3D atmospheric pollution fusion and analysis method based on ResNet network for stereo telemetry data

Abstract

The present invention discloses a method for fusion and analysis of 3D atmospheric pollution from stereo telemetry data based on a ResNet network. The method comprises at least the following steps: S1: preprocessing the collected air quality data; S2: performing a gradient node generation and degree generation phase; and S3: performing model training and testing. The present invention relates to the field of firefighting equipment technology, and more specifically, to a method for fusion and analysis of 3D atmospheric pollution from stereo telemetry data based on a ResNet network.

Description

Technical Field

[0001] The present invention relates to the technical field of air quality forecasting, and in particular to a 3D atmospheric pollution fusion and analysis method of stereo telemetry data based on a ResNet network. Background Art

[0002] Environmental protection has become a key concern for my country's livelihood projects. Protecting the environment not only affects people's living environment but also affects economic development, and is of great significance. However, with the continuous acceleration of industrialization and urbanization, the frequent occurrence of severe atmospheric pollution has brought significant impacts on human health, atmospheric visibility, climate, and other aspects. Currently, the more mainstream greenhouse gas monitoring technologies in my country use non-dispersive infrared, tunable semiconductor laser, Fourier infrared, infrared gas filter correlation method and other methods to monitor greenhouse gases in the air; ammonia monitoring technology fully utilizes the characteristic that ammonia is the only high-concentration alkaline gas in the atmosphere. By monitoring and analyzing secondary particulate matter such as sulfate and nitrate produced by the reaction of ammonia escaping into the atmosphere with acidic gases such as nitric acid or sulfuric acid, it can assess and predict air quality; regional carbon sink inversion numerical simulation technology has strengthened the research on assimilation inversion models of greenhouse gas concentrations and emissions, and clarified the temporal changes and spatial distribution characteristics of carbon sources and carbon sinks.

[0003] However, existing domestic and international air quality forecasting technologies generally have the problem that air quality prediction is affected by instrument performance, power maintenance capabilities, etc. When the instrument fails or the power is cut off, it is impossible to collect the current air quality, so it is very necessary to adopt the ResNet network. Summary of the Invention

[0004] In view of the above situation, in order to make up for the above existing defects, this solution provides a 3D atmospheric pollution fusion and analysis method of stereo telemetry data based on the ResNet network.

[0005] The technical solution adopted by the present invention is as follows: a 3D atmospheric pollution fusion and analysis method based on stereo telemetry data of ResNet network, comprising at least the following steps:

[0006] S1: Preprocess the collected air quality data. Due to the large size of the collected data set and the large differences between the data, the impact of these differences will make the model fitting process difficult. Therefore, these raw data sets need to be preprocessed. Common preprocessing operations such as data normalization are used to control the raw data to a smaller specific range, thereby better training and fitting the deep learning model. A shortcut connection method used by the residual structure:

[0007]

[0008] S2: Gradient generation node. In the degree generation phase, we first use a ResNet3D network pre-trained on kinetic data to fine-tune it on the training dataset so that the ResNet3D network has preliminary recognition capabilities on the dataset. Then, we fix the network parameters and use the network prediction label to derive the input:

[0009]

[0010] The gradient generation process is based on a network that already has basic resolving power. This network is used to derive the input and normalize the derived derivatives. In this way, for each spatial position of the input video, the normalized gradient value corresponding to that position is obtained. These gradient information will then be used to guide the feature enhancement module. The specific process is as follows:

[0011] First, a network pre-trained on the Kinetics dataset is fine-tuned on the specified dataset to enable the network to have the most basic recognition ability on the experimental dataset. The currently trained network is defined as f, which has a preliminary resolution ability for the specified dataset. Then f represents the input image x a The mapping relationship between and image label y:

[0012] y=f(x a )

[0013] Among them, y represents the predicted label of the image segment by the fine-tuned network.

[0014] Fix the trained network parameters and use the predicted output of the final network to derive the input image and generate gradients. Define the gradient value of the network for a pixel in the video as i, j, T. The expressions of i, j, T are as follows:

[0015]

[0016] in, Represents the value of the coordinate point (i, j, T) in the image,

[0017] The generated gradient is then normalized using the normalization formula:

[0018]

[0019] The final θ generated by the normalized gradient value of the position in the original image is a matrix of the same size as the input image, in which the elements are the normalized gradient values ​​of each network for the pixel point at that position in the original image.

[0020] S3: Perform model training and testing. After the model is defined, the image dataset is divided into a training set, a validation set, and a test set in a 6:2:2 ratio. The initial deep learning model is trained to obtain prediction results. After calculating the error using the loss function, the optimizer adjusts the learning model parameters to improve data accuracy.

[0021] The ResNet3D network is divided into two channels. The first channel is used to perform feature selection on environmental data, filtering out the implicit features inherent in the environmental data that have the greatest impact on pollutant emissions, and providing these features to the second channel as parameter variables for the pollutant prediction model. The second channel is used to predict pollutants, that is, to predict the detection results of unknown road sections based on the pollutant detection data at known points.

[0022] Among them, the 3D spatial feature learning of the ResNet network includes the ResNet3D network deep network, which can be understood as a feature extractor. The deep network extracts different features from the shallow layer to the deep layer, and finally sends the features to a classifier or performs other processing through different combinations. In the process of feature extraction of the deep neural network, the overall feature generation process is generated as the network deepens. The input of each layer is the feature of all the outputs of the previous layer. This layer needs to process all this information. These features are global features of the input image.

[0023] Among them, the data collection uses the Kinetics dataset, and the ResNet3D network is pre-trained on the Kinetics dataset.

[0024] RestNet, as the underlying framework, implements convolutional neural networks and performs end-to-end training for tasks such as image classification and object detection. ResNet, which employs the design concept of residual networks and builds a deep network by stacking multiple residual blocks, effectively alleviates the vanishing gradient problem of deep neural networks and improves network performance.

[0025] Among them, a deep neural network is used to conduct in-depth analysis of the data, which can model the uncertainty part (model representation error) and the random part (model parameter error) in the model inversion data, that is, to mine the inherent change rules from the observation data of limited points (urban observations, vehicle-mounted observations). Specifically, the data change rules of different locations in the same period are mined from urban observation data. In the process of deep data mining, we will input geographic information, meteorological information and other data that may affect the observation results into the network for comprehensive analysis. It is not difficult to see that the deep neural network contains hundreds of thousands or even millions of network parameters. Its model representation capability far exceeds the existing modeling methods, and it can use all the existing information in a timely and efficient manner to define a state of atmospheric motion with the greatest possible accuracy.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. The present invention adopts the 3D spatial feature learning of the ResNet network. The input and output of ResNet are usually images or image features. The specific input and output sizes and number of channels depend on the specific network structure and task. In ResNet, the input image first passes through a convolutional layer and a pooling layer, then passes through multiple residual modules, and finally outputs the final classification result through global average pooling and a fully connected layer. It can train very deep neural networks, avoid the gradient vanishing problem, and improve the expression ability and performance of the model; the use of residual connections can retain the original features, making the network learning smoother and more stable, further improving the accuracy and generalization ability of the model; it can avoid the gradient vanishing and gradient explosion problems during training, and accelerate network convergence.

[0028] 2. The present invention performs model training and testing. After the model is defined, the image dataset is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The initial deep learning model is trained to obtain prediction results. After calculating the error using a loss function, the parameters of the learning model are adjusted using an optimizer to improve data accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0030] Figure 1 This is a diagram showing the regression learning principle of the ResNet network-based 3D atmospheric pollution fusion and analysis method for stereo telemetry data provided by the present invention;

[0031] Figure 2This is a ResNet data sample preprocessing diagram of the ResNet network-based stereo telemetry data 3D atmospheric pollution fusion and analysis method provided by the present invention. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] like Figures 1 to 2 As shown, the technical solution adopted by the invention is as follows: a 3D atmospheric pollution fusion and analysis method based on stereo telemetry data of ResNet network, comprising at least the following steps:

[0034] S1: First, preprocess the collected air quality data. Since the collected data set is large in scale and the differences between each data are also large, the impact of these differences will make the model fitting process difficult. Therefore, these original data sets need to be preprocessed. Use common preprocessing operations such as data normalization to control the original data to a smaller specific range, thereby better training and fitting the deep learning model.

[0035] A shortcut connection method used by the residual structure:

[0036]

[0037] S2: Gradient generation node. In the degree generation phase, we first use a ResNet3D network pre-trained on kinetic data to fine-tune it on the training dataset so that the ResNet3D network has preliminary recognition capabilities on the dataset. Then, we fix the network parameters and use the network prediction label to derive the input:

[0038]

[0039] The gradient generation process is based on a network that already has basic resolving power. This network is used to derive the input and normalize the derived derivatives. In this way, for each spatial position of the input video, the normalized gradient value corresponding to that position is obtained. These gradient information will then be used to guide the feature enhancement module. The specific process is as follows:

[0040] First, a network pre-trained on the Kinetics dataset is fine-tuned on the specified dataset to enable the network to have the most basic recognition ability on the experimental dataset. The currently trained network is defined as f, which has a preliminary resolution ability for the specified dataset. Then f represents the input image x a The mapping relationship between and image label y:

[0041] y=f(x a )

[0042] Among them, y represents the predicted label of the image segment by the fine-tuned network.

[0043] Fix the trained network parameters and use the predicted output of the final network to derive the input image and generate gradients. Define the gradient value of the network for a pixel in the video as i, j, T. The expressions of i, j, T are as follows:

[0044]

[0045] in, Represents the value of the coordinate point (i, j, T) in the image,

[0046] The generated gradient is then normalized using the normalization formula:

[0047]

[0048] The final θ generated by the normalized gradient value of the position in the original image is a matrix of the same size as the input image, in which the elements are the normalized gradient values ​​of each network for the pixel point at that position in the original image.

[0049] S3: Perform model training and testing. After the model is defined, the image dataset is divided into a training set, a validation set, and a test set in a 6:2:2 ratio. The initial deep learning model is trained to obtain prediction results. After calculating the error using the loss function, the optimizer adjusts the learning model parameters to improve data accuracy.

[0050] Among them, the S3ResNet data sample preprocessing steps are:

[0051] 1. Based on the collected data, a data feature analysis tool was developed to automatically analyze the characteristics of different types of pollutants at each site;

[0052] 2. Based on the statistical characteristics of the radar echo signal, the standard deviation is calculated to set an appropriate threshold to identify and delete outliers. Outliers are identified and processed through information processing and feature extraction methods such as wavelet transform. Abnormal data exceeding the threshold is eliminated.

[0053] 3. The dataset's time range can be updated in real time, up to the day before the forecast, or even up to the hour before. The collected dataset is divided into three parts: training, validation, and test sets. The training set is used to train the model (fitting parameters), the validation set is used to determine the network structure or hyperparameters that control model complexity (fitting hyperparameters), and the test set is used to evaluate the performance of the final model. Finally, the training, validation, and test sets are divided into a commonly used 6:2:2 ratio based on the sample size, and then the model training and learning is carried out.

[0054] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, material, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, material, or apparatus.

[0055] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A 3D atmospheric pollution fusion and analysis method based on stereo telemetry data from a ResNet network, characterized by: S1: Preprocess the collected air quality data. Since the collected data set is large and the differences between the data are large, the impact of these differences will make the model fitting process difficult. Therefore, it is necessary to preprocess the raw data set. Common preprocessing operations such as data normalization are used to control the raw data to a smaller specific range, thereby better training and fitting the deep learning model. A shortcut connection method used by the residual structure: S2: Gradient generation node. In the degree generation phase, we first use a ResNet3D network pre-trained on kinetic data to fine-tune it on the training dataset so that the ResNet3D network has preliminary recognition capabilities on the dataset. Then, we fix the network parameters and use the network prediction label to derive the input: The gradient generation process is based on a network that already has basic resolving power. This network is used to derive the input and normalize the derived derivatives. In this way, for each spatial position of the input video, the normalized gradient value corresponding to that position is obtained. These gradient information will then be used to guide the feature enhancement module. The specific process is as follows: First, a network pre-trained on the Kinetics dataset is fine-tuned on the specified dataset to enable the network to have the most basic recognition ability on the experimental dataset. The currently trained network is defined as f, which has a preliminary resolution ability for the specified dataset. Then f represents the input image x a The mapping relationship between and image label y: y=f(x a ) where y represents the predicted label of the image segment by the fine-tuned network; Fix the trained network parameters and use the predicted output of the final network to derive the input image and generate gradients. Define the gradient value of the network for a pixel in the video as i, j, T. The expressions of i, j, T are as follows: in, Represents the value of the coordinate point (i, j, T) in the image, The generated gradient is then normalized using the normalization formula: The final θ generated by the normalized gradient value of the position in the original image is a matrix of the same size as the input image, in which the elements are the normalized gradient values ​​of each network for the pixel point at that position in the original image; S3: Perform model training and testing. After completing the model definition, the image dataset is divided into training set, validation set, and test set in a ratio of 6:2:

2. The initial deep learning model is trained to obtain the prediction results. After calculating the error through the loss function, the parameters of the learning model are adjusted through the optimizer to improve data accuracy.

2. The ResNet-based 3D atmospheric pollution fusion and analysis method for stereo telemetry data according to claim 1 is characterized by: The ResNet3D model is divided into two channels. The first channel is used to perform feature selection on environmental data, filter out the implicit features inherent in the environmental data that have the greatest impact on pollutant emissions, and provide them to the second channel as parameter variables of the pollutant prediction model; the second channel is used to predict pollutants, that is, to predict the detection results of unknown road sections based on the pollutant detection data at known points.

3. The ResNet-based 3D atmospheric pollution fusion and analysis method for stereo telemetry data according to claim 1 is characterized by: The 3D spatial feature learning of the ResNet network includes the ResNet3D network deep network, which can be understood as a feature extractor. The deep network extracts different features from shallow to deep layers, and finally sends the features to a classifier or performs other processing through different combinations. In the process of feature extraction by deep neural networks, the overall feature generation process is generated as the network deepens. The input of each layer is the features of all outputs of the previous layer. This layer needs to process all this information. These features are global features of the input image.

4. The ResNet-based 3D atmospheric pollution fusion and analysis method for stereo telemetry data according to claim 1 is characterized by: The data is collected using the Kinetics dataset, and the ResNet3D network is pre-trained on the Kinetics dataset.

5. The ResNet-based 3D atmospheric pollution fusion and analysis method for stereo telemetry data according to claim 1 is characterized by: ResNet regression adopts the design of residual network.

6. The ResNet-based 3D atmospheric pollution fusion and analysis method for stereo telemetry data according to claim 1 is characterized by: Deep neural networks are used to conduct in-depth analysis of the data.

Images