Plateau railway icing early warning method and system based on sequence learning

Abstract

The invention discloses a plateau railway icing early warning method and system based on sequence learning, and solves the problem of data labeling when no direct icing record exists by collecting and preprocessing daily scale meteorological parameter data and applying an innovative method based on meteorological experience criteria to generate icing labels according to multiple meteorological parameter thresholds and logic judgment. A layered self-attention mechanism is utilized, meteorological data features are extracted by a meteorological feature layer, cross-day meteorological dependence features are extracted by a time layer through sliding window self-attention, observation point geographic association feature mining is realized by fusing longitude and latitude position coding in a space layer, daily scale data are efficiently adapted, and extra terrain information is not needed; and constructing a time-space fusion icing early warning model to output an icing probability, and setting a dynamic threshold value based on observation point clustering to carry out graded early warning decision making. According to the method, the precision and efficiency of icing early warning of the plateau railway can be remarkably improved, and a powerful guarantee is provided for safe operation of the railway.

Description

Technical Field

[0001] The present invention relates to the technical field of plateau railway icing warning technology, and in particular to a plateau railway icing warning method and system based on sequence learning. Background Art

[0002] During high-altitude railway operations, icing poses a serious threat to railway safety and normal operation. However, current early warning methods for icing on high-altitude railways have many shortcomings. Traditional methods often rely on direct icing records, but in many cases, the lack of direct icing record data makes data annotation difficult, affecting the accuracy and reliability of early warning models. Furthermore, existing early warning methods may not fully exploit the temporal and spatial characteristics of meteorological data, making it difficult to efficiently adapt daily-scale meteorological data, thereby reducing the accuracy and efficiency of early warnings. Therefore, it is of great practical significance to develop an early warning method for icing on high-altitude railways that can effectively address the data annotation problem and fully utilize the characteristics of meteorological data. Summary of the Invention

[0003] The present invention provides a plateau railway icing early warning method and system based on sequence learning, so as to solve the problems of data labeling difficulty, low warning accuracy and efficiency in the existing plateau railway icing early warning methods.

[0004] According to the first aspect, an embodiment provides a sequence learning-based plateau railway icing early warning method, the method comprising: Collect geographical location information of the plateau railway area and multiple meteorological parameter data on a daily scale, and perform pre-processing; Set thresholds for various meteorological parameters and use logical judgment to label meteorological parameter data and generate icing labels; A feature extraction and fusion model based on a hierarchical self-attention mechanism is established to extract features from daily-scale meteorological parameter data, and the extracted features are fused with geographic location information to obtain spatiotemporal fusion features. Establishing and training a spatiotemporal fusion icing warning model based on a deep learning network, inputting the obtained spatiotemporal fusion features into the spatiotemporal fusion icing warning model, and outputting icing prediction results; A graded warning mechanism is established to classify the obtained icing prediction results into warning levels.

[0005] Furthermore, geographical location information of the plateau railway area and multiple meteorological parameter data at a daily scale are collected and pre-processed, including: The geographical location information includes longitude and latitude information, which is used to identify the specific geographical location of the observation point; The meteorological parameter data includes temperature, precipitation, wind speed, humidity, air pressure and sunshine; temperature includes minimum temperature, average temperature and average surface temperature; precipitation refers to precipitation amount; wind speed refers to the average wind speed at a preset height; humidity refers to relative humidity; air pressure refers to surface air pressure; sunshine refers to sunshine hours; The collected data are preprocessed, including data cleaning, missing value processing and data standardization.

[0006] Furthermore, thresholds for various meteorological parameters are set, and logical judgment is used to annotate meteorological parameter data to generate icing labels, including: For precipitation scenarios, the freezing conditions are: when there is effective precipitation, that is, the precipitation exceeds the preset precipitation threshold, and the surface temperature is less than or equal to the preset surface temperature threshold, it is directly judged as freezing; or when there is effective precipitation, that is, the precipitation exceeds the preset precipitation threshold, the minimum temperature is less than or equal to the preset minimum temperature threshold, and the surface temperature is less than or equal to the preset surface temperature threshold, it is judged as freezing, to take into account the situation of secondary freezing due to snowmelt; For high humidity without precipitation, the freezing conditions are: when there is no effective precipitation, that is, the precipitation does not exceed the precipitation threshold, the surface temperature is less than or equal to the preset surface temperature threshold, the relative humidity is greater than or equal to the high humidity threshold, and the wind speed is less than the preset wind speed threshold, it is determined to be freezing; For extreme low temperature scenarios, the freezing condition is: when the minimum temperature is less than or equal to the extreme low temperature threshold, it is considered freezing even if there is no effective precipitation, that is, the precipitation does not exceed the precipitation threshold; If the icing conditions in each scene are not met, it is determined to be non-icing; In addition, suppression conditions are also considered: if the sunshine hours are greater than or equal to the preset sunshine hours threshold, or the average wind speed is greater than or equal to the preset wind speed threshold, it is considered that the icing risk will be offset. Even if the icing conditions in each scenario are met, it is still judged as non-icing.

[0007] Furthermore, a feature extraction and fusion model based on the hierarchical self-attention mechanism is established to extract features from daily-scale meteorological parameter data, and the extracted features are fused with geographic location information to obtain spatiotemporal fusion features, specifically including: The feature extraction and fusion model based on the hierarchical self-attention mechanism includes a meteorological feature extraction layer, a temporal feature extraction layer, and a spatial feature fusion layer, and the output of the previous layer will serve as the input of the next layer; The meteorological feature extraction layer is used to extract single-day meteorological features from the collected daily-scale meteorological parameter data, and the output single-day meteorological feature vector is used as the input of the time feature extraction layer; The temporal feature extraction layer is used to extract inter-day meteorological dependency features using sliding window self-attention, and the output inter-day time-dependent meteorological features serve as input to the spatial feature fusion layer; The spatial feature fusion layer is used to splice or weightedly fuse the cross-daily time-dependent meteorological features with the latitude and longitude position codes, and the final obtained spatiotemporal fusion features contain time-dependent and geographically associated meteorological information.

[0008] Furthermore, the meteorological feature extraction layer is used to extract meteorological features from the collected daily-scale meteorological parameter data, specifically including: The meteorological feature extraction layer extracts features from the collected daily-scale meteorological data based on the improved VGG16 model. First, the meteorological parameter data is encoded and converted into a format suitable for convolutional neural network processing. The change patterns of different meteorological parameters in the data are obtained as underlying features. The improved VGG16 model is used to extract deep features of the data and explore the potential correlations between different meteorological parameters. Finally, the underlying features are fused with the deep features and output on the Softmax.

[0009] Furthermore, the improved VGG16 model includes five convolutional blocks, namely C1, C2, C3, C4 and C5, and the output of C1 is used as the input of C2, the output of C2 is used as the input of C3, the output of C3 is used as the input of C4, and the output of C4 is used as the input of C5; Among them, the convolution blocks C1 and C2 both contain two convolution layers B, two activation functions and a pooling layer P. Specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the second convolution layer B is used as the input of the pooling layer P. The C3 convolutional block contains three convolutional layers B, three activation functions, and a pooling layer P. Specifically, an activation function is added after each convolutional layer B, and a pooling layer P is added at the end of the convolutional block, so that the output of the third convolutional layer B serves as the input of P. The convolutional layer of the C3 convolutional block uses 256 3*3 convolution kernels to better capture the complex characteristics of meteorological data and the potential correlation between different meteorological parameters. The C4 and C5 convolution blocks both contain three convolution layers B, three activation functions, and a pooling layer P. Specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the third convolution layer B serves as the input of the pooling layer P. The convolution kernel size of the convolution layer B is 3*3, and the moving step is 1; the pooling layer P uses the maximum pooling method, sets the pooling window size to 2*2, and the moving step is 2; the activation function uses the SeLU function.

[0010] Furthermore, a spatiotemporal fusion icing warning model based on a deep learning network is established and trained, the obtained spatiotemporal fusion features are input into the spatiotemporal fusion icing warning model, and an icing prediction result is output, specifically including: The spatiotemporal fusion icing warning model adopts a fully connected neural network architecture, in which the number of neurons in the input layer is the same as the dimension of the spatiotemporal fusion feature vector; multiple hidden layers use the ReLU activation function, and the number of neurons is dynamically adjusted according to the characteristics of meteorological data and warning needs; the output layer uses the Sigmoid activation function to output the icing probability; the number of hidden layers and neurons is determined by hyperparameter tuning methods; and the binary cross entropy loss function is used to measure the difference between the predicted results and the true labels. Construct training sets, validation sets, and test sets to train, validate, and test the model. Use the Adam optimizer to update parameters, and the learning rate is dynamically adjusted according to the validation set loss.

[0011] Furthermore, a graded warning mechanism is established to classify the obtained icing prediction results into warning levels, specifically including: Dynamic thresholds are set based on observation point clustering to make graded warning decisions. The warnings are divided into different levels according to the icing probability output by the spatiotemporal fusion icing warning model and the dynamic thresholds set based on the observation point clustering results.

[0012] According to the second aspect, an embodiment provides a plateau railway icing warning system based on sequence learning, the system comprising: The data acquisition and processing module is used to collect geographical information of the plateau railway area and multiple meteorological parameter data on a daily scale, and perform pre-processing; The icing label generation module is used to set the thresholds of various meteorological parameters and combine logical judgment to annotate the meteorological parameter data to generate icing labels; The spatiotemporal feature extraction module is used to establish a feature extraction and fusion model based on the hierarchical self-attention mechanism to extract features from daily meteorological parameter data, and fuse the extracted features with geographic location information to obtain spatiotemporal fusion features; An icing prediction module is used to establish and train a spatiotemporal fusion icing warning model based on a deep learning network, input the obtained spatiotemporal fusion features into the spatiotemporal fusion icing warning model, and output an icing prediction result; The graded warning module is used to establish a graded warning mechanism and perform warning classification on the obtained icing prediction results.

[0013] According to a third aspect, an embodiment provides an electronic device, the device comprising: a processor and a memory; The memory is used to store one or more program instructions; The processor is used to run one or more program instructions to execute the steps of a plateau railway icing warning method based on sequence learning as described in any one of the above items.

[0014] According to the fourth aspect, an embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of a sequence learning-based plateau railway icing warning method as described in any one of the above items are implemented.

[0015] The present invention provides a sequence learning-based plateau railway icing early warning method and system, which has the following beneficial effects: (1) This paper addresses the issue of icing on plateau railways by using an innovative method based on meteorological experience to generate icing labels based on multiple meteorological parameter thresholds and logical judgment. This provides accurate training data for the subsequent spatiotemporal fusion icing warning model, solves the data labeling problem, and lays a solid foundation for effective model training.

[0016] (2) The hierarchical self-attention mechanism can efficiently adapt to daily-scale meteorological data without the need for additional terrain information. In practical applications, obtaining terrain information may be difficult or costly. However, the present invention can achieve accurate icing warnings by relying solely on collected daily-scale meteorological data and latitude and longitude information, reducing the difficulty and cost of data acquisition and improving the practicality and operability of the method.

[0017] (3) The spatiotemporal fusion icing warning model constructed by the present invention can fully learn the spatiotemporal characteristics of meteorological data, accurately output the icing probability, and improve the accuracy of the warning.

[0018] (4) Based on the clustering of observation points, dynamic thresholds are set to make graded warning decisions. Based on the icing probability output by the spatiotemporal fusion icing warning model, dynamic thresholds are set in combination with the clustering results of observation points to divide warnings into different levels. This graded warning decision-making can take corresponding preventive measures according to the different degrees of icing risk, enabling railway operating departments to respond more targetedly, improving the efficiency of warnings and providing a strong guarantee for safe railway operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of a sequence learning-based early warning method for plateau railway icing provided by one embodiment of the present invention; Figure 2 A schematic diagram of the improved VGG16 model structure in a sequence learning-based plateau railway icing warning method provided by one embodiment of the present invention; Figure 3 A schematic diagram of the logical structure of a sequence learning-based plateau railway icing warning method provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described in detail below by means of specific embodiments in conjunction with the accompanying drawings. Similar elements in different embodiments are numbered with associated similar elements. In the following embodiments, many detailed descriptions are provided to enable the present invention to be better understood. However, those skilled in the art will readily appreciate that some of the features may be omitted under different circumstances, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the present invention are not shown or described in the specification. This is to avoid the core of the present invention being overwhelmed by excessive descriptions, and for those skilled in the art, it is not necessary to describe these related operations in detail. They can fully understand the related operations based on the description in the specification and the general technical knowledge in the art.

[0021] In addition, the features, operations, or characteristics described in the specification may be combined in any appropriate manner to form various embodiments. Furthermore, the steps or actions in the method description may be reordered or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various sequences in the specification and drawings are provided solely for the purpose of clearly describing a particular embodiment and are not intended to be mandatory, unless otherwise specified.

[0022] The first embodiment of the present invention provides a plateau railway icing warning method based on sequence learning. Figure 1 Provide detailed explanation.

[0023] like Figure 1 As shown, in step S100, geographical location information of the plateau railway area and multiple meteorological parameter data on a daily scale are collected and pre-processed.

[0024] The above steps specifically include: In this embodiment, the collected data includes three elements, namely time elements, geographical elements and meteorological elements. For the time element, the time scale of the collected data is the daily scale, which ensures that the changes in meteorological conditions on different dates can be obtained, providing a basis for subsequent ice warnings. For the geographical element, longitude and latitude information is collected to identify the specific geographical location of the observation point, providing necessary geographic data support for spatial layer feature extraction. For the meteorological element, a variety of meteorological parameters are collected. In order to ensure the consistency and accuracy of the data, the collected data is preprocessed as follows, including data cleaning, missing value processing and data standardization, to ensure the consistency and accuracy of the data.

[0025] Various meteorological parameters include: temperature, precipitation, wind speed, humidity, air pressure and sunshine.

[0026] Temperature includes minimum temperature, average temperature and average surface temperature, among which: Minimum temperature records the daily minimum temperature and is used to assess the impact of extreme low temperatures on icing; Average temperature records the average daily temperature, taking into account the potential impact of temperature changes on icing; Mean surface temperature records the average temperature of the Earth's surface, which has a direct impact on the formation and persistence of ice.

[0027] Precipitation, i.e., the amount of precipitation, is recorded daily to analyze the potential contribution of precipitation to ice formation, especially in low-temperature environments.

[0028] The wind speed is the average wind speed at a height of 10 meters. The daily average wind speed at a height of 10 meters is recorded, taking into account the impact of wind speed on ice formation and duration, as well as the effect on the dispersion of ice and snow.

[0029] Humidity uses relative humidity to record daily relative humidity. A high humidity environment is more conducive to ice formation, especially under low temperature and precipitation conditions.

[0030] The air pressure is the surface air pressure, which is recorded daily. Changes in air pressure may affect the probability of freezing. Low pressure systems are usually accompanied by more precipitation and lower temperatures.

[0031] Sunshine refers to the number of sunshine hours, which records the daily sunshine duration. Insufficient sunshine may increase the risk of icing, especially at night and in cloudy conditions.

[0032] By comprehensively collecting the above-mentioned detailed meteorological parameters, we can more accurately capture and analyze the various meteorological factors affecting icing, providing rich data support for the generation of icing labels and the construction of spatiotemporal fusion models, thereby significantly improving the accuracy and reliability of icing warnings on plateau railways.

[0033] like Figure 1 As shown, in step S200, the thresholds of various meteorological parameters are set, and the meteorological parameter data are annotated in combination with logical judgment to generate an icing label.

[0034] The above steps specifically include: In this embodiment, a series of core thresholds are set to clearly define the key meteorological conditions for icing. Next, a basic condition judgment is performed to evaluate whether the various meteorological parameters meet the basic conditions for icing. In the logical judgment process, an evaluation is performed from high to low priority to determine whether an icing label is generated. Finally, an inhibition condition judgment and a comprehensive judgment are performed to determine whether there are factors that offset the icing risk. Through the above-mentioned detailed judgment logic based on meteorological experience criteria, an icing label is generated to solve the data labeling problem when there is no direct icing record, and to provide accurate training data for the subsequent spatiotemporal fusion model, thereby improving the accuracy and reliability of the plateau railway icing warning.

[0035] In this embodiment, the parameter thresholds specifically include: the surface temperature critical value and the minimum air temperature critical value are both set to 273.15K (i.e., 0°C) as the critical point for temperature freezing; the effective precipitation threshold is set to 0.1mm to account for the impact of trace precipitation on freezing; the high humidity threshold is set to 80% because high humidity environments easily reach condensation and frost conditions; the wind speed suppression threshold is set to 5.0m / s. Wind speeds exceeding this value will suppress the surface cooling rate, thereby reducing the probability of freezing; and the sunshine suppression threshold is set to 2.0 hours. Sunshine duration exceeding this value will cause the surface to warm up through solar radiation, reducing the risk of freezing.

[0036] Basic conditions are judged to assess whether various meteorological parameters meet the basic conditions for freezing, including: whether there is effective precipitation, that is, whether the precipitation is greater than 0.1 mm; whether the surface temperature meets the freezing conditions, that is, whether the surface temperature is less than or equal to 273.15 K; whether the minimum temperature meets the freezing conditions, that is, whether the minimum temperature is less than or equal to 273.15 K; whether the ambient humidity reaches the high humidity condition, that is, whether the relative humidity is greater than or equal to 80%; whether the wind speed is conducive to freezing, that is, whether the average wind speed is less than 5.0 m / s; and whether the sunshine conditions are conducive to freezing, that is, whether the sunshine hours are less than 2.0 hours.

[0037] During the logic judgment process, the priority is evaluated from high to low to determine whether to generate an icing label, as follows: First, consider the precipitation scenario: when there is effective precipitation and the surface temperature is less than or equal to 273.15K, it is directly judged as freezing. When there is effective precipitation, the minimum temperature is less than or equal to 273.15K, and the surface temperature is less than or equal to 275.15K (i.e., 2°C), it is judged as freezing to account for the situation of secondary freezing due to snowmelt. Secondly, for high humidity without precipitation: when there is no effective precipitation, the surface temperature is less than or equal to 273.15K, the relative humidity is greater than or equal to 80%, and the wind speed is less than 5.0m / s, it is judged as icing; In addition, extreme low temperature scenarios are also considered: when the minimum temperature is less than or equal to 268.15K (i.e. -5°C), it is considered as freezing even if there is no effective precipitation; If none of the above icing conditions are met, it is determined to be non-icing; A comprehensive assessment is made based on the suppression conditions to determine whether there are factors that offset the icing risk, specifically: Suppression conditions: If the sunshine hours are greater than or equal to the preset sunshine hours threshold, or the average wind speed is greater than or equal to the preset wind speed threshold, the icing risk is considered to be offset. Even if the icing conditions in each scenario are met, it is still determined to be non-icing.

[0038] like Figure 1As shown, in step S300, a feature extraction and fusion model based on a hierarchical self-attention mechanism is established to extract features from daily-scale meteorological parameter data, and the extracted features are fused with geographic location information to obtain spatiotemporal fusion features.

[0039] The above steps specifically include: The feature extraction and fusion model based on the hierarchical self-attention mechanism consists of a meteorological feature extraction layer, a temporal feature extraction layer, and a spatial feature fusion layer. Its data flow follows a hierarchical structure, with the output of the previous layer serving as the input to the next layer. The meteorological feature extraction layer extracts meteorological data features and outputs a single-day meteorological feature vector, which serves as the input to the temporal feature extraction layer. The temporal feature extraction layer extracts cross-daily meteorological dependency features using sliding window self-attention, outputting cross-daily time-dependent meteorological features that serve as the input to the spatial feature fusion layer. The spatial feature fusion layer fuses longitude and latitude location codes to mine geographic correlation features of observation points. This layer efficiently adapts to daily-scale data without requiring additional terrain information, ultimately outputting spatiotemporal fusion features that contain both temporally dependent and geographically correlated meteorological information.

[0040] In this embodiment, the meteorological feature extraction layer extracts features from collected daily meteorological data based on an improved VGG16 model. First, the meteorological data is encoded and converted into a format suitable for convolutional neural network processing. Variation patterns in temperature, wind speed, and other parameters are captured as underlying features. The improved VGG16 model is then used to extract deeper features from the data, exploring potential correlations between different meteorological parameters. Finally, the underlying features are fused with the deeper features and outputted through a softmax filter.

[0041] The improved VGG16 model in this embodiment is as follows: Figure 2 As shown in Figure 1, it includes five convolution blocks: C1, C2, C3, C4, and C5. The five convolution blocks are connected in series, and the output of each convolution block serves as the input of the next convolution block. The structure of each convolution block is as follows: The C1 convolution block includes two convolutional layers (B1 and B2). Each convolutional layer uses a 3*3 convolution kernel with a stride of 1, and a SeLU activation function is connected after each convolutional layer; a maximum pooling layer P1 uses a 2*2 pooling window with a stride of 2.

[0042] The C2 convolution block includes two convolutional layers (B3 and B4). Each convolutional layer uses a 3*3 convolution kernel with a stride of 1, and a SeLU activation function is connected after each convolutional layer; a maximum pooling layer P2 uses a 2*2 pooling window with a stride of 2.

[0043] The C3 convolution block includes three convolutional layers (B5, B6, and B7). Each convolutional layer uses a 3*3 convolution kernel with a stride of 1, and a SeLU activation function is connected after each convolutional layer. Each convolutional layer uses 256 3*3 convolution kernels to enhance the model's ability to extract complex features of meteorological data. There is also a maximum pooling layer P3 with a 2*2 pooling window and a stride of 2.

[0044] The C4 convolution block includes two convolutional layers (B8, B9, B10). Each convolutional layer uses a 3*3 convolution kernel with a stride of 1, and a SeLU activation function is connected after each convolutional layer; a maximum pooling layer P4 uses a 2*2 pooling window with a stride of 2.

[0045] The C5 convolution block includes two convolutional layers (B11, B12, and B13). Each convolutional layer uses a 3*3 convolution kernel with a stride of 1, and a SeLU activation function is connected after each convolutional layer; a maximum pooling layer P5 uses a 2*2 pooling window with a stride of 2.

[0046] The calculation formula of the activation function SeLU is as follows:

[0047] Where, For the data to be processed, is the result after activation function processing, and is a parameter; In this embodiment, the temporal feature extraction layer extracts the cross-day weather dependency features through the sliding window self-attention, specifically extracts the cross-day weather dependency features through the sliding window self-attention, and uses the weather features. Specifically, the fixed size is set to A window slides across the daily meteorological data series. In addition to the original time series data, the window also considers meteorological characteristics. This captures temporal variations and correlations in meteorological data when analyzing changes in meteorological characteristics such as temperature and precipitation over multiple consecutive days. Meteorological characteristics are considered an important factor in calculating attention scores and weights. If the meteorological characteristics at two specific time points show a trend associated with icing risk, the corresponding attention weight is increased.

[0048] Assume that the input meteorological data sequence is ,in represents the meteorological feature vector after processing by the meteorological feature extraction layer on day i, including meteorological feature vectors such as temperature and precipitation. For day t, the data in the window is:

[0049] in, is the window size, Indicates a floor operation.

[0050] Calculating attention score The formula is:

[0051] in and Is the first i 、 j The feature vector output by the meteorological feature extraction layer of the day is the meteorological features of a single day after processing by the meteorological feature extraction layer; Q and K are the mapping functions of query and key respectively.

[0052] Attention weight for:

[0053] Where i is the core time point of current concern, that is, the meteorological feature vector of the i-th day in the window; j is the target time point to be compared in the window; and k is used to traverse and sum the index. is the attention score for the i-th and k-th day of the window.

[0054] In this embodiment, the spatial feature fusion layer integrates latitude and longitude position codes to mine geographic correlation features of observation points. Specifically, when generating the query matrix Q, key matrix K, and value matrix V, the cross-day time-dependent meteorological features output by the temporal feature extraction layer are combined with information such as longitude and latitude as part of the input sequence Y. This allows the model to consider the differences and correlations in meteorological conditions across different observation points when calculating attention scores and weights. In the longitude and latitude position codes, the meteorological features are encoded and then spliced ​​or weightedly fused with the longitude and latitude position codes to enhance the model's perception of geographic spatial relationships and meteorological conditions.

[0055] Assume that in the input sequence Y, the longitude and latitude information is represented as , the meteorological characteristics are expressed as ,in is the time-dependent meteorological feature of the i-th observation point output by the time feature extraction layer. Then Y= ,in It is a set of latitude and longitude codes of observation points, recording geographic spatial location information. It is the cross-day time-dependent meteorological characteristic data of the corresponding observation point after being processed by the time feature extraction layer.

[0056] Calculating attention score The formula is:

[0057] in 、 is the fused feature of the i-th and j-th elements in the input sequence Y; Q and K are the mapping functions of query and key respectively.

[0058] Attention weight for:

[0059] in, is the number of observation points, indicating that the attention scores of all observation points are normalized; k is the index of all observation points, ensuring that the sum of the attention scores of the global observation points is normalized.

[0060] In the longitude and latitude position coding, meteorological features are encoded and then spliced ​​or weighted fused with the longitude and latitude position coding to enhance the model's perception of geographic spatial relationships and meteorological conditions.

[0061] Assume that the longitude and latitude position is coded as , meteorological characteristics are coded as , then the fused code F can be expressed as:

[0062] in, is a weight coefficient used to control the fusion ratio of longitude and latitude location codes and meteorological feature codes. This coefficient can be adjusted according to actual conditions to flexibly adjust the influence of geographic information and meteorological characteristics on the model in different application scenarios.

[0063] like Figure 1 As shown, in step S400, a spatiotemporal fusion icing warning model based on a deep learning network is established and trained, the obtained spatiotemporal fusion features are input into the spatiotemporal fusion icing warning model, and an icing prediction result is output.

[0064] The above steps specifically include: The spatiotemporal fusion icing warning model uses a fully connected neural network architecture. The number of neurons in the input layer matches the dimension of the fused feature vector. Multiple hidden layers use the Reluctant Unit (ReLU) activation function, with the number of neurons dynamically adjusted based on the characteristics of the meteorological data and the warning requirements. The output layer uses a Sigmoid activation function to output the icing probability. The number of hidden layers and neurons is determined through hyperparameter tuning. A binary cross-entropy loss function is used to measure the difference between the predicted results and the true labels. The Adam optimizer is used to update parameters, and the learning rate is dynamically adjusted based on the validation set loss.

[0065] Among them, the ReLU activation function calculation formula is:

[0066] Among them, the formula of the Sigmoid activation function is:

[0067] The labeled dataset was divided into training set, validation set and test set by stratified sampling in the ratio of 7:2:1. The early stopping strategy was used to prevent overfitting during training. The training set was used to update parameters, the validation set was used to adjust hyperparameters, and the test set was used to evaluate performance. Accuracy, recall rate and F1 value were used for comprehensive evaluation.

[0068] like Figure 1 As shown, in step S500, a graded warning mechanism is established to perform warning classification on the obtained icing prediction results.

[0069] The above steps specifically include: In this embodiment, dynamic thresholds are set based on observation point clustering to make graded warning decisions. According to the icing probability output by the spatiotemporal fusion model and combined with the observation point clustering results, dynamic thresholds are set to divide warnings into different levels.

[0070] Corresponding to the above-disclosed method for warning ice formation on plateau railways based on sequence learning, the embodiment of the present invention further discloses a system for warning ice formation on plateau railways based on sequence learning, such as Figure 3 As shown, it specifically includes: The data acquisition and processing module is used to collect geographical information of the plateau railway area and multiple meteorological parameter data on a daily scale, and perform pre-processing; The icing label generation module is used to set the thresholds of various meteorological parameters and combine logical judgment to annotate the meteorological parameter data to generate icing labels; The spatiotemporal feature extraction module is used to establish a feature extraction and fusion model based on the hierarchical self-attention mechanism to extract features from daily meteorological parameter data, and fuse the extracted features with geographic location information to obtain spatiotemporal fusion features; An icing prediction module is used to establish and train a spatiotemporal fusion icing warning model based on a deep learning network, input the obtained spatiotemporal fusion features into the spatiotemporal fusion icing warning model, and output an icing prediction result; The graded warning module is used to establish a graded warning mechanism and perform warning classification on the obtained icing prediction results.

[0071] It should be noted that for a detailed description of a plateau railway icing warning method based on sequence learning provided in an embodiment of the present invention, reference can be made to the relevant description of a plateau railway icing warning method based on sequence learning provided in an embodiment of the present application, which will not be repeated here.

[0072] In addition, an embodiment of the present invention also provides an electronic device, comprising: a processor and a memory; the memory is used to store one or more program instructions; the processor is used to run one or more program instructions to execute the steps of a sequence learning-based plateau railway icing warning method as described in any of the above items.

[0073] It should be noted that, for a detailed description of an electronic device provided in an embodiment of the present invention, reference can be made to the relevant description of a plateau railway icing warning method based on sequence learning provided in an embodiment of the present application, which will not be repeated here.

[0074] In addition, an embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a sequence learning-based plateau railway icing warning method as described in any of the above items are implemented.

[0075] It should be noted that, for a detailed description of a computer-readable storage medium provided in an embodiment of the present invention, reference can be made to the relevant description of a plateau railway icing warning method based on sequence learning provided in an embodiment of the present application, which will not be repeated here.

[0076] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer program. When all or part of the functions in the above embodiments are implemented by computer program, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to implement the above functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented. In addition, when all or part of the functions in the above embodiments are implemented by computer program, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash disk or mobile hard disk, and saved in the memory of the local device by downloading or copying, or the system of the local device is updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0077] The above examples are used to illustrate the present invention, which are only used to help understand the present invention and are not intended to limit the present invention. Those skilled in the art can make several simple deductions, modifications or substitutions based on the concept of the present invention.

Claims

1. A plateau railway icing early warning method based on sequence learning, characterized in that: The method comprises: Collect geographical location information of the plateau railway area and multiple meteorological parameter data on a daily scale, and perform pre-processing; Set thresholds for various meteorological parameters and use logical judgment to label meteorological parameter data and generate icing labels; A feature extraction and fusion model based on a hierarchical self-attention mechanism is established to extract features from daily-scale meteorological parameter data, and the extracted features are fused with geographic location information to obtain spatiotemporal fusion features. Establishing and training a spatiotemporal fusion icing warning model based on a deep learning network, inputting the obtained spatiotemporal fusion features into the spatiotemporal fusion icing warning model, and outputting icing prediction results; A graded warning mechanism is established to classify the obtained icing prediction results into warning levels.

2. The plateau railway icing early warning method based on sequence learning according to claim 1, characterized in that: Collect geographical location information of the plateau railway area and multiple daily meteorological parameter data, and perform pre-processing, including: The geographical location information includes longitude and latitude information, which is used to identify the specific geographical location of the observation point; The meteorological parameter data includes temperature, precipitation, wind speed, humidity, air pressure and sunshine; temperature includes minimum temperature, average temperature and average surface temperature; precipitation refers to precipitation amount; wind speed refers to the average wind speed at a preset height; humidity refers to relative humidity; air pressure refers to surface air pressure; sunshine refers to sunshine hours; The collected data are preprocessed, including data cleaning, missing value processing and data standardization.

3. The plateau railway icing early warning method based on sequence learning according to claim 2, characterized in that: Set thresholds for various meteorological parameters and use logical judgment to label meteorological parameter data to generate icing labels, including: For precipitation scenarios, the freezing conditions are: when there is effective precipitation, that is, the precipitation exceeds the preset precipitation threshold, and the surface temperature is less than or equal to the preset surface temperature threshold, it is directly judged as freezing; or when there is effective precipitation, that is, the precipitation exceeds the preset precipitation threshold, the minimum temperature is less than or equal to the preset minimum temperature threshold, and the surface temperature is less than or equal to the preset surface temperature threshold, it is judged as freezing, to take into account the situation of secondary freezing due to snowmelt; For high humidity without precipitation, the freezing conditions are: when there is no effective precipitation, that is, the precipitation does not exceed the precipitation threshold, the surface temperature is less than or equal to the preset surface temperature threshold, the relative humidity is greater than or equal to the high humidity threshold, and the wind speed is less than the preset wind speed threshold, it is determined to be freezing; For extreme low temperature scenarios, the freezing condition is: when the minimum temperature is less than or equal to the extreme low temperature threshold, it is considered freezing even if there is no effective precipitation, that is, the precipitation does not exceed the precipitation threshold; If the icing conditions in each scene are not met, it is determined to be non-icing; In addition, suppression conditions are also considered: if the sunshine hours are greater than or equal to the preset sunshine hours threshold, or the average wind speed is greater than or equal to the preset wind speed threshold, it is considered that the icing risk will be offset. Even if the icing conditions in each scenario are met, it is still judged as non-icing.

4. The method for early warning of icing on plateau railways based on sequence learning according to claim 1, characterized in that: A feature extraction and fusion model based on the hierarchical self-attention mechanism is established to extract features from daily meteorological parameter data, and the extracted features are fused with geographic location information to obtain spatiotemporal fusion features, including: The feature extraction and fusion model based on the hierarchical self-attention mechanism includes a meteorological feature extraction layer, a temporal feature extraction layer, and a spatial feature fusion layer, and the output of the previous layer will serve as the input of the next layer; The meteorological feature extraction layer is used to extract single-day meteorological features from the collected daily-scale meteorological parameter data, and the output single-day meteorological feature vector is used as the input of the time feature extraction layer; The temporal feature extraction layer is used to extract inter-day meteorological dependency features using sliding window self-attention, and the output inter-day time-dependent meteorological features serve as input to the spatial feature fusion layer; The spatial feature fusion layer is used to splice or weightedly fuse the cross-daily time-dependent meteorological features with the latitude and longitude position codes, and the final obtained spatiotemporal fusion features contain time-dependent and geographically associated meteorological information.

5. The method for early warning of icing on plateau railways based on sequence learning according to claim 4, characterized in that: The meteorological feature extraction layer is used to extract meteorological features from the collected daily-scale meteorological parameter data, specifically including: The meteorological feature extraction layer extracts features from the collected daily-scale meteorological data based on the improved VGG16 model. First, the meteorological parameter data is encoded and converted into a format suitable for convolutional neural network processing. The change patterns of different meteorological parameters in the data are obtained as underlying features. The improved VGG16 model is used to extract deep features of the data and explore the potential correlations between different meteorological parameters. Finally, the underlying features are fused with the deep features and output on the Softmax.

6. The plateau railway icing early warning method based on sequence learning according to claim 5, characterized in that: The improved VGG16 model includes five convolutional blocks: C1, C2, C3, C4, and C5, and the output of C1 is used as the input of C2, the output of C2 is used as the input of C3, the output of C3 is used as the input of C4, and the output of C4 is used as the input of C5; Among them, the convolution blocks C1 and C2 both contain two convolution layers B, two activation functions and a pooling layer P. Specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the second convolution layer B is used as the input of the pooling layer P. The C3 convolutional block contains three convolutional layers B, three activation functions, and a pooling layer P. Specifically, an activation function is added after each convolutional layer B, and a pooling layer P is added at the end of the convolutional block, so that the output of the third convolutional layer B serves as the input of P. The convolutional layer of the C3 convolutional block uses 256 3*3 convolution kernels to better capture the complex characteristics of meteorological data and the potential correlation between different meteorological parameters. The C4 and C5 convolution blocks both contain three convolution layers B, three activation functions, and a pooling layer P. Specifically, an activation function is added after each convolution layer B, and a pooling layer P is added at the end of the convolution block, so that the output of the third convolution layer B serves as the input of the pooling layer P. The convolution kernel size of the convolution layer B is 3*3, and the moving step is 1; the pooling layer P uses the maximum pooling method, sets the pooling window size to 2*2, and the moving step is 2; the activation function uses the SeLU function.

7. The method for early warning of icing on plateau railways based on sequence learning according to claim 1, characterized in that: Establish and train a spatiotemporal fusion icing warning model based on a deep learning network, input the obtained spatiotemporal fusion features into the spatiotemporal fusion icing warning model, and output icing prediction results, specifically including: The spatiotemporal fusion icing warning model adopts a fully connected neural network architecture, in which the number of neurons in the input layer is the same as the dimension of the spatiotemporal fusion feature vector; multiple hidden layers use the ReLU activation function, and the number of neurons is dynamically adjusted according to the characteristics of meteorological data and warning needs; the output layer uses the Sigmoid activation function to output the icing probability; the number of hidden layers and neurons is determined by hyperparameter tuning methods; and the binary cross entropy loss function is used to measure the difference between the predicted results and the true labels. Construct training sets, validation sets, and test sets to train, validate, and test the model. Use the Adam optimizer to update parameters, and the learning rate is dynamically adjusted according to the validation set loss.

8. The method for early warning of icing on plateau railways based on sequence learning according to claim 1, characterized in that: Establish a graded warning mechanism to classify the icing prediction results, specifically including: Dynamic thresholds are set based on observation point clustering to make graded warning decisions. The warnings are divided into different levels according to the icing probability output by the spatiotemporal fusion icing warning model and the dynamic thresholds set based on the observation point clustering results.

9. A plateau railway icing warning system based on sequence learning, characterized by: The system comprises: The data acquisition and processing module is used to collect geographical information of the plateau railway area and multiple meteorological parameter data on a daily scale, and perform pre-processing; The icing label generation module is used to set the thresholds of various meteorological parameters and combine logical judgment to annotate the meteorological parameter data to generate icing labels; The spatiotemporal feature extraction module is used to establish a feature extraction and fusion model based on the hierarchical self-attention mechanism to extract features from daily meteorological parameter data, and fuse the extracted features with geographic location information to obtain spatiotemporal fusion features; An icing prediction module is used to establish and train a spatiotemporal fusion icing warning model based on a deep learning network, input the obtained spatiotemporal fusion features into the spatiotemporal fusion icing warning model, and output an icing prediction result; The graded warning module is used to establish a graded warning mechanism and perform warning classification on the obtained icing prediction results.

10. An electronic device, characterized in that: The device includes: a processor and a memory; The memory is used to store one or more program instructions; The processor is used to run one or more program instructions to execute the steps of a plateau railway icing warning method based on sequence learning as described in any one of claims 1 to 8.

Images