Mobile equipment vibration monitoring and alarming system based on convolutional neural network

Abstract

The invention discloses a mobile equipment vibration monitoring and alarming system based on a convolutional neural network, and the system comprises the following steps: collecting the abnormal vibration data of mobile equipment, and arranging the abnormal vibration data into a data set; the method comprises the following steps: constructing a convolutional neural network, acquiring vibration signals through a six-axis inertial measurement unit of an MEMS, performing short-time Fourier transform on data, extracting feature signals, and classifying the data according to time-frequency features of the short-time Fourier transform; inputting the processed data into a trained convolutional neural network, performing adaptive extraction on abnormal vibration by the convolutional neural network, and outputting an abnormal state of the mobile equipment; and step 6, carrying out positioning monitoring on the mobile equipment through a GPS positioning module, and uploading the abnormal state, abnormal data and positioning information data of the mobile equipment to a monitoring alarm platform. The system has the advantages that the running state of the mobile equipment can be monitored in real time, and an alarm is given out in time when abnormal vibration occurs.

Description

Technical Field

[0001] The invention relates to the technical field of vibration monitoring, and in particular to a vibration monitoring and alarm system for mobile equipment based on a convolutional neural network. Background Art

[0002] In production and life, most faults and disasters can be detected from noise fluctuations and abnormal vibration information before they occur. In particular, vibration monitoring and sound monitoring play an important role in equipment and product testing. When the equipment is working normally, a database of normal vibration frequencies is established. When the equipment is working abnormally, the faulty part will generate abnormal high-frequency or low-frequency vibration frequency information and noise frequency information. The receiving terminal will detect the problem in real time, analyze the collected abnormal data, and take countermeasures in advance, so as to avoid losses in time before the equipment is seriously damaged or even an accident occurs.

[0003] At present, the vibration sensor systems used to monitor equipment abnormalities are mainly large-volume, high-cost, and high-precision instruments and equipment, which are costly and cumbersome to install and use. In addition, the monitoring environment of most equipment is mainly monitored in real time by cameras, which cannot accurately judge the operating status of facilities and equipment (especially the status of internal operating mechanisms) in a multi-dimensional and three-dimensional manner; in addition, the traditional system monitoring method is completed manually, and experienced staff independently judge whether the equipment is in a normal state based on factors such as sound and vibration frequency. This judgment method has a very large self-subjective factor, and the conclusions drawn are not only unreliable and poorly standardized, but also require staff to have extremely strong professionalism and rich experience, which is difficult to promote.

[0004] In addition, there are some fault monitoring systems for single components of equipment, but such single fault monitoring systems can only reflect the local health status of the equipment and cannot comprehensively evaluate the overall operating status of the equipment. In addition, most current equipment still relies on traditional regular inspections or passive responses after a fault occurs, and fails to achieve large-scale real-time monitoring in the IoT environment. This lagging fault discovery mode undoubtedly exacerbates potential safety hazards and uncertainties. Summary of the invention

[0005] The purpose of the present invention is to make up for the above-mentioned deficiencies and to disclose to the public a mobile device vibration monitoring and alarm system based on a convolutional neural network, which can monitor the operating status of the mobile device in real time and issue an alarm in time when abnormal vibration occurs.

[0006] The technical solution of the present invention is achieved in this way: A mobile device vibration monitoring and alarm system based on convolutional neural network includes the following steps: Step 1: Collect abnormal vibration data of mobile devices, organize them into a data set, and divide the data set into a training set and a test set; Step 2: Build a convolutional neural network, and train, test and tune the convolutional neural network; Step 3: Collect vibration signals through the six-axis inertial measurement unit of MEMS and pre-process the collected data; Step 4: Perform short-time Fourier transform on the data, extract characteristic signals, and classify the data according to the time-frequency characteristics of the short-time Fourier transform; Step 5: Input the data processed in step 4 into the trained convolutional neural network, and the convolutional neural network adaptively extracts the abnormal vibration and outputs the abnormal state of the mobile device; Step 6: Use the GPS positioning module to monitor the location of the mobile device, and upload the abnormal status, abnormal data, and positioning information of the mobile device to the monitoring and alarm platform.

[0007] The measures to further optimize this technical solution are: As an improvement, the abnormal vibration signal is preprocessed, including noise reduction of the data, and the data is filtered using a Butterworth filter:

[0008] Where H(s) is the transfer function of the filter, s is the data to be processed, ωc is the cutoff frequency, and n is the order of the filter.

[0009] As an improvement, the convolutional neural network adopts an 8-layer structure, namely input layer In, pooling layer P0, convolution layer C1, pooling layer P1, convolution layer C2, pooling layer P2, fully connected layer F and output layer Out.

[0010] As an improvement, the convolution kernel size of the convolutional neural network is 4×4, and the time-frequency spectrum resolution of the convolutional neural network is 64×64.

[0011] As an improvement, the ReLU function is used as the activation function in the convolutional layer of the convolutional neural network.

[0012] As an improvement, Softmax is used as the activation function in the fully connected layer of the convolutional neural network.

[0013] As an improvement, in the convolutional neural network, a cross entropy function is used as a loss function.

[0014] As an improvement, in the step 2, 80% of the data set is used as a training set and 20% is used as a test set.

[0015] As an improvement, in the step 2, the convolutional neural network is converged to the optimal value by using the gradient descent method.

[0016] The advantages of the present invention compared with the prior art are: The present invention provides a mobile device vibration monitoring and alarm system based on a convolutional neural network. The system uses a sophisticated algorithm framework to perform in-depth analysis on vibration data collected from the device to achieve real-time monitoring of the device's operating status. The system can quickly upload the analyzed data to a cloud monitoring platform. Once an abnormal vibration occurs on the device, a real-time alarm mechanism can be triggered immediately. The GPS positioning module can be used to quickly obtain the device's accurate location information, thereby achieving abnormal vibration monitoring of the mobile device. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of abnormal vibration judgment of the convolutional neural network of the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below in conjunction with the accompanying drawings: The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be used for other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0019] Those skilled in the art should understand that, in the disclosure of the present invention, the directions or positions indicated by the terms "longitudinal", "lateral", "up", "down", "left", "right", "front", "back", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc. are based on the directions or positional relationships shown in the accompanying drawings, which are only simplified descriptions for the convenience of describing the present invention, and do not indicate or imply that the device or component referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, the above terms should not be understood as limitations on the present invention.

[0020] It is to be understood that the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element may be one, while in another embodiment, the number of the element may be multiple, and the term "one" should not be understood as a limitation on the quantity.

[0021] like Figure 1 As shown, a mobile device vibration monitoring and alarm system based on convolutional neural network includes the following steps: Step 1: Collect abnormal vibration data of mobile devices, organize them into a data set, and divide the data set into a training set and a test set. In the data set, 80% is used as a training set and 20% is used as a test set.

[0022] Step 2: Build a convolutional neural network and train, test and tune the convolutional neural network.

[0023] The convolutional neural network adopts an 8-layer structure, which includes input layer In, pooling layer P0, convolution layer C1, pooling layer P1, convolution layer C2, pooling layer P2, fully connected layer F and output layer Out. The convolution kernel size of the convolutional neural network is 4 × 4, and the time-frequency spectrum resolution of the convolutional neural network is 64 × 64. The ReLU function is used as the activation function in the convolutional layer of the convolutional neural network. Softmax is used as the activation function in the fully connected layer of the convolutional neural network. In the convolutional neural network, the cross entropy function is used as the loss function. When the convolutional neural network is trained, the convolutional neural network is converged to the optimal value by the gradient descent method.

[0024] Step 3: Collect vibration signals through the MEMS six-axis inertial measurement unit and pre-process the collected data; pre-process abnormal vibration signals, including noise reduction and filtering the data using a Butterworth filter:

[0025] Where H(s) is the transfer function of the filter, s is the data to be processed, ωc is the cutoff frequency, and n is the order of the filter.

[0026] Step 4: Perform short-time Fourier transform on the data, extract characteristic signals, and classify the data according to the time-frequency characteristics of the short-time Fourier transform; Step 5: Input the data processed in step 4 into the trained convolutional neural network, and the convolutional neural network adaptively extracts the abnormal vibration and outputs the abnormal state of the mobile device; Step 6: Use the GPS positioning module to monitor the location of the mobile device, and upload the abnormal status, abnormal data, and positioning information of the mobile device to the monitoring and alarm platform.

[0027] The six-axis inertial sensor of MEMS is mainly composed of three-axis acceleration sensors and three-axis gyroscopes. The accelerometer is used to measure the linear acceleration of an object in three-dimensional space, and the gyroscope is used to measure the angular velocity of an object in three-dimensional space. It can obtain the posture, position and velocity information of the object in three-dimensional space and provide the real-time motion status of the device. Route deviation, longitudinal and lateral swing angular velocity, longitudinal, lateral and vertical acceleration, vibration and other signals can be accurately collected and transmitted to the data bus through a standard interface. In addition, the six-axis inertial sensor of MEMS has a built-in attitude fusion algorithm based on extended Kalman filtering, which can effectively suppress the influence of dynamic interference on attitude accuracy and ensure the stability of attitude accuracy.

[0028] Short-time Fourier transform (STFT) is a joint time-frequency analysis method for time-varying and non-stationary signals. It can transform a one-dimensional abnormal vibration signal into a two-dimensional matrix suitable for convolutional neural network (CNN) processing: a characteristic spectrum containing time domain and frequency domain information.

[0029] Definition of Short-time Fourier Transform: The basic idea of ​​STFT is to use a fixed-length window function to intercept the time domain signal, and perform Fourier transform on the intercepted signal to obtain the local spectrum in a small time period near time t. By translating the window function on the entire time axis, the final transformation is to obtain a set of local spectra in each time period. Therefore, STFT is a two-dimensional function of time and frequency. The basic calculation formula is as follows: (1) Where: f(t) is the time domain signal, that is, the data value obtained by the six-axis inertial sensor; g(t-τ) is the time window centered at time τ. It can be seen that STFT is the Fourier transform of the signal f(t) multiplied by a window function g(t-τ) centered at τ.

[0030] Parameter selection: For STFT, there are two factors that affect its transformation effect: the window function type and the window width. Selecting a good window function can effectively reduce spectrum leakage and inter-spectral interference, and the width of the window affects the relative resolution of the time domain and the frequency domain: the wider the window, the higher the frequency resolution, and the narrower the window, the higher the time resolution. In this embodiment, the Hamming window function is used as the window function. Therefore, it is necessary to reasonably select the window function width according to the characteristics of the specific signal. The resolution of the time domain and the frequency domain can be calculated by the following formula: (2) when When is an even number, ;when When is an odd number, (3) Where T and F represent the time resolution and frequency resolution of the time-spectrogram, respectively, and [·] represents the rounding down operation. is the sample length involved in the STFT, is the window width, is the window overlap width, is the number of points involved in Fourier transform. According to the definition of image resolution, it is generally taken as equal , the resolution P of the time-frequency spectrum is defined as: (4) The value of P determines the amount of information contained in the time-frequency spectrum, which includes both useful abnormal vibration information and noise interference. Therefore, reasonable T and F can make the abnormal vibration signal significant and reduce noise interference.

[0031] In this mobile device vibration monitoring and alarm system, the fault characteristic frequency is an important reference indicator. For the vibration signal of the mobile device, the sample length is selected based on the first frequency of the fault characteristic frequency, and the window width of the STFT is selected based on the second frequency.

[0032] Convolutional neural network (CNN) is a feedforward neural network that has performed well in image recognition and target detection in recent years. It can handle overfitting problems well and enable larger-scale deep learning. The basic structure of a convolutional neural network mainly consists of convolutional layers, pooling layers, and fully connected layers.

[0033] In this mobile device vibration monitoring and alarm system, the input of CNN is the time-frequency spectrum. After passing through the convolution layer, pooling layer and fully connected layer, the output is the class label of the abnormal vibration type. The learning process of CNN is mainly divided into the forward propagation process and the reverse parameter update process.

[0034] Forward propagation phase: The forward propagation process of a convolutional neural network is mainly divided into three parts: convolutional layer, pooling layer, and fully connected layer.

[0035] Convolution layer: In the convolution layer, multiple convolution kernels are used to convolve with the input image, and after adding a bias, a series of feature maps can be obtained through an activation function. The mathematical expression of the convolution process is: (5) Where: is the jth element of the Lth layer, is the jth convolution region of the L-1 layer feature map, For the elements, is the weight matrix corresponding to the convolution kernel, is the bias term. As the activation function, the ReLU function is used as the activation function in this system, and the mathematical expression is: (6).

[0036] Pooling layer: The pooling layer is often connected after the convolution layer to reduce the dimension of the feature map while maintaining the invariance of the feature scale to a certain extent. Commonly used pooling methods include: Max pooling, Mean pooling, stochastic pooling, etc. The pooling layer only performs dimensionality reduction operations, has no parameters, and does not require weight updates.

[0037] In this system, at the pooling layer, the feature map output by the convolutional layer is pooled in each non-overlapping area of ​​size n × n, and the maximum value or average value in each area is selected (that is, maximum pooling or average pooling is used), and finally the output image is reduced by n times in both dimensions.

[0038] Fully connected layer: After the input image has been propagated alternately through multiple convolutional layers and pooling layers, the fully connected layer network is used to classify the extracted features. In the fully connected layer, the input is the one-dimensional feature vector of all feature maps expanded by weighted summation and then passed through the activation function: (7) Where: k is the sequence number of the network layer, is the output of the fully connected layer, is the expanded one-dimensional eigenvector, is the weight coefficient; is the bias term. As the activation function, the Softmax function is used as the activation function in this system. The Softmax function is an activation function suitable for classification tasks.

[0039] Reverse parameter update For a specific classification task, the training goal of CNN is to minimize the network's loss function, so it is very important to choose a suitable loss function. Common loss functions include mean square error function, cross entropy function, negative log-likelihood function, etc. This system uses the cross entropy function with better effect as the loss function, and the expression is as follows: (8) Where n is the number of samples of this type of fault, t is the predicted value, and y is the true value. During the training process, the method to minimize the loss function is the gradient descent method. By calculating the first-order partial derivative of the above cross entropy function, the learnable parameters w and b of the convolutional neural network can be updated layer by layer. (9) (10) Where: w' and b' are the updated weights and biases; w and b are the existing weights and biases; is the learning rate parameter, which is used to control the step size of weight update. If it is too large, the network will fall into a local optimum. If it is too small, the training time of the network will be increased. Take 2×10 -3 .

[0040] Abnormal vibration diagnosis process based on CNN: The abnormal vibration diagnosis method based on CNN can integrate signal preprocessing, abnormal vibration feature extraction and abnormal vibration pattern classification, and realize the specific process of adaptive extraction of abnormal vibration features and intelligent diagnosis. The collected vibration signal is divided into a training set and a test set after STFT. First, the training set is input into CNN for parameter learning, and the weight w and bias b are continuously updated using the gradient descent method. Then the trained parameters are applied to the test set to obtain the abnormal vibration diagnosis results of the equipment.

[0041] GPS positioning module: GPS positioning module is an existing technology. In addition to the basic device positioning function, this module also has track return (including direction, speed, mileage, longitude, latitude), ACC detection, anti-dismantling alarm, equipment overspeed alarm, electronic fence (allows users to pre-set a specific area or route. When the mobile device enters or leaves this area or deviates from the predetermined route, the system will automatically send an alarm message) and remote oil and power cut-off functions.

[0042] The mobile device vibration monitoring and alarm system can be applied to various mobile devices. Different mobile devices have different fault classifications. The fault types can be set when the convolutional neural network is constructed, and the convolutional neural network can be trained through the corresponding fault classification data set.

[0043] For example, when monitoring abnormal vibration of a car, five types of faults can be set. After the MEMS six-axis inertial sensor obtains the monitoring data, it is input into the trained convolutional neural network after short-time Fourier transform. The convolutional neural network outputs a matrix [A1, A2, A3, A4, A5], where A1 represents the axis misalignment of the six-axis inertial sensor, A2 represents the zero drift of the six-axis inertial sensor, A3 represents the engine failure of the car, A4 represents the tire problem of the car, and A5 represents the abnormality of the suspension system of the car; when the value is greater than 0.5, it is determined that an abnormality has occurred; for example, when the A3 value is greater than 0.5, it means that the car's engine has failed, and the GPS positioning module will send an alarm signal.

[0044] For example, when monitoring abnormal vibration of rollers, 7 types of faults can be set. After the six-axis inertial sensor of MEMS obtains the monitoring data, it is input into the trained convolutional neural network after short-time Fourier transform. The convolutional neural network outputs a matrix of [B1, B2, B3, B4, B5, B6, B7], where B1 represents the axis misalignment of the six-axis inertial sensor, B2 represents the zero drift of the six-axis inertial sensor, B3 represents the jamming of the roller, B4 represents the fracture of the roller, B5 represents the damage of the roller bearing, B6 represents the abnormal position of the roller, and B7 represents the wear of the roller surface. Similarly, when the value is greater than 0.5, it is determined that an abnormality has occurred. For example, when the B6 value is greater than 0.5, it means that the roller position is abnormal.

[0045] The number of abnormal classifications determines the number of coefficients in the matrix output by the convolutional neural network. Each coefficient corresponds to an abnormal situation. When it is greater than 0.5, the abnormal situation is considered to have occurred.

[0046] Feasibility and effectiveness verification: Dataset Description: The vibration signal of a mobile device is divided into 10 fault modes, which are the fault modes in normal state and abnormal vibration state. The vibration signal of each type of fault is divided into 400 samples, and a fault data set with 4,000 samples is established. The "one-hot" encoding method is used for each fault class label, that is, a vector is set, whose dimension is the same as the number of fault categories, except that a certain digit is 1 and the rest of the digits are 0. For example, if the class label of the normal vibration signal is set to 0, it can be encoded as [1,0,0,0,0,0,0,0,0,0].

[0047] According to the characteristic periods corresponding to the first and second frequencies, the sample length is determined to be 0.085 s (1024 sampling points) and the window width of the STFT is 0.01 s (128 sampling points).

[0048] STFT uses two different overlapping points, 64 and 114, to obtain two time-frequency spectra with different resolutions. From equations (2) to (4), it can be calculated that the resolutions are 65 × 15 and 65 × 65 respectively. Both time-frequency spectra are input into CNN for fault identification, and one of them is selected as input based on the accuracy rate.

[0049] Convolutional neural network model establishment: The selection of parameters such as the size and number of convolution kernels in the convolution layer, the size and number of pooling layers, and the number of neurons in the fully connected layer all affect the recognition effect of CNN.

[0050] Basic model: LetNel-5 model, its parameters are as follows, see Table 1, Table 1:

[0051] Where C represents the convolutional layer, P represents the pooling layer, and F represents the fully connected layer.

[0052] The number of convolutional layers C1 and C2, the size of the convolution kernel (same for C1 and C2), and the number of neurons in the fully connected layer F are taken as variables, and the value range is defined as {'C1': 2, 4, 8, 16, 32, 64; 'C2': 2, 4, 8, 16, 32, 64; 'F': 128, 256, 512, 1024, 2 048; 'Convolution kernel size': 2×2, 4×4, 5×5, 6×6;} These parameters can be combined into 576 CNN models, and the optimal network structure is determined in these combinations by taking into account both the recognition accuracy and the training time. During the experiment, it was found that removing the C3 layer in the LeNet-5 model and adding a pooling layer PO between the input layer and the convolution layer C1 can greatly reduce the training time while achieving a high fault recognition accuracy.

[0053] Taking the case with the highest accuracy, we determine the optimal network structure with a 65×65 input time-frequency spectrum, as shown in the following table. For an 8-layer CNN model "In-PO-C1-P1-C2-P2-F-Out", the convolution kernel size is 4×4, and the time-frequency spectrum resolution is modified to 64×64 (eliminating one row and one column), which are rounded to multiples of 2. This can reduce a certain amount of training time and has little effect on the final accuracy.

[0054] The structure of the convolutional neural network is as follows, see Table 2. Table 2:

[0055] The structure of the convolutional neural network of the present invention is compared with the 5-layer BP neural network structure of "In-H1-H2-F-Out" and the 7-layer convolutional neural network structure of "In-C1-P1-C2-P2-F-Out". In the 5-layer BP neural network structure, a one-dimensional time series is used as input data, and H1 and H2 represent hidden layers; in the 7-layer convolutional neural network structure, a time map is used as input data. The time map is a two-dimensional matrix obtained by rearranging the time series by rows. Here, 1024 sampling points are reconstructed into a 32×32 time map as input data.

[0056] The input data formats of the three network structures are set as follows, see Table 3. Table 3:

[0057] In the data set, there are 10 types of faults, each with 400 samples, a total of 4,000 samples, divided into a training set of 3,200 samples and a test set of 800 samples. The training process adopts the form of "packet capture", that is, a subset is captured from the training set at each step and input into the network for training, and after each training, the obtained parameters are applied to the test set, the identified fault class label is compared with the correct class label, and the accuracy of each subset is calculated as follows.

[0058] (11) Where: Accuracy is the fault identification accuracy, is the recognition result of the i-th sample in the subset. If the recognition is correct, it is equal to 1, and if the recognition is wrong, it is equal to 0. N is the number of samples in the subset.

[0059] Through experimental comparison, the convolutional neural network structure of the present invention has a higher accuracy rate than the 5-layer BP neural network structure with the structure of "In-H1-H2-F-Out" and the 7-layer convolutional neural network structure with the structure of "In-C1-P1-C2-P2-F-Out".

[0060] Those skilled in the art should understand that the embodiments of the present invention described above are only examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and the embodiments of the present invention may be deformed or modified in any way without departing from the principles.

Claims

1. A mobile device vibration monitoring and alarm system based on convolutional neural network, characterized by: The following steps are involved: Step 1: Collect abnormal vibration data of mobile devices, organize them into a data set, and divide the data set into a training set and a test set; Step 2: Build a convolutional neural network, and train, test and tune the convolutional neural network; Step 3: Collect vibration signals through the six-axis inertial measurement unit of MEMS and pre-process the collected data; Step 4: Perform short-time Fourier transform on the data, extract characteristic signals, and classify the data according to the time-frequency characteristics of the short-time Fourier transform; Step 5: Input the data processed in step 4 into the trained convolutional neural network, and the convolutional neural network adaptively extracts the abnormal vibration and outputs the abnormal state of the mobile device; Step 6: Use the GPS positioning module to monitor the location of the mobile device, and upload the abnormal status, abnormal data, and positioning information of the mobile device to the monitoring and alarm platform.

2. According to claim 1, a mobile device vibration monitoring and alarm system based on convolutional neural network is characterized in that: Preprocess the abnormal vibration signal, including noise reduction and filtering the data using a Butterworth filter: ； Where H(s) is the transfer function of the filter, s is the data to be processed, ωc is the cutoff frequency, and n is the order of the filter.

3. According to claim 1, a mobile device vibration monitoring and alarm system based on convolutional neural network is characterized in that: The convolutional neural network adopts an 8-layer structure, namely input layer In, pooling layer P0, convolution layer C1, pooling layer P1, convolution layer C2, pooling layer P2, fully connected layer F and output layer Out.

4. The mobile device vibration monitoring and alarm system based on convolutional neural network according to claim 3 is characterized in that: The convolution kernel size of the convolutional neural network is 4×4, and the time-frequency spectrum resolution of the convolutional neural network is 64×64.

5. The mobile device vibration monitoring and alarm system based on convolutional neural network according to claim 4 is characterized in that: The ReLU function is used as the activation function in the convolutional layer of the convolutional neural network.

6. The mobile device vibration monitoring and alarm system based on convolutional neural network according to claim 4 is characterized in that: Softmax is used as the activation function in the fully connected layer of the convolutional neural network.

7. The mobile device vibration monitoring and alarm system based on convolutional neural network according to claim 4 is characterized in that: In the convolutional neural network, a cross entropy function is used as a loss function.

8. The mobile device vibration monitoring and alarm system based on convolutional neural network according to claim 1, characterized in that: In the step 2, 80% of the data set is used as a training set and 20% is used as a test set.

9. The mobile device vibration monitoring and alarm system based on convolutional neural network according to claim 8, characterized in that: In the step 2, the convolutional neural network is converged to the optimal value by the gradient descent method.

Images