Visibility perception method of tower probe under complex meteorological conditions of sea and land

Abstract

The present application belongs to the field of meteorological monitoring and intelligent sensing technology, and discloses a visibility sensing algorithm for tower probe under complex sea-land meteorological conditions based on historical similarity and deep learning. First, the probe on the tower is used to collect meteorological data, image data and environmental data in the sea-land area to construct a multi-dimensional data sample library. Through analysis of the meteorological data and image data, key features are extracted, a historical similar data matching model is established, and historical data similar to the current meteorological conditions are found out. At the same time, based on deep learning technology, a hybrid neural network model for visibility sensing is built, the real-time collected data and the historical similar data are input into the hybrid neural network model for visibility sensing, and training and optimization are carried out. Finally, according to the trained hybrid neural network model for visibility sensing, the visibility under the current complex sea-land meteorological conditions is sensed and predicted in real time, and accurate visibility information support is provided for the fields of sea navigation and land transportation.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological monitoring and intelligent perception technology, and relates to a visibility perception method of an iron tower probe based on historical similarity and deep learning under complex meteorological conditions at sea and on land. Background Art

[0002] Visibility, as an important indicator of atmospheric transparency, plays a key role in maritime and land transportation, disaster warning, and production safety. At sea, visibility directly impacts ship navigation safety, waterway scheduling, and port operation efficiency. On land, particularly in coastal areas and areas with complex terrain, low visibility significantly increases the risk of traffic accidents and impacts the normal operation of transportation networks. Accurate visibility perception can effectively protect the lives and property of participants in maritime and land transportation, improve the efficiency of transportation systems, and reduce potential risks in adverse weather conditions, thus playing a vital role in the stable development of the economy and society.

[0003] In existing technologies, visibility perception in land and sea areas mostly relies on traditional single-point visibility meters, weather radars and other equipment. These devices can often only provide local, discrete visibility data, making it difficult to meet the needs of large-scale, continuous, and real-time monitoring under complex land and sea meteorological conditions. Visibility prediction methods based on historical data usually only consider the statistical laws of a single meteorological element and cannot fully capture the interactive effects of multiple meteorological factors in land and sea environments. Simple deep learning algorithms are prone to overfitting when processing complex and changeable land and sea meteorological data, and lack effective utilization of data features under similar historical meteorological conditions. (Ma Chuyan, Zu Jian, Fu Qingpan, Luo Lingxiao. Air visibility prediction based on genetic neural network model [J]. Journal of Environmental Engineering, 2015, 9(4): 1905-1910)

[0004] Existing technologies make it difficult to effectively integrate multi-source heterogeneous data obtained by tower probes, and are unable to fully utilize the characteristics of the towers' wide distribution and obvious height advantages. As a result, in complex meteorological scenarios at sea and on land, the accuracy, reliability and real-time performance of visibility perception are difficult to meet actual application needs. Summary of the Invention

[0005] The purpose of the present invention is to provide a visibility perception method of an iron tower probe based on historical similarity and deep learning under complex meteorological conditions at sea and on land, aiming to solve the problems raised in the background technology.

[0006] The technical solution of the present invention:

[0007] A visibility perception method for tower probes under complex weather conditions on land and sea based on historical similarity and deep learning is proposed. The steps are as follows:

[0008] (1) Obtain historical and real-time multidimensional data and build a multidimensional data sample library containing various meteorological conditions and corresponding visibility data;

[0009] Visibility data: The actual visibility is obtained using visibility meters installed on towers in both land and sea areas;

[0010] Various meteorological condition data: Using probes installed on towers in both land and sea areas, multi-dimensional data is collected in real time, including meteorological data, image data, and environmental data. Meteorological data includes wind speed, wind direction, humidity, temperature, and air pressure; image data includes visible light and infrared images; and environmental data includes wave height and topography.

[0011] Data preprocessing: Unify the format of the collected multidimensional data, perform preliminary cleaning, remove noise and outliers, and store them in time series;

[0012] Construction of a multi-dimensional data sample library: Each sample contains the collection time, tower coordinates, meteorological data, image data, environmental data, and the corresponding measured visibility;

[0013] (2) Extract key features of meteorological data, environmental data, and image data from historical multidimensional data, establish a historical similarity data matching model, and form a historical similarity data set;

[0014] Feature extraction: Conduct in-depth analysis of meteorological data and image data in the multidimensional data sample library, and use feature engineering methods to extract key features of meteorological data and image data. For meteorological data and environmental data, extract meteorological and environmental features such as the changing trends, extreme values, and periodicity of different meteorological and environmental elements. For image data, use image processing technology to extract image features such as texture, brightness, and color distribution.

[0015] Build and train a historical similar data matching model: Use a graph neural network (GNN) model to input the extracted meteorological, environmental, and image features. These features are processed through an embedding layer, then fused through a fusion layer to output fused features. In the GNN model, the cosine similarity of the fused features is used to measure sample similarity. Use historical multidimensional data to train the GNN model, and optimize the expressive power of the fused features using a mean square error loss function, enabling the GNN model to more accurately capture sample similarity.

[0016] Forming a historical similarity dataset: Calling the trained historical similarity data matching model, inputting the query sample to be matched, the historical similarity data matching model automatically calculates the cosine similarity of the fusion features of the matching historical samples and the query sample, sorts them from high to low by similarity, and takes the top 5 samples; extracting the visibility values ​​and corresponding meteorological conditions of these 5 samples to form a historical similarity dataset corresponding to the current query sample;

[0017] (3) Build a hybrid neural network model for visibility perception, taking historical multidimensional data and its historical similarity datasets as input and using measured visibility as labels for training;

[0018] Based on deep learning technology, a hybrid neural network model for visibility perception was built, combining a convolutional neural network (CNN) with a long short-term memory (LSTM) network. The CNN branch used two convolutional layers with ReLU activation and 2×2 max pooling to output image features. The LSTM branch used two LSTM layers to output meteorological features. A fusion layer was set up to fuse image and meteorological features, and the visibility prediction value was output through a fully connected layer.

[0019] The hybrid neural network model for visibility perception is trained by taking historical multidimensional data and its historical similar data sets as input and visibility values ​​as output labels.

[0020] During the training process, appropriate loss functions (such as mean square error loss function) and optimization algorithms (such as stochastic gradient descent algorithm) are used to continuously adjust model parameters, optimize the structure of the hybrid neural network model of visibility perception, and improve the prediction accuracy and generalization ability of the hybrid neural network model of visibility perception for visibility.

[0021] (4) Based on the construction of the hybrid neural network model of visibility perception, the key features of real-time data are input into the hybrid neural network model of visibility perception, and combined with the historical similar data matching model, real-time visibility prediction is performed under the current complex meteorological conditions of sea and land;

[0022] The meteorological data and image data collected in real time by probes on iron towers in the sea and land areas are processed using the same feature extraction method as historical data to obtain the key features of the real-time data;

[0023] Call the historical similarity data matching model to output the historical similarity dataset of real-time data; input the key features of the real-time multidimensional data and its historical similarity dataset into the trained visibility perception hybrid neural network model, make real-time predictions of visibility under the current complex meteorological conditions of land and sea, and output the prediction results;

[0024] (5) Establish a visibility prediction and evaluation mechanism, compare and analyze the prediction results of the hybrid neural network model of visibility perception with the measured visibility, use a variety of evaluation indicators to quantitatively evaluate the prediction results, and continuously optimize the hybrid neural network model of visibility perception in practical applications.

[0025] Establish a visibility prediction result evaluation mechanism, and use evaluation indicators such as accuracy, mean square error, and mean absolute error to compare and analyze the prediction results of the hybrid neural network model of visibility perception with the actual measured visibility values;

[0026] Based on the evaluation results, the hybrid neural network model of visibility perception is dynamically optimized regularly, such as adjusting the parameters of the hybrid neural network model of visibility perception, updating training data, improving the network structure, etc., so as to continuously improve the accuracy and reliability of the model's visibility perception under complex meteorological conditions at sea and on land, and provide more accurate visibility information for practical applications.

[0027] Beneficial effects of the present invention: The embodiment of the present invention uses tower probes to collect real-time information such as meteorological data and image data in sea and land areas to build a data sample library; by analyzing historical meteorological data and image data, key features are extracted, and a historical similarity data matching model is established to find historical data similar to the current meteorological conditions; at the same time, based on deep learning technology, a visibility perception neural network model is built, and the real-time collected data and historical similarity data are input into the model for training and optimization; finally, based on the trained model, visibility under the current complex sea and land meteorological conditions is perceived and predicted in real time, providing accurate visibility information support for fields such as maritime navigation and land transportation. By combining historical similarity data matching with deep learning, the algorithm can effectively cope with complex sea and land meteorological conditions and achieve accurate perception and prediction of visibility. DETAILED DESCRIPTION

[0028] The specific implementation of the present invention is further described below in conjunction with the technical solution.

[0029] Understandably, existing technologies for sensing visibility in land and sea areas rely heavily on traditional single-point visibility meters, weather radars, and other equipment. These devices often only provide localized, discrete visibility data, making it difficult to meet the needs of large-scale, continuous, and real-time monitoring under complex land and sea meteorological conditions. Visibility prediction methods based on historical data typically only consider the statistical patterns of a single meteorological element and fail to fully capture the interactive influence of multiple meteorological factors in land and sea environments. Simple deep learning algorithms are prone to overfitting when processing complex and variable land and sea meteorological data, and lack effective utilization of data features from similar historical meteorological conditions. Furthermore, existing technologies struggle to effectively integrate multi-source, heterogeneous data captured by tower probes, failing to fully leverage the widespread distribution and significant height advantages of towers. Consequently, the accuracy, reliability, and real-time performance of visibility perception in complex land and sea meteorological scenarios struggle to meet practical application requirements.

[0030] To address the above issues, the present invention first uses tower probes to collect real-time information such as meteorological data and image data in land and sea areas to build a data sample library. By analyzing historical meteorological and image data, key features are extracted and a historical similarity data matching model is established to identify historical data similar to the current meteorological conditions. Simultaneously, based on deep learning technology, a visibility perception neural network model is constructed, and the real-time collected data and historical similarity data are input into the model for training and optimization. Finally, based on the trained model, visibility under the current complex meteorological conditions at sea and land is perceived and predicted in real time, providing accurate visibility information support for fields such as maritime navigation and land transportation. By combining historical similarity data matching with deep learning, this algorithm can effectively cope with complex meteorological conditions at sea and land, achieving accurate perception and prediction of visibility.

[0031] Specifically, in a preferred embodiment provided by the present invention, a method for perceiving visibility of a tower probe using historical similarity and deep learning under complex meteorological conditions at sea and on land comprises the following steps:

[0032] Step S101 : Using tower probes to collect multi-dimensional data such as meteorological data, image data, and environmental data in real time, and constructing a multi-dimensional data sample library containing various meteorological conditions and corresponding visibility data.

[0033] In an embodiment of the present invention, probes installed on towers in both land and sea areas are used to collect meteorological data (including wind speed, wind direction, humidity, temperature, air pressure, etc.), image data (visible light images, infrared images, etc.), and environmental data (such as wave height, topography, etc.) in real time. The collected raw data is formatted and preliminarily cleaned to remove noise and outliers. The data is then stored in time series to construct a multidimensional data sample library containing various meteorological conditions and corresponding visibility data, providing basic data support for subsequent analysis.

[0034] The method of using the tower probe to collect multidimensional data such as meteorological data, image data, and environmental data in real time and constructing a multidimensional data sample library containing various meteorological conditions and corresponding visibility data specifically includes the following steps:

[0035] Step S1011: using a tower probe to collect meteorological data, image data, and environment-related data.

[0036] Step S1012: unify the format of the original data.

[0037] Step S1013: Preliminary data cleaning (removing noise and outliers).

[0038] Step S1014: store data in time series.

[0039] Step S1015: construct a multi-dimensional data sample library (including various meteorological conditions and corresponding visibility data).

[0040] Furthermore, the visibility perception method based on historical similarity and deep learning of tower probes under complex sea and land meteorological conditions also includes the following steps:

[0041] Step S102, extracting key features of meteorological data, environmental data, and image data from the historical multidimensional data using the graph neural network method, establishing a historical similar data matching model, and forming a historical similar data set;

[0042] In an embodiment of the present invention, an in-depth analysis is performed on meteorological data and image data in a multidimensional data sample library, and a feature engineering method is used to extract key features of the meteorological data and image data. For meteorological data and environmental data, meteorological features and environmental features such as changing trends, extreme values, and periodicity of different meteorological and environmental elements are extracted; for image data, image processing technology is used to extract image features such as texture, brightness, and color distribution; based on the extracted features, a historical similar data matching model is established, and by calculating feature similarity, such as using cosine similarity, Euclidean distance, and other measurement methods, the historical data that is most similar to the current meteorological conditions is found to form a historical similar data set.

[0043] In a preferred embodiment of the present invention, the graph neural network method is used to extract key features of historical meteorological data and image data, and based on the extracted features, a historical similar data matching model is established, which specifically includes the following steps:

[0044] Step S1021: Acquire meteorological data and image data from a multidimensional data sample library.

[0045] Step S1022: extracting features such as change trends, extreme values, and periodicity from meteorological data and environmental data using feature engineering methods.

[0046] Step S1023: Apply image processing technology to the image data to extract features such as texture, brightness, and color distribution.

[0047] In step S1024, the trained historical similarity data matching model is called, and the query sample to be matched is input. The historical similarity data matching model automatically calculates the cosine similarity of the fusion features of the matching historical samples and the query sample, sorts them from high to low by similarity, and takes the first 5 samples; extracts the visibility values ​​and corresponding meteorological conditions of these 5 samples to form a historical similarity data set corresponding to the current query sample.

[0048] Furthermore, the visibility perception method based on historical similarity and deep learning of tower probes under complex sea and land meteorological conditions also includes the following steps:

[0049] In step S103 , a neural network model for visibility perception is constructed, which takes real-time multidimensional data and historical multidimensional data sets as inputs and is trained using actual visibility values ​​as labels.

[0050] In an embodiment of the present invention, a neural network model for visibility perception is constructed based on deep learning technology, such as a hybrid network structure combining a convolutional neural network (CNN) and a long short-term memory network (LSTM). Historical multidimensional data and its historical similar data sets are used as input, and the actually measured visibility values ​​are used as output labels. The hybrid neural network model for visibility perception is trained. During the training process, the parameters of the hybrid neural network model for visibility perception are continuously adjusted, the network structure is optimized, and a suitable loss function (such as a mean square error loss function) and an optimization algorithm (such as a stochastic gradient descent algorithm) are adopted to improve the model's prediction accuracy and generalization ability for visibility.

[0051] In a preferred embodiment of the present invention, the neural network model specifically for visibility perception is constructed, which uses real-time collected data and historical similar data sets as inputs and uses actual visibility values ​​as labels for training, and specifically includes the following steps:

[0052] Step S1031: Build a hybrid neural network model for visibility perception based on deep learning technology (such as a hybrid network structure combining CNN and LSTM).

[0053] Step S1032: Acquire real-time multi-dimensional data and its historical similar data sets as input.

[0054] Step S1033 : Using the actually measured visibility values ​​as output labels, the hybrid neural network model for visibility perception is trained.

[0055] Step S1034 , adjusting the parameters of the hybrid neural network model for visibility perception, optimizing the network structure, and selecting a suitable loss function (such as mean square error loss function) and optimization algorithm (such as stochastic gradient descent algorithm).

[0056] Step S1035 , improving the prediction accuracy and generalization ability of the hybrid neural network model for visibility perception.

[0057] Furthermore, the visibility perception method based on historical similarity and deep learning of tower probes under complex sea and land meteorological conditions also includes the following steps:

[0058] In step S104, based on the construction of the hybrid neural network model for visibility perception, real-time multi-dimensional data is input into the hybrid neural network model for visibility perception, and combined with the historical similar data matching model, real-time prediction of visibility under the current complex sea and land meteorological conditions is performed.

[0059] Among them, in the preferred embodiment provided by the present invention, the real-time data features are input into the hybrid neural network model of visibility perception based on the model construction, and combined with the historical similar data matching model to perform real-time prediction of visibility under the current complex meteorological conditions of sea and land, specifically including the following steps:

[0060] Step S1041: The tower probe collects meteorological data and image data in real time.

[0061] Step S1042 : Process the real-time data using the same feature extraction method as the historical data to obtain key features of the real-time data.

[0062] Step S1043: input the key features of the real-time data into the trained hybrid neural network model of visibility perception.

[0063] Step S1044: combining the similar historical data information provided by the historical similar data matching model.

[0064] In step S1045 , the model makes a real-time prediction of visibility under the current complex sea and land meteorological conditions.

[0065] Furthermore, the visibility perception method based on historical similarity and deep learning of tower probes under complex sea and land meteorological conditions also includes the following steps:

[0066] Step S104: Establish a visibility prediction and evaluation mechanism, compare and analyze the model prediction results with the actual measured visibility data, use multiple evaluation indicators to quantify the prediction results, and continuously optimize the model in actual application.

[0067] Among them, a visibility prediction and evaluation mechanism is established, the prediction results of the hybrid neural network model of visibility perception are compared and analyzed with the actual measured visibility data, and multiple evaluation indicators are used to quantify the prediction results. The model is continuously optimized in actual application. The specific steps include the following:

[0068] Step S1051: Establish a visibility prediction result evaluation mechanism.

[0069] Step S1052 : Compare and analyze the prediction results of the hybrid neural network model of visibility perception with the actually measured visibility data.

[0070] Step S1053: Use evaluation indicators such as accuracy, mean square error, and mean absolute error to quantitatively evaluate the prediction results.

[0071] Step S1054: determine whether the hybrid neural network model of visibility perception needs to be optimized.

[0072] Step S1055 , regularly performing dynamic optimization to improve the accuracy and reliability of visibility perception of the hybrid neural network model under complex meteorological conditions at sea and land.

Claims

1. A visibility perception method for tower probes under complex sea and land weather conditions based on historical similarity and deep learning, characterized by: Here are the steps: (1) Obtain historical and real-time multidimensional data and build a multidimensional data sample library containing various meteorological conditions and corresponding visibility data; (2) Extract key features of meteorological data, environmental data, and image data from historical multidimensional data, establish a historical similarity data matching model, and form a historical similarity data set; (3) Build a hybrid neural network model for visibility perception, taking historical multidimensional data and its historical similarity datasets as input and using measured visibility as labels for training; (4) Based on the construction of the hybrid neural network model of visibility perception, the key features of real-time data are input into the hybrid neural network model of visibility perception, and combined with the historical similar data matching model, real-time visibility prediction is performed under the current complex meteorological conditions of sea and land; (5) Establish a visibility prediction and evaluation mechanism, compare and analyze the prediction results of the hybrid neural network model of visibility perception with the measured visibility, use a variety of evaluation indicators to quantitatively evaluate the prediction results, and continuously optimize the hybrid neural network model of visibility perception in practical applications.

2. The visibility perception method of tower probes based on historical similarity and deep learning under complex sea and land weather conditions according to claim 1 is characterized in that: The specific implementation process of step (1) is as follows: Visibility data: The actual visibility is obtained using visibility meters installed on towers in both land and sea areas; Various meteorological condition data: Using probes installed on towers in both land and sea areas, multi-dimensional data is collected in real time, including meteorological data, image data, and environmental data. Meteorological data includes wind speed, wind direction, humidity, temperature, and air pressure; image data includes visible light and infrared images; and environmental data includes wave height and topography. Data preprocessing: Unify the format of the collected multidimensional data, perform preliminary cleaning, remove noise and outliers, and store them in time series; Construction of a multi-dimensional data sample library: Each sample contains the collection time, tower coordinates, meteorological data, image data, environmental data and the corresponding measured visibility.

3. The visibility perception method of tower probes based on historical similarity and deep learning under complex sea and land weather conditions according to claim 1 is characterized in that: The specific implementation process of step (2) is as follows: Feature extraction: Conduct in-depth analysis of meteorological data and image data in the multidimensional data sample library, and use feature engineering methods to extract key features of meteorological data and image data. For meteorological data and environmental data, extract the changing trends, extreme values, and periodic meteorological and environmental features of different meteorological and environmental elements. For image data, use image processing technology to extract image features such as texture, brightness, and color distribution. Construct and train a historical similar data matching model: Using a graph neural network model, the extracted meteorological, environmental, and image features are input. These features are processed through an embedding layer, and then fused through a fusion layer to output fused features. The graph neural network model uses the cosine similarity of the fused features to measure sample similarity. The graph neural network model is trained using historical multidimensional data, and the mean square error loss function is used to optimize the expressive power of the fused features, enabling the graph neural network model to more accurately capture sample similarity. Form a historical similarity data set: call the trained historical similarity data matching model, input the query sample to be matched, and the historical similarity data matching model automatically calculates the cosine similarity of the fusion features of the matching historical samples and the query samples, sorts them from high to low by similarity, and takes the top 5 samples; extract the visibility values ​​and corresponding meteorological conditions of these 5 samples to form a historical similarity data set corresponding to the current query sample.

4. The visibility perception method of tower probes based on historical similarity and deep learning under complex sea and land weather conditions according to claim 1 is characterized in that: The specific implementation process of step (3) is as follows: Based on deep learning technology, a hybrid neural network model for visibility perception was built, combining a convolutional neural network with a long short-term memory network. The convolutional neural network branch has two convolutional layers, with a ReLU activation function and 2×2 maximum pooling to output image features. The long short-term memory network branch sets two LSTM layers to output meteorological features; Set the fusion layer to fuse image features and meteorological features, and output the visibility prediction value through the fully connected layer; The hybrid neural network model for visibility perception is trained by taking historical multidimensional data and its historical similar data sets as input and visibility values ​​as output labels. During the training process, the loss function and optimization algorithm are used to continuously adjust the model parameters, optimize the structure of the hybrid neural network model of visibility perception, and improve the prediction accuracy and generalization ability of the hybrid neural network model of visibility perception.

5. The visibility perception method of tower probes based on historical similarity and deep learning under complex sea and land weather conditions according to claim 1 is characterized in that: The specific implementation process of step (4) is as follows: The meteorological data and image data collected in real time by probes on iron towers in the sea and land areas are processed using the same feature extraction method as historical data to obtain the key features of the real-time data; Call the historical similarity data matching model to output the historical similarity data set of real-time data; The key features of real-time multidimensional data and its historical similar data sets are input into the trained visibility perception hybrid neural network model to make real-time predictions of visibility under the current complex sea and land meteorological conditions and output the prediction results.

6. The visibility perception method of tower probes based on historical similarity and deep learning under complex sea and land weather conditions according to claim 1 is characterized in that: The specific implementation process of step (5) is as follows: Establish a visibility prediction result evaluation mechanism and compare the prediction results of the hybrid neural network model of visibility perception with the actual measured visibility values ​​using accuracy, mean square error, and mean absolute error; Based on the evaluation results, the hybrid neural network model of visibility perception is dynamically optimized regularly.