Road surface icing monitoring system based on LSTM

Abstract

The invention discloses a road surface icing monitoring system based on LSTM, and the system comprises a sensor sensing layer which comprises a temperature and humidity probe and is used for collecting the temperature and humidity data of a road surface; the communication transmission layer is used for uploading the data acquired by the sensor sensing layer to the cloud processing layer; an LSTM-based prediction model is built in the cloud processing layer, and the icing probability in a certain time in the future is predicted; and the user interaction layer adopts a dynamic electronic map to mark monitoring point positions and dynamically display monitoring data in real time. According to the invention, software and hardware are combined to construct an integrated system for monitoring, transmitting, applying and interacting road surface temperature and humidity data; through multi-factor composite analysis, a long-short-term memory network model is constructed, and the icing probability in a certain time in the future is predicted in combination with monitoring data; monitoring point positions, data and a GIS map are combined, and a visual interaction interface is established to achieve one-map unified management. The dynamic electronic map displays the state of the monitoring point in real time and can backtrack the change of monitoring data, so that dynamic map interaction is realized and the management efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of highway management, and in particular to a road icing monitoring system based on LSTM. Background Art

[0002] Pavement temperature is a key factor affecting road safety. When temperatures are low, ice and snow may form on the road surface, significantly reducing the friction coefficient and significantly increasing vehicle braking distances. This can easily lead to vehicle loss of control, skidding, and subsequent rear-end collisions. Traffic accidents caused by icy roads are frequent in winter. Existing methods for monitoring road surface temperature suffer from numerous shortcomings. Traditional contact thermometers are inefficient and require regular, on-site data collection. This is labor-intensive and inefficient, as well as incapable of real-time, continuous monitoring. Rapid changes in road surface temperature are difficult to detect. In situations with significant diurnal temperature swings or sudden weather changes, road maintenance departments cannot be provided with up-to-date temperature information to facilitate the timing of maintenance operations. Furthermore, while some monitoring equipment can capture temperature data, data transmission is untimely and processing capabilities are limited. This makes it impossible to deeply analyze historical road surface temperature changes and predict the likelihood of icing based on historical data. Consequently, this equipment fails to effectively support traffic management decisions. Consequently, response measures to snow and icing are delayed, posing a threat to road use and driving safety.

[0003] Currently, a Chinese patent with the announcement number CN201710031475.0 discloses an intelligent road surface monitoring and early warning system. The technical problem to be solved by the present invention is to monitor the snow accumulation on the road in real time and clear the snow in time when the snow reaches a certain thickness.

[0004] The above-mentioned existing technologies can monitor road conditions, which is conducive to timely clearing snow from the road after snowfall. However, the above-mentioned inventions cannot monitor road icing conditions, which is not conducive to monitoring road conditions when the temperature reaches below zero. Therefore, it is necessary to provide a road icing monitoring and early warning method based on a long short-term memory network (LSTM). Summary of the Invention

[0005] The purpose of this invention is to provide a road icing monitoring system based on LSTM, which solves the defects of the existing technology through the following innovations:

[0006] 1. Build an integrated system for road surface temperature and humidity data monitoring, transmission, application, and interaction through the combination of software and hardware;

[0007] 2. Through multi-factor composite analysis (temperature, humidity, historical data), a long-short-term memory network model is constructed and combined with monitoring data to predict the probability of ice formation within a certain period of time in the future;

[0008] 3. Combine monitoring points, data and GIS maps to establish a visual interactive interface to achieve "one map, one management".

[0009] The technical solution of the present invention is:

[0010] The LSTM-based road icing monitoring system includes:

[0011] The sensor perception layer includes temperature and humidity probes to collect road surface temperature and humidity data;

[0012] The communication transmission layer uploads the data collected by the sensor perception layer to the cloud processing layer;

[0013] The cloud processing layer has a built-in LSTM-based prediction model to predict the probability of icing within a certain period of time in the future;

[0014] The user interaction layer uses a dynamic electronic map to mark monitoring points and display monitoring data in real time.

[0015] Preferably, the prediction model of the cloud processing layer is trained by:

[0016] Data acquisition: Collect historical temperature and humidity data from the sensor perception layer, and record whether there is ice formation within a certain period of time in the future at the corresponding moment;

[0017] Data preprocessing: Detect and process missing values and outliers in historical temperature and humidity data;

[0018] LSTM layer design: Build a neural network consisting of an input layer, multiple LSTM hidden layers, and an output layer. The input layer receives temperature and humidity data over a certain period of time. The LSTM hidden layer captures the long-term and short-term dependencies of the time series. The output layer uses a Sigmoid activation function to output a freezing probability value between 0 and 1.

[0019] Model compilation: select binary cross entropy as the loss function; use the Adam optimizer to update the model parameters; and use accuracy as the evaluation metric;

[0020] Model training: The model is trained over multiple rounds using the training set, with the number of rounds and batch size set. In each round, the model calculates a predicted value based on the input data, then uses a loss function to calculate the error between the predicted value and the true value. Finally, the optimizer is used to update the model parameters.

[0021] Hyperparameter tuning: Use grid search to tune the hyperparameters of the LSTM layer. For each combination in the hyperparameter grid, train the model using the training set and evaluate the model performance on the validation set.

[0022] Model evaluation: Use the test set to perform a final evaluation of the trained model, calculate the model's recall rate, precision rate, and F1 value, and comprehensively evaluate the model's performance;

[0023] Save Model: Compare the performance of models with different hyperparameters and save the prediction model with the best performance.

[0024] Preferably, the prediction model of the cloud processing layer includes the following prediction method:

[0025] Data acquisition: Get the temperature and humidity data of the current time period;

[0026] Data preprocessing: Clean and normalize the acquired data to make it have the same format and range as the training data;

[0027] Load model: load the previously trained and saved LSTM model;

[0028] Probabilistic prediction: input the preprocessed data into the model to obtain the probability of icing at a certain time in the future;

[0029] Trigger condition judgment: If the predicted icing probability is higher than the set threshold, a preliminary alarm is triggered; otherwise, no alarm is triggered;

[0030] Data update: Based on the results of manual on-site confirmation, new data is added to the training dataset;

[0031] Model optimization: Retrain the model using the updated training dataset to improve the model's predictive accuracy.

[0032] Preferably, the user interaction layer adopts a dynamic electronic map, introduces a GIS map, and the digital display color follows the pollution level color specification; the monitoring time bar is pulled to retrospectively display the data of the monitoring point, providing intuitive materials for reviewing the evolution of road conditions.

[0033] Preferably, the user interaction layer is further provided with a data analysis panel to continuously collect monitoring data, form historical data and outline a historical trend graph to show the changing patterns of temperature and humidity in the road section.

[0034] Preferably, in data preprocessing, missing values are filled using linear interpolation; outliers are corrected or eliminated using a set threshold range;

[0035] Missing data points are filled using linear interpolation, the formula is:

[0036] ;

[0037] in, is the missing value at time t, is the missing time interval;

[0038] Based on the 3σ principle, data outside the following ranges are eliminated:

[0039] ;

[0040] in, 、 is the mean, 、 is the standard deviation;

[0041] After data preprocessing, a sliding window is constructed with a window length of L hours, a step size of S = 1 hour, an input dimension of D = 2 (temperature T, humidity H), and an output label of a certain future freezing state (0 / 1).

[0042] The data format of each window is:

[0043] ;

[0044] Use Min-Max scaling to normalize the temperature and humidity data to the [0,1] interval to speed up the convergence of the model. The formula is:

[0045] ;

[0046] in, 、 is the maximum and minimum value of the training set;

[0047] Divide the processed data into the following proportions: 7:1.5:1.5 in chronological order:

[0048] Training set: used for model training;

[0049] Validation set: used for hyperparameter tuning and early stopping;

[0050] Test set: used for final performance evaluation.

[0051] Preferably, during model training, after each round of training, the model is verified using a validation set to monitor the model performance; when the performance of the validation set does not decrease over multiple consecutive training rounds, the training is terminated early and the model parameters are restored to the minimum loss function value of the validation set;

[0052] In hyperparameter tuning, grid search is used to tune the number of neurons, dropout rate, and window length of the LSTM layer. The value ranges of each hyperparameter are as follows:

[0053] Number of LSTM units: 64, 128, 256;

[0054] Dropout rate: 0.2, 0.3;

[0055] Window length: 48h, 72h, 96h;

[0056] For each combination of hyperparameters in the grid, the model is trained using the training set and the performance of the model is evaluated on the validation set.

[0057] Preferably, in model evaluation, the method for calculating the recall rate, precision rate, and F1 value of the model includes:

[0058] List the confusion matrix: true positives TP, false negatives FN, false positives FP, false negatives TN;

[0059] Recall : measure the model's ability to detect actual icing events and avoid missed reports,

[0060] ;

[0061] Accuracy : Measure the proportion of real icing events predicted as icing events to control false alarms,

[0062] ,

[0063] F1 value: The harmonic mean of precision and recall, which comprehensively considers the accuracy and recall ability of the model.

[0064] ;

[0065] Based on the evaluation indicators on the validation set, the hyperparameter combination with the best performance is selected.

[0066] Preferably, in the prediction method of the LSTM model,

[0067] First, obtain the current 72-hour historical temperature, humidity, and time series characteristics, temperature and humidity sequence ;

[0068] Fill missing values and remove outliers from the data to generate input tensors ;

[0069] Use the Min-Max parameters of the training set to perform real-time data scaling and normalization to make it have the same format and range as the training data;

[0070] Load the previously trained and saved LSTM model, input the preprocessed data into the loaded model, and predict the icing probability. The prediction result is a probability value between 0 and 1, indicating the icing probability in the next two hours.

[0071] If the predicted icing probability is ≥70%, a preliminary alarm is triggered; otherwise, no alarm is triggered.

[0072] Preferably, the prediction method of the LSTM model further includes a manual confirmation mechanism:

[0073] When a preliminary alarm is triggered, a work order is automatically generated and pushed to the mobile terminal, notifying relevant personnel to retrieve road surveillance cameras, view real-time images at monitoring points, and check for road reflections or signs of vehicle slippage. Alternatively, personnel can carry out on-site measurements with a portable temperature and humidity meter to verify temperature and humidity errors and freezing time.

[0074] The system records the response time and initiates the de-icing plan. If there is a false alarm, the result will be recorded.

[0075] Preferably, the temperature and humidity probe of the sensor sensing layer is buried at a depth of 30-50 cm, protected by a stainless steel shell, with a built-in temperature compensation algorithm, and the probe surface is covered with an anti-radiation coating to reduce the temperature drift error caused by direct sunlight. It has a built-in self-test chip to monitor the battery power, voltage, and signal strength in real time, and trigger a local sound and light alarm when an abnormality occurs.

[0076] Preferably, the communication transmission layer adopts NB-IoT / 4G dual-mode communication module, supports Band5 / Band8 frequency bands, and adapts to a working environment of -40℃~85℃.

[0077] Preferably, the cloud processing layer also includes a data storage module, which uses a time series database and stores data partitioned by road section and timestamp.

[0078] The advantages of the present invention are:

[0079] 1. The present invention combines software and hardware to build an integrated system for monitoring, transmitting, applying and interacting with road surface temperature and humidity data;

[0080] 2. This invention uses a multi-factor composite analysis (temperature, humidity, historical data) to construct a long-short-term memory network model, combined with monitoring data, to predict the probability of freezing within a certain period of time in the future;

[0081] 3. The present invention combines monitoring points, data and GIS maps to establish a visual interactive interface to achieve "one map for all management". The dynamic electronic map displays the status of monitoring points in real time (color grading), can trace back changes in monitoring data within two hours, realize dynamic map interaction, and automatically generate historical data and alarm tables to improve management efficiency.

[0082] 4. The present invention adopts NB-IoT / 4G dual-mode communication module to support full network coverage and low-power transmission, and can achieve continuous monitoring.

[0083] 5. The icing risk prediction based on the LSTM model of the present invention can issue an alarm 2 hours in advance, leaving sufficient time for emergency response to road icing. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0085] Figure 1 This is an architecture diagram of the LSTM-based road icing monitoring system of the present invention;

[0086] Figure 2 Flowchart for LSTM model training;

[0087] Figure 3 Prediction flowchart for the LSTM model. DETAILED DESCRIPTION

[0088] Example 1

[0089] like Figure 1 As shown, the LSTM-based road icing monitoring system of the present invention has a system composition and functions including a sensor perception layer, a communication transmission layer, a cloud processing layer and a user interaction layer.

[0090] 1. Sensor perception layer

[0091] (1) The sensor layer uses a highly sensitive temperature and humidity probe: the probe is buried at a depth of 30-50 cm and is protected by a stainless steel casing (IP68 rating). It collects data every 5 minutes. It has a built-in temperature compensation algorithm and a measurement accuracy of ±0.3°C for temperature and ±2%RH for humidity.

[0092] (2) The probe is designed to be anti-interference: the probe surface is covered with an anti-radiation coating to reduce the temperature drift error caused by direct sunlight.

[0093] (3) Equipment health monitoring: Built-in self-test chip monitors battery power, voltage, and signal strength in real time, and triggers local sound and light alarms when abnormalities occur.

[0094] 2. Communication transport layer

[0095] The communication transmission layer adopts NB-IoT / 4G dual-mode communication module: supports Band5 / Band8 frequency bands and adapts to -40℃~85℃ working environment;

[0096] 3. Cloud processing layer

[0097] (1) Data storage: Use a time series database (InfluxDB 2.0) and partition the data by road section and timestamp;

[0098] (2) Road icing prediction: including LSTM model construction and LSTM model prediction.

[0099] like Figure 2 As shown in Figure 2, the method flow for building the LSTM model is as follows.

[0100] Data collection: Collects historical temperature and humidity data from sensors, and also records whether ice will form in the next two hours. Label: Whether ice will form in the next two hours (0 - no, 1 - yes).

[0101] Data preprocessing: Detect and handle missing values and outliers in the data.

[0102] LSTM layer design: Build a neural network consisting of an input layer, multiple LSTM hidden layers, and an output layer. The input layer receives 72 hours of temperature and humidity data (dimension 72×2). The LSTM hidden layer captures the long-term and short-term dependencies of the time series. The output layer uses a sigmoid activation function to output a freezing probability value between 0 and 1.

[0103] Model compilation: Select binary cross entropy as the loss function; use the Adam optimizer to update the model parameters; and use accuracy as the evaluation metric.

[0104] Model training: The model is trained over multiple rounds using the training set, with a set number of epochs and batch size. In each round, the model calculates a predicted value based on the input data, then uses a loss function to calculate the error between the predicted value and the true value. Finally, an optimizer is used to update the model parameters.

[0105] If the model's performance on the validation set no longer improves, overfitting may occur. In this case, you can use the early stopping strategy to terminate training early.

[0106] Hyperparameter tuning: Grid search is used to tune the hyperparameters of the LSTM layer. For each combination of hyperparameters in the grid, the model is trained using the training set and the performance of the model is evaluated on the validation set.

[0107] Model evaluation: Use the test set to perform a final evaluation of the trained model, calculate the model's recall rate, precision rate, F1 value and other indicators, and comprehensively evaluate the model's performance.

[0108] Save Model: Compare the performance of models with different hyperparameters and save the prediction model with the best performance.

[0109] like Figure 3 As shown in Figure 2, the LSTM model prediction method flow is as follows.

[0110] S1. Data input and preprocessing

[0111] Data acquisition: Get the current 72-hour historical temperature and humidity data.

[0112] Data preprocessing: Clean and normalize the acquired data so that it has the same format and range as the training data.

[0113] S2. Model loading and prediction

[0114] Load model: Load the previously trained and saved LSTM model.

[0115] Probabilistic prediction: The preprocessed data is input into the model to obtain the probability of icing in the next 2 hours.

[0116] S3. Alarm judgment

[0117] Trigger condition judgment: If the predicted icing probability is ≥ 70%, the preliminary alarm is triggered; otherwise, no alarm is triggered.

[0118] S4. Manual confirmation

[0119] On-site inspection: When a preliminary alarm is triggered, relevant personnel are notified to check through surveillance cameras or on-site to see if there is indeed an icing risk.

[0120] Result record: record the results of manual on-site confirmation (whether there is ice).

[0121] S5. Result Feedback and Model Update

[0122] Data update: Based on the results of manual on-site confirmation, new data is added to the training dataset.

[0123] Model optimization: Retrain the model using the updated training dataset to improve the model's predictive accuracy.

[0124] 4. User interaction layer

[0125] The user interaction layer includes dynamic electronic maps, data analysis panels and export reports.

[0126] Dynamic electronic maps: GIS maps are introduced to accurately mark monitoring points and display monitoring data in real time. The digital display colors follow the pollution level color standard (green: normal; yellow: warning; red: alarm), allowing users to instantly focus on road sections with poor environmental conditions. Pulling the monitoring time bar displays data from monitoring points within two hours, providing intuitive materials for reviewing road condition evolution.

[0127] Data analysis panel: Continuously collect monitoring data, form historical data and outline historical trend charts, showing the changing patterns of temperature and humidity in road sections, providing a scientific basis for road maintenance, traffic planning, etc.

[0128] Export reports: Generate daily, weekly, monthly, and annual reports in PDF / Excel format, including mean, extreme values, and forecast results.

[0129] Example 2

[0130] This embodiment introduces an example of building an LSTM model.

[0131] 1. Data preprocessing:

[0132] Detect and process missing values and outliers in the data. For missing values, linear interpolation can be used to fill them; for outliers, they can be corrected or removed by setting a threshold range.

[0133] Missing data points are filled using linear interpolation, the formula is:

[0134] ;

[0135] in, is the missing value at time t, is the missing time interval.

[0136] Based on the 3σ principle, data outside the following ranges are eliminated:

[0137] ;

[0138] in, 、 is the mean, 、 is the standard deviation.

[0139] 2. Sliding window construction

[0140] The window length L = 72 hours, the step size S = 1 hour, the input dimension D = 2 (temperature, humidity), and the output label is the freezing state in the next 2 hours (0 / 1).

[0141] The data format of each window is:

[0142] ;

[0143] Use Min-Max scaling to normalize the temperature and humidity data to the [0,1] interval to speed up the convergence of the model. The formula is:

[0144] ;

[0145] in, 、 are the maximum and minimum values of the training set.

[0146] Divide the processed data into the ratio of 7:1.5:1.5:

[0147] Training set (data from the first 8 months of history) 70%: used for model training;

[0148] Validation set (data from the middle two months) 15%: used for hyperparameter tuning and early stopping;

[0149] Test set (data from the last two months) 15%: used for final performance evaluation.

[0150] 3. Model construction:

[0151] 3.1 Network Structure Design

[0152] Build a neural network consisting of an input layer, two LSTM hidden layers, and an output layer.

[0153] The input layer receives 72 hours of temperature and humidity data (dimension is 72×2).

[0154] The LSTM hidden layer is used to capture the long-term and short-term dependencies of the time series. The first layer of 128 units extracts short-term dependencies (hourly temperature and humidity fluctuations); the second layer of 64 units captures long-term trends (diurnal temperature difference, seasonal changes).

[0155] Dropout layer: random dropout rate of 0.2 to prevent overfitting and improve generalization ability;

[0156] Output layer: Use the Sigmoid activation function to output an icing probability value between 0 and 1.

[0157] 3.2 Loss Function and Optimizer

[0158] Binary Cross Entropy

[0159] ;

[0160] Where N is the number of samples, is the true label (0 or 1), is the predicted probability.

[0161] Adam optimizer, parameters are:

[0162] Learning rate ; Momentum parameter ; Weight decay .

[0163] 4. Model training

[0164] 4.1 Use the training set to train the model for multiple rounds, setting the number of training rounds (epochs) and batch size (batchsize). In each round of training, the model calculates the predicted value based on the input data, then uses the loss function to calculate the error between the predicted value and the true value, and finally uses the optimizer to update the model parameters.

[0165] After each round of training, the model is validated using the validation set to monitor its performance. If the performance on the validation set does not decrease over 10 consecutive training rounds (with a patience value of 10), training is terminated early and the model parameters are restored to the minimum loss function value on the validation set.

[0166] 4.2 Hyperparameter Tuning

[0167] Use grid search to tune the number of neurons, dropout rate, and window length of the LSTM layer. The value ranges of each hyperparameter are as follows:

[0168] Number of LSTM units: 64, 128, 256;

[0169] Dropout rate: 0.2, 0.3;

[0170] Window length: 48h, 72h, 96h.

[0171] For each combination of hyperparameters in the grid, the model is trained using the training set and the performance of the model is evaluated on the validation set.

[0172] 5. Model Evaluation

[0173] Use the test set to perform a final evaluation of the trained model, calculate the model's recall rate, precision rate, F1 value and other indicators, and comprehensively evaluate the model's performance.

[0174] Construct confusion matrix: as shown in Table 1.

[0175] Table 1 Confusion matrix

[0176] Prediction is positive (freezing) The prediction is negative (no ice) Actual positive example TP (True Positive) FN (False Negative) Actually a negative example FP (False Positive) TN (true negative example)

[0177] Recall: measures the model's ability to detect actual icing events and avoid false negatives.

[0178] ;

[0179] Precision: measures the proportion of actual icing events predicted as icing events and controls false alarms.

[0180] ;

[0181] F1 value: The harmonic mean of precision and recall, which comprehensively considers the accuracy and recall ability of the model.

[0182] ;

[0183] Based on the evaluation indicators on the validation set, the hyperparameter combination with the best performance is selected.

[0184] Example 3

[0185] This example introduces an example of LSTM model prediction.

[0186] 1. Data input and preprocessing:

[0187] Get the current 72-hour historical temperature, humidity, time series characteristics, and temperature and humidity sequence ;

[0188] Fill missing values and remove outliers from the data to generate input tensors .

[0189] The data is scaled and normalized on the fly using the Min-Max parameters of the training set to have the same format and range as the training data.

[0190] 2. Model loading and prediction:

[0191] Load the previously trained and saved LSTM model and input the preprocessed data into the loaded model to predict the icing probability. The prediction result is a probability value between 0 and 1, indicating the icing probability in the next two hours.

[0192] 3. Alarm judgment:

[0193] Trigger condition judgment: If the predicted icing probability is ≥ 70%, the preliminary alarm is triggered; otherwise, no alarm is triggered.

[0194] 4. Manual confirmation mechanism:

[0195] When a preliminary alarm is triggered, a work order is automatically generated and pushed to the mobile terminal, notifying relevant personnel to retrieve the road monitoring camera, view the real-time image of the monitoring point, check for road reflections or signs of vehicle slippage; or carry a portable temperature and humidity meter for on-site measurement to verify temperature and humidity errors and freezing time.

[0196] The system records the response time and initiates the de-icing plan. If there is a false alarm, the result will be recorded.

[0197] 5. Result feedback and model update:

[0198] Data update: Based on the results of manual on-site confirmation, new temperature and humidity data and the corresponding icing labels (0 or 1) are added to the training dataset.

[0199] Model optimization: Retrain the model using the updated training dataset to improve the model's predictive accuracy.

[0200] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any modifications made based on the spirit of the main technical solution of the present invention shall be included in the scope of protection of the present invention.

Claims

1. The LSTM-based road icing monitoring system is characterized by: include: The sensor perception layer includes temperature and humidity probes to collect road surface temperature and humidity data; The communication transmission layer uploads the data collected by the sensor perception layer to the cloud processing layer; The cloud processing layer has a built-in LSTM-based prediction model to predict the probability of icing within a certain period of time in the future; The user interaction layer uses a dynamic electronic map to mark monitoring points and display monitoring data in real time.

2. The LSTM-based road icing monitoring system according to claim 1, characterized in that: The prediction model of the cloud processing layer, and its training method includes: Data acquisition: Collect historical temperature and humidity data from the sensor perception layer, and record whether there is ice formation within a certain period of time in the future at the corresponding moment; Data preprocessing: Detect and process missing values and outliers in historical temperature and humidity data; LSTM layer design: Build a neural network consisting of an input layer, multiple LSTM hidden layers, and an output layer. The input layer receives temperature and humidity data over a certain period of time. The LSTM hidden layer captures the long-term and short-term dependencies of the time series. The output layer uses a Sigmoid activation function to output a freezing probability value between 0 and 1. Model compilation: select binary cross entropy as the loss function; use the Adam optimizer to update the model parameters; and use accuracy as the evaluation metric; Model training: The model is trained over multiple rounds using the training set, with the number of rounds and batch size set. In each round, the model calculates a predicted value based on the input data, then uses a loss function to calculate the error between the predicted value and the true value. Finally, the optimizer is used to update the model parameters. Hyperparameter tuning: Use grid search to tune the hyperparameters of the LSTM layer. For each combination in the hyperparameter grid, train the model using the training set and evaluate the model performance on the validation set. Model evaluation: Use the test set to perform a final evaluation of the trained model, calculate the model's recall rate, precision rate, and F1 value, and comprehensively evaluate the model's performance; Save Model: Compare the performance of models with different hyperparameters and save the prediction model with the best performance.

3. The LSTM-based road icing monitoring system according to claim 2, characterized in that: The prediction model of the cloud processing layer, and its prediction method include: Data acquisition: Get the temperature and humidity data of the current time period; Data preprocessing: Clean and normalize the acquired data to make it have the same format and range as the training data; Load model: load the previously trained and saved LSTM model; Probabilistic prediction: input the preprocessed data into the model to obtain the probability of icing at a certain time in the future; Trigger condition judgment: If the predicted icing probability is higher than the set threshold, a preliminary alarm is triggered; otherwise, no alarm is triggered; Data update: Based on the results of manual on-site confirmation, new data is added to the training dataset; Model optimization: Retrain the model using the updated training dataset to improve the model's predictive accuracy.

4. The LSTM-based road icing monitoring system according to claim 1, characterized in that: The user interaction layer adopts a dynamic electronic map, introduces a GIS map, and the digital display color follows the pollution level color specification; by pulling the monitoring time bar, the data of the monitoring point is displayed retrospectively, providing intuitive materials for reviewing the evolution of road conditions.

5. The LSTM-based road icing monitoring system according to claim 4, characterized in that: The user interaction layer also provides a data analysis panel to continuously collect monitoring data, form historical data and outline historical trend charts to show the temperature and humidity change patterns of road sections.

6. The LSTM-based road icing monitoring system according to claim 2, characterized in that: In data preprocessing, missing values are filled using linear interpolation; outliers are corrected or eliminated using the set threshold range; Missing data points are filled using linear interpolation, the formula is: ； in, is the missing value at time t, is the missing time interval; Based on the 3σ principle, data outside the following ranges are eliminated: ； in, 、 is the mean, 、 is the standard deviation; After data preprocessing, a sliding window is constructed with a window length of L hours, a step size of S = 1 hour, an input dimension of D = 2 (temperature T, humidity H), and an output label of a certain future freezing state (0 / 1). The data format of each window is: ； Use Min-Max scaling to normalize the temperature and humidity data to the [0,1] interval to speed up the convergence of the model. The formula is: ； in, 、 is the maximum and minimum value of the training set; Divide the processed data into the following proportions: 7:1.5:1.5 in chronological order: Training set: used for model training; Validation set: used for hyperparameter tuning and early stopping; Test set: used for final performance evaluation.

7. The LSTM-based road icing monitoring system according to claim 6, characterized in that: During model training, after each round of training, the model is verified using the validation set to monitor the model performance. If the performance of the validation set does not decrease over multiple consecutive training rounds, the training is terminated early and the model parameters are restored to the minimum loss function value of the validation set. In hyperparameter tuning, grid search is used to tune the number of neurons, dropout rate, and window length of the LSTM layer. The value ranges of each hyperparameter are as follows: Number of LSTM units: 64, 128, 256; Dropout rate: 0.2, 0.3; Window length: 48h, 72h, 96h; For each combination of hyperparameters in the grid, the model is trained using the training set and the performance of the model is evaluated on the validation set.

8. The LSTM-based road icing monitoring system according to claim 7, characterized in that: In model evaluation, the methods for calculating the model's recall, precision, and F1 value include: List the confusion matrix: true positives TP, false negatives FN, false positives FP, false negatives TN; Recall : measure the model's ability to detect actual icing events and avoid missed reports, ； Accuracy : Measure the proportion of real icing events predicted as icing events to control false alarms, ， F1 value: The harmonic mean of precision and recall, which comprehensively considers the accuracy and recall ability of the model. ； Based on the evaluation indicators on the validation set, the hyperparameter combination with the best performance is selected.

9. The LSTM-based road icing monitoring system according to claim 3, characterized in that: In the prediction method of the LSTM model, First, obtain the current 72-hour historical temperature, humidity, and time series characteristics, temperature and humidity sequence ; Fill missing values and remove outliers from the data to generate input tensors ; Use the Min-Max parameters of the training set to perform real-time data scaling and normalization to make it have the same format and range as the training data; Load the previously trained and saved LSTM model, input the preprocessed data into the loaded model, and predict the icing probability. The prediction result is a probability value between 0 and 1, indicating the icing probability in the next two hours. If the predicted icing probability is ≥70%, a preliminary alarm is triggered; otherwise, no alarm is triggered.

10. The LSTM-based road icing monitoring system according to claim 9, characterized in that: The prediction method of the LSTM model also includes a manual confirmation mechanism: When a preliminary alarm is triggered, a work order is automatically generated and pushed to the mobile terminal, notifying relevant personnel to retrieve road surveillance cameras, view real-time images at monitoring points, and check for road reflections or signs of vehicle slippage. Alternatively, personnel can carry out on-site measurements with a portable temperature and humidity meter to verify temperature and humidity errors and freezing time. The system records the response time and initiates the de-icing plan. If there is a false alarm, the result will be recorded.

11. The LSTM-based road icing monitoring system according to claim 1, characterized in that: The temperature and humidity probes in the sensor sensing layer are buried at a depth of 30-50 cm, protected by a stainless steel shell, equipped with a built-in temperature compensation algorithm. The probe surface is covered with an anti-radiation coating to reduce temperature drift errors caused by direct sunlight. A built-in self-test chip monitors battery power, voltage, and signal strength in real time, triggering local sound and light alarms in the event of an abnormality.

12. The LSTM-based road icing monitoring system according to claim 1, characterized in that: The communication transmission layer adopts NB-IoT / 4G dual-mode communication module, supports Band5 / Band8 frequency bands, and adapts to -40℃~85℃ working environment.

13. The LSTM-based road icing monitoring system according to claim 1, characterized in that: The cloud processing layer also includes a data storage module, which uses a time series database and stores data partitioned by road section and timestamp.

Images