(X i )) and f l (X j ) is taken as the loss term represents the distribution difference between data sets and minimizes each class f l (X i ),f l (Aug(X i )) and f l (X j ) To reduce the distribution differences between data sets, the calculation formula is as follows:

[0020]

[0021] in, is the final loss, C is the number of health status categories, N batch is the number of samples selected for each batch, f l (X i ),f l (Aug(X i )) and f l (X j ) are respectively the high-level features of the time-frequency domain signal, the enhanced time-frequency domain signal, and the data prototype in the network, and the subscript l represents the lth layer of the model;

[0022] Step 7: Use the loss calculated in step 6 As the optimization objective function, back propagation is used to calculate the gradient value to update the data prototype:

[0023]

[0024] Where η represents the learning rate;

[0025] Step 8: Loop through steps 3 to 7. Each loop performs real-time enhancement on the time-frequency domain signals and updates the data prototype. When the number of iterations reaches the maximum number of iterations I, the update ends, and the final lightweight representation dataset is obtained.

[0026] The step 5 is specifically as follows:

[0027] 5.1) The time-frequency domain signal, the enhanced time-frequency domain signal, and the data prototype are input into the data lightweight representation method model. First, feature extraction is performed. The feature extraction module is structured as three stacked sub-units, each of which includes a convolution layer, a group normalization layer, a leaky relu activation function layer, and an average pooling layer.

[0028] 5.2) After passing through the feature extraction module, each sample enters the state classification module. Each sample needs to be flattened into one-dimensional data. The flattened data then passes through two fully connected layers to identify the health category. The first fully connected layer uses the LeakyReLU activation function and a Dropout layer with a probability of 0.5. The second fully connected layer outputs the predicted value of each health state. Finally, it is normalized through the Softmax function to output the probability value of each category.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention proposes a lightweight characterization and intelligent diagnosis method for big data of rotating machinery equipment status monitoring. By using a deep convolutional neural network to align the feature distribution of the data prototype and the original data set, a large amount of original data is used to obtain a final extremely small amount of information-rich lightweight characterization data set after continuous updating and iteration. This greatly reduces the data storage cost, data transmission limitation, data redundancy interference, modeling time cost and high computing cost caused by massive data in data storage, data transmission, data analysis and data modeling. In addition, through real-time enhancement technology, the original data set is enhanced in each cycle, thereby greatly improving the anti-interference ability of the data prototype, and the data lightweight characterization method model can be better generalized for data in a noisy environment. Finally, in addition to realizing the fault diagnosis of rotating machinery equipment, the lightweight characterization data set constructed by the present invention has the characteristics of small data volume and non-real data, providing data support for downstream tasks such as data privacy and collaborative modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flowchart of the present invention.

[0032] Figure 2 Schematic diagram of a data lightweight characterization method model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention is described in further detail below with reference to the embodiments and accompanying drawings.

[0034] like Figure 1As shown, a method for lightweight characterization and intelligent diagnosis of rotating machinery equipment condition monitoring big data includes the following steps:

[0035] Step 1: Given a time domain vibration signal dataset with Class C health status in represents the data point of the i-th sample, N represents the number of data points of each sample, y i ={1,2,3,...,C} represents the sample data point x i The corresponding label, superscript N real Represents the total number of samples in the data set D1; Short Time Fourier Transform (STFT) is used to perform signal conversion on the time domain vibration signal data set D1, converting the one-dimensional vibration time domain signal into a two-dimensional time-frequency domain signal; the converted data set is a time-frequency domain dataset, and the time-frequency domain dataset is called the original dataset, and the corresponding label y after conversion i And the number of data set samples N real Remain unchanged, the time-frequency domain signal X i The calculation formula is as follows:

[0036]

[0037] Among them, x i (t) is the time domain signal, Represents the time-frequency domain signal, L and M are the dimensions of the signal in the time domain and frequency domain respectively, h(t-τ)e -j2πft is the basis function, h(t-τ) is the window function, f is the frequency in Fourier transform, j is the imaginary unit, τ is the time, and e is the natural base.

[0038] By using short-time Fourier transform, the time-domain vibration signal dataset is converted into a time-frequency domain dataset, which contains both time domain and frequency domain information, making the dataset more informative.

[0039] Step 2: Use the randomly selected data prototype X for each type of health status in the time-frequency domain dataset D2 j Initialize and get the initialized lightweight representation dataset in is the data prototype, and X j With X i The dimensions of y are consistent, j ={1,2,3,...,C} represents the sample point X j Corresponding health status label, superscript N prototype Represents the total number of samples in the lightweight representation dataset, and N prototype <<N real ;

[0040] By randomly selecting samples from the original data set to initialize the data prototype, with the help of the basic information of the original data, the initial state of the data prototype can be quickly determined, reducing the number of data prototype iteration steps;

[0041] Step 3: Randomly select batch size N for each health state from the original dataset D2 and the lightweight representation dataset S batch The total number of samples selected from the original dataset D2 and the lightweight representation dataset S is C×N batch ;

[0042] Step 4: Perform data enhancement on the samples selected from the original dataset D2 in step 3. First, the time-frequency domain signal X selected from the original dataset D2 is i The corresponding time domain signal x before conversion i By calculating the specified signal-noise ratio (SNR), we can obtain the signal-noise ratio of the time domain signal x. i One-dimensional noise signal x of the same dimension n ; Then, the one-dimensional noise signal x n Perform short-time Fourier transform to obtain the time-frequency domain signal X i Two-dimensional noise signal X with the same dimension n ; Finally, the time-frequency domain signal X i With the two-dimensional noise signal X n Add together to obtain the enhanced time-frequency domain signal Aug(X i ), the specific calculation formula is as follows:

[0043]

[0044]

[0045] Aug(X i )=X i +X n (4)

[0046] Where p represents the signal x n (t), N represents the length of the signal, SNR represents the signal-to-noise ratio, P s With P n Represent the power of signal and noise respectively, X i represents the time-frequency domain signal, X n Represents a two-dimensional noise signal, Aug(X i ) represents the enhanced time-frequency domain signal;

[0047] By adding noise to the original time-frequency domain signal, an enhanced time-frequency domain dataset is obtained, which expands the number of samples in the original dataset and improves the anti-interference ability of the lightweight representation dataset.

[0048] Step 5: Randomly initialize the data lightweight representation method model and fix the model weights, that is, the network cannot be trained; in each optimization process, the C×N obtained in step 3 is batch The number of time-frequency domain signals X i , data prototype X j The number obtained in step 4 is C×N batch The enhanced time-frequency domain signal Aug(X i ) are simultaneously input into the constructed data lightweight representation method model to obtain high-level feature representations f corresponding to different levels of health status in the model. l (X i ),f l (X j ) and f l (Aug(X i )), where f l (·) represents the mapping relationship of the model, the subscript l represents the lth layer of the model, i, j = {1, 2, 3..., N batch};

[0049] Reference Figure 2 , step 5 is as follows:

[0050] 5.1) After the time-frequency domain signal, the enhanced time-frequency domain signal, and the data prototype are input into the data lightweight representation method model, feature extraction is first performed. The feature extraction module is structured as three sequentially stacked sub-units, each of which includes a convolutional layer, a group normalization layer, a leaky relu activation function layer, and an average pooling layer.

[0051] 5.2) After passing through the feature extraction module, each sample enters the state classification module. Each sample needs to be flattened into one-dimensional data. The flattened data is then passed through two fully connected layers to identify the health category. The first fully connected layer uses the LeakyReLU activation function and a Dropout layer with a probability of 0.5. The second fully connected layer outputs the predicted value of each health state. Finally, it is normalized through the Softmax function to output the probability value of each category.

[0052] By fixing the weights of the lightweight representation method model, the network model cannot be updated, and the data prototype is set to be learnable, thereby obtaining a lightweight representation dataset;

[0053] Step 6: Use the high-level features f of various mini-batch data obtained in step 5 l (X i),f l (Aug(X i )) and f l (X j ) represent the original dataset and the lightweight representation dataset, respectively, with each class f l (X i ),f l (Aug(X i )) and f l (X j ) is taken as the loss term represents the distribution difference between data sets and minimizes each class f l (X i ),f l (Aug(X i )) and f l (X j ) To reduce the distribution differences between data sets, the calculation formula is as follows:

[0054]

[0055] in, is the final loss, C is the number of health status categories, N batch is the number of samples selected for each batch, f l (X i ),f l (Aug(X i )) and f l (X j ) are respectively the high-level features of the time-frequency domain signal, the enhanced time-frequency domain signal, and the data prototype in the network, and the subscript l represents the lth layer of the model;

[0056] By calculating the distance between the original time-frequency domain signal extracted by the model layer, the enhanced time-frequency domain signal, and the high-level features of the data prototype, and using this distance as a loss, the algorithm reduces the distance between data sets to obtain a data prototype with a distribution that is centrally adapted to the data set. While retaining the distribution information of the data prototype, the anti-interference ability of the lightweight representation data set is improved.

[0057] Step 7: Use the loss calculated in step 6 As the optimization objective function, back propagation is used to calculate the gradient value to update the data prototype:

[0058]

[0059] Where η represents the learning rate;

[0060] Step 8: Loop through steps 3 to 7. Each loop performs real-time enhancement on the time-frequency domain signals and updates the data prototype. When the number of iterations reaches the maximum number of iterations I, the update ends, and the final lightweight representation dataset is obtained.

[0061] The time-frequency domain signals are enhanced in real time in each cycle, that is, the enhanced data are different each time, which further improves the anti-interference ability of the final lightweight representation data set.

[0062] The feasibility of the method of the present invention is verified by taking the health status identification of gear data collected in the laboratory as an example.

[0063] The gear dataset used is gear vibration data collected on a laboratory power transmission test bench and contains six health states: first-stage sun gear spalling, first-stage sun gear wear, first-stage planet gear spalling, first-stage planet gear tooth root fracture, first-stage planet gear shaft eccentricity, and healthy. The sampling frequency during the experiment was 25.6 kHz, the motor rotation frequency was 35 Hz, and the magnetic powder brake current was 0.5 A. The specific information of the dataset is shown in Table 1.

[0064] Table 1 Gear vibration signal dataset information

[0065]

[0066] The window length of the short-time Fourier transform in data preprocessing is 1024, and the overlap rate is 0.9; the learning rate used in data prototype update is 3×10 -3 The optimizer uses stochastic gradient descent with a total number of iterations of 2000. After obtaining the final data prototype, in order to evaluate the performance of the constructed data prototype, a new diagnostic model is established using the lightweight representation dataset, and the diagnostic model is used to evaluate the test data. In order to avoid randomness, each group of experiments is repeated 5 times.

[0067] In addition, two sets of experiments were selected to compare the diagnostic results of the present invention. Comparison Method 1 did not use the lightweight data representation method proposed in this invention, and randomly selected one raw data point from each health state category for training and testing the diagnostic model. Comparison Method 2 was conducted in a noisy test environment and did not use the lightweight data representation method proposed in this invention, but randomly selected one raw data point from each health state category for training and testing the diagnostic model. Comparison Method 3 used the present invention's method, lightweighting each health state category into a data prototype without augmenting the raw data. The present invention's method lightweighted each health state category into a data prototype and also augmented the raw data. A comparison of the diagnostic results of the different methods is shown in Table 2. Table 2 shows that Comparison Method 1 had a lower diagnostic accuracy, with a five-test average accuracy of only 76.58%. The average diagnostic accuracy of Comparison Method 2 in the presence of noise further decreased, reaching only 72.93%. Comparison Method 3, using the present invention's method without accounting for noise, achieved an improved average diagnostic accuracy of 88.97% compared to Comparison Method 2. The present invention's method achieved the highest average diagnostic accuracy, at 95.63%.

[0068] Table 2 Comparison of diagnostic results of different methods

[0069]

[0070] By comparing the diagnostic results of methods 1, 2, and 3 with those of the method of the present invention, it can be seen that: first, the method of the present invention only has one data prototype for each type of health state, which greatly reduces the number of training sets compared to traditional fault diagnosis methods; then, the lightweight data representation dataset obtained by the present invention is used for subsequent diagnostic tasks, and the diagnostic accuracy is significantly improved compared to the comparison method 1 that does not use the method of the present invention; finally, the real-time data enhancement technology used by the present invention is compared with comparison methods 2 and 3, and it can be seen that it still performs well in noise interference conditions and can generalize well in noisy environments.

Claims

1. A method for lightweight characterization and intelligent diagnosis of rotating machinery equipment condition monitoring big data, characterized in that: The following steps are involved: Step 1: Given a time domain vibration signal dataset with Class C health status in represents the data point of the i-th sample, N represents the number of data points of each sample, y i ={1,2,3,...,C} represents the sample data point x i The corresponding label, superscript N real Represents the total number of samples in the data set D1; Short Time Fourier Transform (STFT) is used to perform signal conversion on the time domain vibration signal data set D1, converting the one-dimensional vibration time domain signal into a two-dimensional time-frequency domain signal; the converted data set is a time-frequency domain dataset, and the time-frequency domain dataset is called the original dataset, and the corresponding label y after conversion i And the number of data set samples N real Remain unchanged, the time-frequency domain signal X i The calculation formula is as follows: Among them, x i (t) is the time domain signal, Represents the time-frequency domain signal, L and M are the dimensions of the signal in the time domain and frequency domain respectively, h(t-τ)e -j2πft is the basis function, h(t-τ) is the window function, f is the frequency in Fourier transform, j is the imaginary unit, τ is the time, and e is the natural base; Step 2: Randomly select samples from the time-frequency domain dataset D2 to complete the data prototype X for each type of health status j Initialization, get the initialized lightweight representation dataset in is the data prototype, and X j With X i The dimensions of y are consistent, j ={1,2,3,...,C} represents the sample point X j Corresponding health status label, superscript N prototype Represents the total number of samples in the lightweight representation dataset, and N prototype <<N real ; Step 3: Randomly select batch size N for each health state from the original dataset D2 and the lightweight representation dataset S batch The total number of samples selected from the original dataset D2 and the lightweight representation dataset S is C×N batch ; Step 4: Perform data enhancement on the samples selected from the original dataset D2 in step 3. First, the time-frequency domain signal X selected from the original dataset D2 is i The corresponding time domain signal x before conversion i By calculating the specified signal-noise ratio (SNR), we can obtain the signal-noise ratio of the time domain signal x. i One-dimensional noise signal x of the same dimension n ; Then, the one-dimensional noise signal x n Perform short-time Fourier transform to obtain the time-frequency domain signal X i Two-dimensional noise signal X with the same dimension n ; Finally, the time-frequency domain signal X i With the two-dimensional noise signal X n Add and obtain the enhanced time-frequency domain signal Aug(X i ), the specific calculation formula is as follows: Aug(X i )=X i +X n (4) Where p represents the signal The power of the signal, N represents the length of the signal, SNR represents the signal-to-noise ratio, P s With P n Represent the power of signal and noise respectively, X i represents the time-frequency domain signal, X n Represents a two-dimensional noise signal, Aug(X i ) represents the enhanced time-frequency domain signal; Step 5: Randomly initialize the data lightweight representation method model and fix the model weights, that is, the network cannot be trained; in each optimization process, the C×N obtained in step 3 is batch The number of time-frequency domain signals X i , data prototype X j The number obtained in step 4 is C×N batch The enhanced time-frequency domain signal Aug(X i ) are simultaneously input into the constructed data lightweight representation method model to obtain high-level feature representations f corresponding to different levels of the model for each health state. l (X i ),f l (X j ) and f l (Aug(X i )), where f l (·) represents the mapping relationship of the model, the subscript l represents the lth layer of the model, i,j={1,2,3...,N batch }; Step 6: Use the high-level features f of various mini-batch data obtained in step 5 l (X i ),f l (Aug(X i )) and f l (X j ) represent the original dataset and the lightweight representation dataset, respectively, with each class f l (X i ),f l (Aug(X i )) and f l (X j ) is taken as the loss term L d represents the distribution difference between data sets and minimizes each class f l (X i ),f l (Aug(X i )) and f l (X j ) distance L d To reduce the distribution differences between data sets, the calculation formula is as follows: in, is the final loss, C is the number of health status categories, N batch is the number of samples selected for each batch, f l (X c,i ),f l (Aug(X c,i )) and f l (X c,j ) are respectively the time-frequency domain signal of class c health state, the enhanced time-frequency domain signal, and the high-level features of the data prototype in the network. The subscript l represents the lth layer of the model. Step 7: Use the loss calculated in step 6 As the optimization objective function, back propagation is used to calculate the gradient value to update the data prototype: Where η represents the learning rate; Step 8: Loop through steps 3 to 7. Each loop performs real-time enhancement on the time-frequency domain signals and updates the data prototype. When the number of iterations reaches the maximum number of iterations I, the update ends, and the final lightweight representation dataset is obtained.

2. The method according to claim 1, characterized in that The step 5 is specifically as follows: 5.1) The time-frequency domain signal, the enhanced time-frequency domain signal, and the data prototype are input into the data lightweight representation method model. First, feature extraction is performed. The feature extraction module is structured as three stacked sub-units, each of which includes a convolutional layer, a group normalization layer, a leaky relu activation function layer, and an average pooling layer. 5.2) After passing through the feature extraction module, each sample enters the state classification module. Each sample needs to be flattened into one-dimensional data. The flattened data then passes through two fully connected layers to identify the health category. The first fully connected layer uses the LeakyReLU activation function and a Dropout layer with a probability of 0.

5. The second fully connected layer outputs the predicted value of each health state. Finally, it is normalized through the Softmax function to output the probability value of each category.

Images