A high-resolution perception method for IoT data based on deep learning technology

Abstract

This invention presents a high-resolution sensing method for IoT data based on deep learning technology. First, based on assumed spatial and temporal information of the IoT sensing layer, the collected data undergoes spatial and temporal correlation processing to generate simulated sensing data. Second, simulated sensing data with different temporal and spatial correlations are combined to generate a dataset for model training. Finally, a high-resolution sensing model is constructed. During actual monitoring, a small number of sensor nodes are selected to form the IoT sensing layer to sense the target monitoring area, obtaining low-resolution sensing data. This low-resolution sensing data is preprocessed and then input into the trained high-resolution sensing model for optimization, resulting in high-resolution sensing data, thus completing the high-resolution sensing of IoT data. This method fully utilizes the temporal and spatial correlation between sensing data, achieving high-resolution sensing of the target monitoring area without changing the IoT sensing layer hardware, using a small number of sensor nodes and low-frequency data transmission.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet of Things data collection, and specifically is a high-resolution perception method for Internet of Things data based on deep learning technology. Background Art

[0002] The Internet of Things (IoT) is the third wave of information technology development since the emergence of computers and the Internet. Its applications have spread to many fields. The IoT device layer consists of a large number of sensor nodes capable of sensing and communicating. These sensor nodes are deployed in the monitoring area and can monitor multiple aspects of information in real time. For example, in environmental information monitoring, multiple types of sensor nodes can simultaneously sense temperature, humidity, light, voltage and other multi-attribute data. With the increase in the number of IoT application scenarios, the requirements for perception resolution are gradually increasing. High-resolution perception means more sensor nodes and more frequent data transmission, but also requires higher construction costs, network energy consumption and energy supply.

[0003] There are two ways to achieve high-resolution perception of IoT data. One is to upgrade the hardware equipment, that is, to add sensor nodes and increase the frequency of data transmission so that the hardware equipment can meet the requirements of efficient data collection and transmission. The other is to perform super-resolution processing on the low-resolution perception data collected by the IoT perception layer through algorithms without upgrading the hardware equipment to obtain high-resolution perception data. The existing technology mainly achieves high-resolution perception by upgrading hardware equipment, which not only increases the deployment cost but also increases resource consumption. Therefore, this application proposes a high-resolution perception method for IoT data based on deep learning technology. Without upgrading the hardware equipment, it uses a small number of sensor nodes and low-frequency data transmission to achieve high-resolution perception of IoT data. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the technical problem that the present invention intends to solve is to provide a high-resolution perception method for Internet of Things data based on deep learning technology.

[0005] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0006] A method for high-resolution perception of IoT data based on deep learning technology, characterized by comprising the following steps:

[0007] Step 1: Assume that the IoT perception layer contains n sensor nodes uniformly and randomly deployed in the monitoring area. After a perception cycle T = mΔT, n×m pieces of data are collected and transmitted to the base station. These data form a perception data matrix Among them, ΔT represents the time slot length of a single perception in the perception cycle, m represents the number of perceptions contained in the perception cycle, represents the field of real numbers;

[0008] Step 2: According to the assumed spatial and temporal information of the IoT perception layer, the data collected by the IoT perception layer is processed in spatial correlation and temporal correlation to generate simulated perception data with assumed spatial and temporal correlation;

[0009] Step 2.1, spatial correlation processing: Let p = (x, y) be the position coordinates of any sensor node in the two-dimensional spatial domain D of the monitoring area, randomly select a sensor node as the first sensor node, and the position coordinates are marked as p1 = (x1, y1). The Euclidean distance d between each sensor node in the two-dimensional spatial domain D and the first sensor node is calculated by formula (2);

[0010]

[0011] Step 2.2, time correlation processing: Arrange the Euclidean distances between each sensor node and the first sensor node in the two-dimensional spatial domain D from small to large, and map the two-dimensional spatial domain to the one-dimensional spatial domain C = [n×1]; Generate an independently distributed random Gaussian field at time t = t0 = 0, which represents the number field composed of the collected data of each sensor node at time t0 in the one-dimensional spatial domain; Generate random Gaussian fields at different perception time points according to the perception time sequence, and then add time correlation to the random Gaussian field to obtain a random Gaussian field with time correlation ω(p,t i ),ω(p,t i-1 ) represent the random Gaussian fields at the i-th and i-1-th perception time points, i∈m, ε(p,t i ) represents random noise, γ t ∈(0,1) is the time correlation coefficient;

[0012] Perform a two-dimensional Fourier transform on the random Gaussian field with time correlation to obtain a random Gaussian field with time correlation after Fourier transform; then, calculate the spatial correlation between the sensor nodes in the random Gaussian field and perform a two-dimensional Fourier transform, and then multiply it with the random Gaussian field with time correlation after Fourier transform to obtain the frequency domain expression of the random Gaussian field; finally, convert the frequency domain expression of the random Gaussian field into the time domain expression through the two-dimensional inverse Fourier transform to obtain the perception data of different perception time points with time and space correlation; put the perception data of different perception time points into a matrix according to the perception time order to obtain the simulated perception data;

[0013] The expression for calculating the spatial correlation between sensor nodes is:

[0014] r s (d) = exp - γ s α 2 d 2 (3)

[0015] Among them, γ s ∈[0,1] represents the spatial correlation coefficient between two adjacent sensor nodes, which is an artificially set value; α represents the scale factor of the field size, which is an artificially set value;

[0016] Step 3: prepare a data set for model training; combine simulation perception data with different time correlations and spatial correlations to generate high-resolution simulation perception data; perform equidistant downsampling on the high-resolution simulation perception data to obtain low-resolution perception data. The high-resolution and low-resolution simulation perception data together constitute a data set for model training;

[0017] Step 4: construct a high-resolution perception model based on a convolutional neural network, use the data set obtained in step 3 to train the high-resolution perception model, and use the trained high-resolution perception model to generate high-resolution perception data; in the actual monitoring process, select k (k∈n) sensor nodes to form an Internet of Things perception layer to perceive the target monitoring area, and obtain low-resolution perception data within the perception cycle. After preprocessing, the low-resolution perception data is input into the trained high-resolution perception model for optimization processing to obtain high-resolution perception data;

[0018] At this point, high-resolution perception of IoT data based on deep learning technology has been completed.

[0019] Furthermore, in step 3, the IoT perception layer collects data with days, hours, minutes, and seconds as perception time slot units, obtains perception data of different perception time slot units, and calculates the time correlation coefficient using formula (4);

[0020]

[0021] Where z(p,t1) and z(p,t2) represent the data collected at the sensing time points t1 and t2 respectively; cov(·) represents the covariance, σ z (·) represents variance;

[0022] The value range of the spatial correlation coefficient is divided into multiple intervals, and the data of different spatial correlation coefficient intervals and different perception time slot units are combined to obtain high-resolution simulation perception data.

[0023] Furthermore, in step 4, the low-resolution perception data is upsampled and preprocessed by bicubic interpolation. The high-resolution perception model contains nine convolutional layers, and the first eight convolutional layers are connected to a ReLU activation function. The first convolutional layer has 1 input channel, 64 output channels, and a convolution kernel size of 9. The second convolutional layer has 64 input channels, 32 output channels, and a convolution kernel size of 5. The third convolutional layer has 32 input channels, 16 output channels, and a convolution kernel size of 1. The fourth convolutional layer has 1 input channel, 64 output channels, and a convolution kernel size of 1. 16, the output channel number is 16, and the convolution kernel size is 3; the input channel number of the fifth convolution layer is 16, the output channel number is 16, and the convolution kernel size is 3; the input channel number of the sixth convolution layer is 16, the output channel number is 16, and the convolution kernel size is 3; the input channel number of the seventh convolution layer is 16, the output channel number is 16, and the convolution kernel size is 3; the input channel number of the eighth convolution layer is 16, the output channel number is 32, and the convolution kernel size is 1; the input channel number of the ninth convolution layer is 32, the output channel number is 1, and the convolution kernel size is 5.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. The present invention introduces deep learning technology into high-resolution perception of IoT environmental information. The high-resolution perception model extracts spatiotemporal related information and fully utilizes the spatiotemporal correlation of IoT data collected during the perception cycle. Without changing the hardware equipment of the IoT perception layer, it uses a small number of sensor nodes and low-frequency data transmission to achieve high-resolution perception of the target monitoring area. Compared with the commonly used bicubic interpolation method, the perception error of the method of the present invention is smaller, and high-resolution perception is achieved for various information in various scenarios.

[0026] 2. In order to solve the problem of lack of real IoT data, the present invention generates high-resolution simulated perception data with assumed spatiotemporal correlation by setting the spatial information and temporal information of the assumed IoT perception layer. These simulated perception data constitute the data set for training the model. The method can directly generate high-resolution simulated perception data using the temporal correlation coefficient and the spatial correlation coefficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall flow chart of the present invention;

[0028] Figure 2(a) shows the low-resolution perception visualization result of indoor temperature;

[0029] Figure 2(b) shows the visualization result of indoor temperature perception with double the resolution using bicubic interpolation;

[0030] FIG2( c ) is a visualization result of the indoor temperature using the method of the present invention with double-high resolution perception;

[0031] Figure 2(d) shows the visualization result of the real high-resolution perception of indoor temperature;

[0032] Figure 3(a) shows the low-resolution perception visualization result of forest humidity;

[0033] Figure 3(b) shows the visualization result of forest humidity using bicubic interpolation for three times higher resolution perception;

[0034] FIG3( c ) is a visualization result of forest humidity using the method of the present invention for triple high-resolution perception;

[0035] Figure 3(d) shows the visualization result of the real high-resolution perception of forest humidity. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present application is not limited thereto.

[0037] The present invention is a method for high-resolution perception of IoT data based on deep learning technology, comprising the following steps:

[0038] Step 1: Assume that the IoT perception layer for sensing the monitoring area includes n sensor nodes, which are uniformly and randomly deployed in the monitoring area for data collection. T is the total time length of a sensing cycle, ΔT is the time slot length of a single sensing in a sensing cycle, and m is the number of sensing times in a sensing cycle. After a sensing cycle T = mΔT, a total of n×m pieces of data are collected and transmitted to the base station. These collected data form a sensing data matrix M.

[0039]

[0040] Among them, f(N n ,ΔT m ) represents the nth sensor node N n The data collected during the mth sensing process, represents the field of real numbers;

[0041] Step 2: According to the assumed spatial and temporal information of the IoT perception layer, the data collected by the IoT perception layer is processed in spatial correlation and temporal correlation to generate simulated perception data with assumed spatial and temporal correlation;

[0042] Step 2.1, spatial correlation processing: The two-dimensional spatial domain of the monitoring area is recorded as D = [-x D ,x D ]×[-yD ,y D ], -x D 、x D Respectively represent the starting point and ending point of the horizontal coordinate of the two-dimensional space domain D, -y D ,y D Respectively represent the starting point and the ending point of the vertical coordinate of the two-dimensional space domain; let p = (x, y) represent the position coordinate of any sensor node in the two-dimensional space domain D, randomly select a sensor node as the first sensor node, the position coordinate is marked as p1 = (x1, y1), and calculate the Euclidean distance d between each sensor node and the first sensor node in the two-dimensional space domain D by formula (2);

[0043]

[0044] The spatial correlation between sensor nodes is calculated by formula (3);

[0045] r s (d) = exp - γ s α 2 d 2 (3)

[0046] Among them, γ s ∈[0,1] represents the spatial correlation coefficient between two adjacent sensor nodes, artificially set value, γ s The larger the value, the greater the spatial correlation between sensor nodes, and the denser the sensor nodes in the two-dimensional spatial domain D. On the contrary, the smaller the spatial correlation, the sparser the sensor nodes. α represents the scale factor of the field size, which is an artificially set value.

[0047] Step 2.2, time correlation processing: Arrange the Euclidean distances between each sensor node and the first sensor node in the two-dimensional spatial domain D from small to large, and map the two-dimensional spatial domain D to the one-dimensional spatial domain C = [n×1]; at time t = t0 = T0 = mΔT = 0, (m = 0), generate an independently distributed random Gaussian field, defined as It is a Gaussian random variable, which represents the number field composed of the collected data of each sensor node at time t0 in the one-dimensional space domain C. According to the perception time sequence, the random Gaussian field ω(p,t i ), i∈m, add time correlation to the random Gaussian field, and we get ε(p,t i ) represents random noise; γ t ∈(0,1) is the time correlation coefficient, which indicates the time correlation between the perception data at different time points, γ t The closer it is to 1, the greater the temporal correlation is, that is, the more times the perception is performed in a single perception cycle, and vice versa;

[0048] Perform a two-dimensional Fourier transform on the random Gaussian field with added time correlation to obtain the random Gaussian field after Fourier transform represents the two-dimensional Fourier transform; then, the spatial correlation between the sensor nodes in the random Gaussian field is calculated by equation (3) and a two-dimensional Fourier transform is performed, and then the spatial correlation between the sensor nodes in the random Gaussian field W(p,t i ) to obtain the frequency domain expression of the random Gaussian field Finally, the frequency domain expression of the random Gaussian field is converted into the time domain expression through the two-dimensional inverse Fourier transform to obtain the perception data of different perception time points with spatiotemporal correlation. Since the simulation requires a two-dimensional matrix composed of the perception data of all perception time points in the entire perception cycle, the perception data of different perception time points are put into the matrix in the order of perception time to obtain the simulation perception data z(p,t);

[0049] Step 3: prepare a data set for model training; combine simulation perception data with different time correlations and spatial correlations to generate high-resolution simulation perception data; perform equidistant downsampling on the high-resolution simulation perception data to obtain low-resolution perception data. The high-resolution and low-resolution simulation perception data together constitute a data set for model training;

[0050] The IoT perception layer collects data in perception time slots of day, hour, minute, and second, obtains the perception data of the perception time slots in units of day, hour, minute, and second, and calculates the time correlation coefficient using formula (4). The time correlation coefficients corresponding to different perception time slot units in this example are shown in Table 1;

[0051]

[0052] In the formula, cov(·) represents the covariance, σ z (·) represents variance;

[0053] Table 1 Time correlation coefficients corresponding to different perception time slot units

[0054]

[0055] The value range of the spatial correlation coefficient is evenly divided into five intervals, namely [0.1-0.2], [0.2-0.4], [0.4-0.6], [0.6-0.8] and [0.8-1]. In order to be more in line with the actual perception scenario, the last four spatial correlation coefficient intervals and the perception time slot units of hours, minutes and seconds are selected and combined to obtain a data set with different spatiotemporal correlations. Therefore, the data set contains four groups of data. The first group simulates the indoor environment perception scenario with a high density of sensor nodes, and the spatial correlation coefficient is [0.8-1]. 15 groups of perception data in hours, minutes and seconds are selected respectively, and a total of 45 groups of high-resolution simulation perception data are obtained. ; The second group simulates outdoor environment perception scenarios with high density of sensor node deployment, with a spatial correlation coefficient of [0.6-0.8]. 15 groups of perception data in hours, minutes, and seconds are selected, and a total of 45 groups of high-resolution simulation perception data are obtained; the third group simulates outdoor environment perception scenarios with low density of sensor node deployment. Such scenarios usually do not collect data in seconds, so 15 groups of perception data in minutes and hours are selected, and a total of 30 groups of high-resolution simulation perception data are obtained; the fourth group simulates large-scale outdoor environment perception scenarios with extremely low density of sensor node deployment, and 5 groups of perception data in hours and days are selected, and a total of 10 groups of high-resolution simulation perception data are obtained. According to the current application frequency in the field of the Internet of Things, the first and second scenarios are more frequently used, so they account for a higher proportion; the third and fourth scenarios are less used and account for a smaller proportion.

[0056] In summary, the composition of the dataset is shown in Table 2.

[0057]

[0058]

[0059] By performing equal-interval downsampling on the simulated high-resolution perception data generated above, simulated low-resolution perception data is obtained. The simulated high-resolution perception data and the simulated low-resolution perception data together constitute a data set for high-resolution perception model training.

[0060] Step 4: pre-process the low-resolution perception data collected during the actual monitoring process, and then optimize the pre-processed data through a high-resolution perception model to obtain high-resolution perception data;

[0061] In the actual application process, k (k∈n) sensor nodes are selected to form the IoT perception layer to perceive the target monitoring area and obtain low-resolution perception data M within a perception cycle. LR , the perception data itself has temporal and spatial correlation; the low-resolution perception data M LRFirst, the low-resolution perception data is upsampled and preprocessed by bicubic interpolation to obtain the upsampled low-resolution perception data. Then upsample the low-resolution perception data Input into the trained high-resolution perception model for optimization processing to obtain high-resolution perception data M HR ;

[0062] The high-resolution perception model contains nine convolutional layers. The first eight convolutional layers are followed by a ReLU activation function. The parameters of the high-resolution perception model are shown in Table 3.

[0063] Table 3 Parameters of high-resolution perception model

[0064]

[0065] In order to verify the effectiveness of the method of the present invention in the data perception process of the Internet of Things, the bicubic interpolation method is selected as a comparison, and the multi-parameter data collected by the Berkeley Research Laboratory and the GreenOrbs project are selected as test data; since data loss in the data acquisition process of the Internet of Things perception layer is inevitable, a small part of the complete perception data subset is selected as the real test data; specifically, the original data collected by the Berkeley Research Laboratory contains three parameter data of indoor temperature, humidity and voltage collected by 48 sensor nodes in 138 perception time slots in one perception cycle; these data are processed by 2 times super resolution, that is, the data of 69 perception time slots of 24 sensor nodes in one perception cycle are selected for reconstruction, and the visualization results of indoor temperature are shown in Figures 2(a) to (d); the perception error is represented by the normalized mean absolute error (NMAE), and the perception error comparison results in Table 4 are obtained.

[0066] Table 4 Comparison of perception errors of three parameter data: indoor temperature, humidity and voltage

[0067] Parameter Type Perceiving data size Super resolution multiple NMAE (bicubic interpolation) NMAE (method of the present invention) Indoor temperature 48*138 2 0.150479 0.065671 Indoor humidity 48*138 2 0.138443 0.067211 Indoor voltage 48*138 2 0.141127 0.062224

[0068] The original data of the GreenOrbs project includes three parameter data of forest temperature, humidity and light collected by 132 sensor nodes in 129 sensing time slots in a sensing cycle. These data are processed with 3 times higher resolution, that is, the data of 44 sensor nodes in 43 sensing time slots in a single sensing cycle are selected for reconstruction. The visualization results of forest humidity are shown in Figures 3(a) to (d), and the comparison results of perception errors are shown in Table 5.

[0069] Table 5 Comparison of perception errors of three parameter data: indoor temperature, humidity and voltage

[0070] Parameter Type Perceiving data size Super resolution multiple NMAE (bicubic interpolation) NMAE (method of the present invention) Forest Temperature 132*129 3 0.164087 0.047715 Forest humidity 132*129 3 0.167969 0.061520 Forest Light 132*129 3 0.215147 0.095578

[0071] It can be seen from Figures 2, 3 and Tables 4, 5 that compared with the commonly used bicubic interpolation method, the method of the present invention achieves high-resolution perception of various types of IoT data, and the perception error is significantly smaller than that of the bicubic interpolation method. There are two reasons for the small perception error: first, the spatiotemporal correlation between the perception data is effectively utilized with the help of the convolutional neural network; second, the model is trained by constructing simulated perception data, so that the model can fully extract the spatiotemporal correlation characteristics of the perception data and fully utilize the spatiotemporal correlation of the perception data for reconstruction.

[0072] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A high-resolution perception method for IoT data based on deep learning technology, characterized in that: The following steps are involved: Step 1: Assume that the IoT perception layer contains n sensor nodes uniformly and randomly deployed in the monitoring area. After a perception cycle T = mΔT, n×m pieces of data are collected and transmitted to the base station. These data form a perception data matrix Among them, ΔT represents the time slot length of a single perception in the perception cycle, m represents the number of perceptions contained in the perception cycle, represents the field of real numbers; Step 2: According to the assumed spatial and temporal information of the IoT perception layer, the data collected by the IoT perception layer is processed in spatial correlation and temporal correlation to generate simulated perception data with assumed spatial and temporal correlation; Step 2.1, spatial correlation processing: Let p = (x, y) be the position coordinates of any sensor node in the two-dimensional spatial domain D of the monitoring area, randomly select a sensor node as the first sensor node, and the position coordinates are marked as p1 = (x1, y1). The Euclidean distance d between each sensor node in the two-dimensional spatial domain D and the first sensor node is calculated by formula (2); Step 2.2, time correlation processing: Arrange the Euclidean distances between each sensor node and the first sensor node in the two-dimensional spatial domain D from small to large, and map the two-dimensional spatial domain to the one-dimensional spatial domain C = [n×1]; Generate an independently distributed random Gaussian field at time t = t0 = 0, which represents the number field composed of the collected data of each sensor node at time t0 in the one-dimensional spatial domain; Generate random Gaussian fields at different perception time points according to the perception time sequence, and then add time correlation to the random Gaussian field to obtain a random Gaussian field with time correlation ω(p,t i ),ω(p,t i-1 ) represent the random Gaussian fields at the i-th and i-1-th perception time points, i∈m, ε(p,t i ) represents random noise, γ t ∈(0,1) is the time correlation coefficient; Perform a two-dimensional Fourier transform on the random Gaussian field with time correlation to obtain a random Gaussian field with time correlation after Fourier transform; then, calculate the spatial correlation between the sensor nodes in the random Gaussian field and perform a two-dimensional Fourier transform, and then multiply it with the random Gaussian field with time correlation after Fourier transform to obtain the frequency domain expression of the random Gaussian field; finally, convert the frequency domain expression of the random Gaussian field into the time domain expression through the two-dimensional inverse Fourier transform to obtain the perception data of different perception time points with time and space correlation; put the perception data of different perception time points into a matrix according to the perception time order to obtain the simulated perception data; The expression for calculating the spatial correlation between sensor nodes is: r s (d)=exp-γ s a 2 d 2 (3) Among them, γ s ∈[0,1] represents the spatial correlation coefficient between two adjacent sensor nodes, which is an artificially set value; α represents the scale factor of the field size, which is an artificially set value; Step 3: prepare a data set for model training; combine simulation perception data with different time correlations and spatial correlations to generate high-resolution simulation perception data; perform equidistant downsampling on the high-resolution simulation perception data to obtain low-resolution perception data. The high-resolution and low-resolution simulation perception data together constitute a data set for model training; Step 4: Build a high-resolution perception model based on a convolutional neural network, use the data set obtained in step 3 to train the high-resolution perception model, and use the trained high-resolution perception model to generate high-resolution perception data; in the actual monitoring process, select k (k∈n) sensor nodes to form the Internet of Things perception layer to perceive the target monitoring area, and obtain low-resolution perception data within the perception cycle. After preprocessing, the low-resolution perception data is input into the trained high-resolution perception model for optimization processing to obtain high-resolution perception data; At this point, high-resolution perception of IoT data based on deep learning technology has been completed.

2. The method for high-resolution perception of IoT data based on deep learning technology according to claim 1 is characterized in that: In step 3, the IoT perception layer collects data with days, hours, minutes, and seconds as the perception time slot units, obtains the perception data of different perception time slot units, and calculates the time correlation coefficient using formula (4); Where z(p,t1) and z(p,t2) represent the data collected at the sensing time points t1 and t2 respectively; cov(·) represents the covariance, σ z (·) represents variance; The value range of the spatial correlation coefficient is divided into multiple intervals, and the data of different spatial correlation coefficient intervals and different perception time slot units are combined to obtain high-resolution simulation perception data.

3. The method for high-resolution perception of IoT data based on deep learning technology according to claim 1 or 2, characterized in that: In step 4, the low-resolution perception data is upsampled and preprocessed by bicubic interpolation. The high-resolution perception model contains nine convolutional layers, and the first eight convolutional layers are connected to a ReLU activation function. The first convolutional layer has 1 input channel, 64 output channels, and a convolution kernel size of 9. The second convolutional layer has 64 input channels, 32 output channels, and a convolution kernel size of 5. The third convolutional layer has 32 input channels, 16 output channels, and a convolution kernel size of 1. The fourth convolutional layer has 16 input channels. , the number of output channels is 16, and the convolution kernel size is 3; the number of input channels of the fifth convolution layer is 16, the number of output channels is 16, and the convolution kernel size is 3; the number of input channels of the sixth convolution layer is 16, the number of output channels is 16, and the convolution kernel size is 3; the number of input channels of the seventh convolution layer is 16, the number of output channels is 16, and the convolution kernel size is 3; the number of input channels of the eighth convolution layer is 16, the number of output channels is 32, and the convolution kernel size is 1; the number of input channels of the ninth convolution layer is 32, the number of output channels is 1, and the convolution kernel size is 5.

Images