Rail transit rail surface dehumidification method and system based on multi-modal data fusion

Abstract

The application provides a track surface dehumidification method and system based on multi-modal data fusion, which comprises the following steps: collecting data related to the track surface state, including sensor data, including humidity sensor data, temperature sensor data, thickness detection sensor data; visual detection data, including running computer vision algorithm on real-time track surface image to obtain track surface state and distribution range of ice and water; meteorological data, including real-time data of weather stations along the line or weather forecast services; line topology data, including line information of the current running section of the train; inputting the collected data related to the track surface state into a pre-trained multi-modal data fusion and intelligent decision-making model for prediction, and outputting an optimal dehumidification strategy; inputting the optimal dehumidification strategy into a dehumidification execution module in the form of an instruction to execute high-pressure dry air dehumidification.

Description

Technical Field

[0001] This invention relates to the field of rail transit track surface dehumidification technology, specifically to a rail transit track surface dehumidification method and system based on multimodal data fusion. Background Technology

[0002] In rail transit operations, the accumulation of moisture on the track surface reduces the coefficient of friction between train wheels and the rail, leading to problems such as slippage, wheel spin, and increased braking distance. This can even result in serious accidents that threaten operational safety, such as train collisions or impacts. Simultaneously, moisture accelerates the corrosion of rails and fasteners, shortening equipment lifespan and increasing maintenance costs. Existing dehumidification technologies often employ methods such as mechanical scraping, hot air drying, or chemical coatings, but these generally suffer from limitations including low efficiency, high energy consumption, poor adaptability, and inconvenient maintenance.

[0003] Chinese invention patent application CN113832901A, published on December 24, 2021, provides a dehumidification and cleaning device for the rail surface of rail transit vehicles. The device generates hot air through a fan and a heater, which is blown directly onto the rail surface to evaporate moisture. However, this device has the following disadvantages: (1) high energy consumption and high continuous operating costs; (2) uneven drying, which can easily lead to local moisture residue; (3) strong dependence on ambient temperature, with a significant decrease in effectiveness in low-temperature environments; (4) large equipment size, making installation and maintenance inconvenient; (5) inability to effectively solve the effects of turbulence; and (6) it is a passive response dehumidification method, which cannot preventive intervention before the risk of slippery conditions occurs.

[0004] In recent years, the development of IoT, data fusion, and artificial intelligence technologies has provided new pathways for intelligent operation and maintenance of rail transit. Predictive maintenance based on multi-source information fusion and condition recognition based on machine vision have become research hotspots in the industry. Therefore, developing a high-efficiency, energy-saving, intelligent, and predictive rail surface dehumidification system has both urgent practical needs and significant technological value. Summary of the Invention

[0005] This invention aims to address the problems of high energy consumption, uneven drying, poor adaptability, and lack of predictive capabilities in existing rail surface dehumidification technologies. The purpose of this invention is to provide a rail surface dehumidification method and system for rail transit based on multimodal data fusion. This system can comprehensively integrate multi-dimensional information to accurately perceive and predict the rail surface condition, achieving efficient, energy-saving, adaptive, and preventative rail surface moisture management, thereby improving train safety and equipment durability.

[0006] Specifically, the first aspect of this invention provides a method for dehumidifying the rail surface of rail transit based on multimodal data fusion, comprising:

[0007] Collect data related to the condition of the rail surface, including:

[0008] Sensor data, including humidity sensor data representing the humidity of the rail surface and near-rail space, temperature sensor data representing the ambient and rail surface temperature, and thickness detection sensor data representing the thickness of the water film or thin ice on the rail surface.

[0009] Visual inspection data includes the rail surface status and the distribution range of ice and water obtained by running computer vision algorithms on real-time rail surface images; the rail surface status includes dry rail status, wet rail status, and icy rail status.

[0010] Meteorological data, including real-time data from meteorological stations or weather forecast services along the route;

[0011] Line topology data, including line information for the current train operating section;

[0012] The collected data related to the track surface condition is input into a pre-trained multimodal data fusion and intelligent decision-making model for prediction, and the optimal dehumidification strategy is output.

[0013] The optimal dehumidification strategy is input into the dehumidification execution module as an instruction, and the dehumidification execution module executes high-pressure drying airflow dehumidification according to the instruction.

[0014] This solution organically combines four key elements: multi-source sensing, data fusion, intelligent decision-making, and adaptive execution.

[0015] Multimodal sensing (humidity, temperature, and thickness sensors + computer vision) provides a high-precision description of the track surface condition;

[0016] The data fusion model maps these heterogeneous information into a unified state vector and uses weather forecasts and line topology for short-term predictions.

[0017] Based on this, the intelligent decision-making module outputs execution instructions with minimum energy consumption and fastest dehumidification speed, achieving the dual goals of "high efficiency and energy saving";

[0018] Adjustable execution hardware (variable aperture nozzle array, auxiliary heating, ultrasonic atomization, and moisture-absorbing materials) can be flexibly combined according to different modes, ensuring dehumidification capability even at extreme low temperatures while avoiding energy waste caused by high-power operation throughout the process;

[0019] The predictive-prevention working mode enables the system to proactively intervene before risks occur, truly achieving the goal of "intelligent and predictable" operation and maintenance.

[0020] Based on the above, the multimodal data fusion and intelligent decision-making model includes:

[0021] Data fusion layer: Using Kalman filtering, Bayesian network or deep learning fusion model, sensor data, visual inspection data, meteorological data and track topology data are spatiotemporally aligned and fused to generate an estimate of the track surface status of the current and the section ahead.

[0022] Forecasting and Decision-Making Layer: Based on the fused current status and weather forecasts and track topology, time series analysis or forecasting models are used to determine the trend and risk of track surface condition deterioration in the future.

[0023] The core control algorithm of the prediction and decision layer adopts fuzzy PID control or neural network control algorithm, establishes the rail surface humidity-wheel-rail adhesion coefficient model, and predicts the optimal dehumidification parameters. The algorithm receives the fusion and prediction results, dynamically calculates and outputs the optimal dehumidification strategy, which includes the start and stop, intensity, range and working mode of the dehumidification execution module.

[0024] The dehumidification module has three operating modes: continuous high-intensity operation mode, continuous low-intensity operation mode, and intermittent operation mode.

[0025] By further defining the multimodal data fusion and intelligent decision-making model to include a "data fusion layer" and a "prediction and decision-making layer," and clarifying that the core control algorithm adopts fuzzy PID or neural network control and establishing a rail surface humidity-adhesion coefficient model, the problems of difficulty in uniformly representing heterogeneous data, inability to predict the trend of rail surface state changes, and lack of dynamic adaptability of control strategies have been solved. The technical effects of synergistic state estimation and trend prediction, adaptive adjustment of dehumidification parameters according to operating conditions, and support for switching between multiple working modes as needed have been achieved, thereby improving the accuracy of control and the adaptability of scenarios.

[0026] Based on the above, the algorithms for the data fusion layer include:

[0027] Noise suppression is achieved by applying a Kalman filter algorithm to the sensor data; the state equation and observation equation of the Kalman filter algorithm are as follows;

[0028] State equation: X k =AX k-1 +Bu k-1 +w k-1

[0029] Observation equation: Z k =HX k +v k

[0030] Among them, X k =[h k ,t k [] represents the state vector at time k, h k For the rail surface humidity, t kA is the rail surface temperature; B is the state transition matrix; C is the control matrix; D is the control matrix. k-1 To control the quantity; w k-1 For process noise; Z k v is the observation value at time k; H is the observation matrix; k To observe noise;

[0031] Based on the sensor data acquisition time, linear interpolation is used to align the visual inspection data, and nearest neighbor interpolation is used to match the meteorological data. Spatially, the track surface area in the visual inspection data is spatially associated with the sensor installation location by combining the mileage markers in the line topology data, so that data from different sources correspond to the same track surface section.

[0032] Construct a multimodal fusion model based on an attention mechanism to integrate Kalman-filtered sensor data (h f ,t f The feature vector F extracted by visual detection includes atmospheric humidity h. m and precipitation probability p m External meteorological data, including track gradient s and curve radius r, are used as model inputs. The weights of each modal data are dynamically allocated through an attention module, and the fused track surface state feature vector F is output. fusion ;

[0033]

[0034] Where M is the number of modes, F i Let α be the eigenvector of the i-th mode; i Let be the attention weights for the i-th modality, satisfying It is calculated by the attention module using the sigmoid function:

[0035]

[0036] Among them, W i b i These are the trainable parameters for the attention module, determined through offline training.

[0037] Based on the fused track surface state feature vector F fusion The comprehensive state estimate H of the track surface is output through the fully connected layer. est H est Normalized to the track surface humidity normalized value [0,1], where 0 is completely dry and 1 is completely wet.

[0038] By refining the data fusion layer, limiting the use of Kalman filtering to suppress sensor noise, resolving the inconsistency problem of multi-source data through spatiotemporal alignment, and constructing a multimodal fusion model based on an attention mechanism to dynamically allocate the weights of each modality, the problems of sensor data noise interference, spatiotemporal scale mismatch between visual and meteorological data, and difficulty in quantifying the contribution of each modality's information have been solved. This has resulted in improved state estimation accuracy, fusion results that are sensitive to key modes, and output track surface humidity normalization values ​​that can uniformly represent the degree of slippage.

[0039] Based on the above, the algorithms for prediction and decision-making layers include:

[0040] Rail surface condition trend prediction:

[0041] An improved ARIMA model is used to predict the trend of track surface conditions over a future period. The traditional ARIMA model is optimized to take into account the influence of meteorological data. The expression of the improved ARIMA model is as follows:

[0042]

[0043] Among them, H est (t) represents the estimated state of the track surface at time t, ▽ d Let d be the difference operator, c be the constant term, p be the autoregressive order, q be the moving average order, and Φ be the difference operator. i For autoregressive coefficients, θ j ε is the moving average coefficient, ε(t-j) is the random error term at time j, γ is the meteorological data influence coefficient, and h is the moving average coefficient. m (t) represents the atmospheric humidity at time t;

[0044] The improved ARIMA model outputs the predicted orbital state H at time t+Δt. pred (t+Δt), combined with the preset early warning threshold H th1 Emergency threshold H th2 Determine the risk level of deterioration in the track surface condition:

[0045] When H pred (t+Δt)<H th1 At that time, the risk level was low;

[0046] When H th1 ≤H pred (t+Δt)<H th2 At that time, the risk level was medium.

[0047] When H pred (t+Δt)≥H th2 At that time, the risk level was high;

[0048] A rail surface humidity-wheel-rail adhesion coefficient model is established; the relationship between the two is described by an exponential function model.

[0049] μ(h) = μ0·exp(-k·h) + μ min

[0050] Where μ(h) is the wheel-rail adhesion coefficient, μ0 is the adhesion coefficient in the dry state, k is the attenuation coefficient, and h is the coefficient of adhesion with H. est Consistent track surface humidity normalized value, μ min The limiting adhesion coefficient;

[0051] Set the target adhesion coefficient μ target :

[0052]

[0053] The target rail surface humidity h is obtained by inverse calculation using the rail surface humidity-wheel-rail adhesion coefficient model. target .

[0054] By refining the prediction and decision-making layers, the improved ARIMA model (introducing meteorological data correction terms) is used to predict the trend of track surface condition. A track surface humidity-wheel-rail adhesion coefficient index model is established to infer the target humidity. This solves the problems of traditional time series models not making full use of prior meteorological information and lacking a quantitative mapping relationship between track surface humidity and train safety. It achieves the technical effects of quantifying and classifying the risk of short-term wet skidding, directly linking dehumidification targets with safety indicators, and providing clear setpoints for control decisions.

[0055] Based on the above, the method for calculating the optimal dehumidification parameters using the fuzzy PID control algorithm is as follows:

[0056] The fuzzy PID control algorithm is used as the core control algorithm, with the current rail surface humidity H as the reference. est relative humidity h of the target rail surface target The deviation e and the rate of change of deviation ec are used as inputs to dynamically adjust the proportional coefficient K of the PID controller. p Integral coefficient K i Differential coefficient K d The optimal dehumidification parameters were calculated and obtained:

[0057] Input fuzzification:

[0058] The deviation e, with a value range of [-0.5, 0.5], is divided into 5 fuzzy subsets: negative large NB, negative small NS, zero ZO, positive small PS, and positive large PB;

[0059] The deviation change rate ec, with a value range of [-0.1, 0.1], is divided into 5 fuzzy subsets: negative large nb, negative small ns, zero zo, positive small ps, and positive large pb;

[0060] The fuzzy universes are all set to [-2,-1,0,1,2], and fuzzification is performed using triangular membership functions.

[0061] The following fuzzy control rules are established:

[0062] When e is PB and ec is pb, set K. p For large, K i For small, K d For the middle;

[0063] When e is ZO and ec is zo, K is set. p For middle, K i For middle, K d For the middle;

[0064] When e is NB and ec is nb, set K. p For small, K i For large, K d Small;

[0065] The centroid method is used to clarify the fuzzy inference results, resulting in K. p K i K d The correction amount ΔK p ΔK i ΔK d Combined with the initial proportionality coefficient K p0 Initial integration coefficients K i0 Initial differential coefficients K d0 To obtain the real-time PID parameters:

[0066] K p =K p0 +ΔK p

[0067] K i =K i0 +ΔK i

[0068] K d =K d0 +ΔK d

[0069] The initial parameters are preset according to the line environment, K p0 ∈[5-8], K i0 ∈[0.1-0.3], K d0 ∈[1-2];

[0070] The output U of the PID controller, normalized to [0,1], is mapped to specific dehumidification parameters:

[0071] Start / Stop Status: Dehumidification starts when U > 0.1; dehumidification stops when U ≤ 0.1.

[0072] Dehumidification intensity: U\in(0.1,0.4] corresponds to low intensity, U\in(0.4,0.7] corresponds to medium intensity, and U\in(0.7,1] corresponds to high intensity;

[0073] Dehumidification range: Based on the location of water accumulation areas detected by visual detection, output zoned dehumidification commands for the left rail, right rail, and full rail;

[0074] Operating modes: Low-risk levels use intermittent low-pressure dehumidification mode, medium-risk levels use continuous low-intensity conventional dehumidification mode, and high-risk levels use continuous high-intensity emergency dehumidification mode.

[0075] By employing a fuzzy PID control algorithm, taking humidity deviation and deviation change rate as input, and clarifying and dynamically adjusting PID coefficients through fuzzification, rule tables, and the center of gravity method, the controller output is mapped to specific dehumidification parameters such as start / stop, intensity, range, and working mode. This solves the problems of nonlinearity, time-varying nature, and difficulty in effectively controlling the rail surface dehumidification process with fixed PID parameters. It achieves the technical effects of real-time self-tuning of control parameters, dehumidification commands that can be refined to zones and modes, and a balance between response speed and steady-state accuracy.

[0076] Based on the above, the computer vision algorithm for processing real-time track surface images is as follows:

[0077] First, the acquired track surface image is converted to grayscale, and an adaptive threshold segmentation algorithm is used to remove background noise while preserving the track surface area. Then, morphological opening operations are used to eliminate small interference points in the image and restore the track surface contour.

[0078] The adaptive threshold segmentation algorithm is: T(x,y)=μ(x,y)+k*σ(x,y);

[0079] Where T(x,y) is the adaptive threshold of pixel (x,y), μ(x,y) is the gray mean of the neighborhood of the pixel, σ(x,y) is the standard deviation of the gray value of the neighborhood, and k is the adjustment coefficient, which is dynamically adjusted according to the light intensity.

[0080] An improved Canny edge detection algorithm is used to extract edge features of the track surface region. Combined with Hough line transform to locate track boundaries, the effective track surface region is segmented. Based on the track surface condition, color, texture, and shape features are extracted to construct a multi-dimensional visual feature vector F=[f1,f2,...,f...]. n], where f1 is the gray-level mean, f2 is the saturation, f3 is the entropy value of the gray-level co-occurrence matrix, f n This represents the percentage of the area affected by water accumulation.

[0081] A track surface state classification model is constructed based on the lightweight convolutional neural network MobileNet-V2. The classification and recognition results of the track surface state are inferred and output based on the constructed track surface state classification model. The classification and recognition results are output as discrete state values ​​Sv∈0,1,2,3, where 0 represents dry track surface, 1 represents slight condensation on track surface, 2 represents moderate water accumulation on track surface, and 3 represents severe water accumulation on track surface.

[0082] The specific implementation steps of the computer vision algorithm include adaptive threshold segmentation, improved Canny edge detection, multi-dimensional visual feature extraction, and a lightweight classification model based on MobileNet-V2. This solves the problem of unstable recognition of track surface images caused by illumination, noise, and background interference. It achieves technical effects such as accurate segmentation of track surface areas, real-time classification of water accumulation / icing conditions, high classification accuracy and low inference latency, and applicability to vehicle-mounted embedded environments.

[0083] A second aspect of the present invention provides a rail transit track surface dehumidification system based on multimodal data fusion, comprising:

[0084] A humidity sensor, connected to the intelligent control unit, is used to detect and upload the humidity of the track surface and the near-track space;

[0085] A temperature sensor, connected to the intelligent control unit, is used to detect and upload the temperature of the environment and the rail surface;

[0086] A thickness detection sensor, connected to the intelligent control unit, is used to detect and upload the thickness of the water film or thin ice on the rail surface.

[0087] High-definition camera used to capture real-time images of the track surface;

[0088] An image processor, connected to a high-definition camera and an intelligent control unit, has built-in computer vision algorithms to obtain and upload track surface status and the distribution range of ice water.

[0089] An external data access interface connects to external databases and intelligent control units to acquire and upload meteorological data, line topology data, and historical operational data.

[0090] The dehumidification execution module is connected to the intelligent control unit and is used to execute high-pressure drying airflow dehumidification according to instructions;

[0091] The intelligent control unit has a pre-set computer program / instruction that, when executed, implements the steps of the rail transit track surface dehumidification method based on multimodal data fusion as described above.

[0092] Based on the above, the dehumidification execution module includes a main dehumidification device and an auxiliary electric heating device;

[0093] The main dehumidification device includes an air compressor, a high-pressure airflow injection device, and an angle adjustment mechanism;

[0094] The air compressor is used to generate high-pressure dry airflow;

[0095] The high-pressure airflow injection device is equipped with a nozzle array, which is used to spray the high-pressure drying airflow generated by the air compressor onto the rail surface to be dried.

[0096] The angle adjustment mechanism, on which the high-pressure airflow injection device is disposed, is used to adjust the injection angle of the nozzle array of the high-pressure airflow injection device;

[0097] The auxiliary electric heating device is integrated near the nozzle array or in the airflow channel of the high-pressure airflow jet device, and is used to preheat the jetted high-pressure drying airflow.

[0098] This invention has outstanding substantive features and significant progress compared to the prior art, specifically:

[0099] (1) Track surface state perception and prediction model based on multimodal data fusion;

[0100] This invention's method simultaneously receives and processes information from sensor arrays, visual inspection and recognition units, external meteorological services, and a track topology database. Through a specific data fusion algorithm, it generates a unified and accurate estimate and prediction of the current and short-term track surface dryness and wetness conditions. Reasoning: Single data sources have significant limitations; complementary fusion of multi-source heterogeneous data can significantly improve the comprehensiveness, accuracy, and foresight of perception.

[0101] (2) A dual-mode intelligent dehumidification control method based on fusion prediction: "response-prevention";

[0102] Based on the output of the predictive model, dehumidification can be triggered immediately when the current track surface condition exceeds the standard, and the control logic and decision-making algorithm for dehumidification can also be activated in advance when it is predicted that the train is about to enter a high-risk wet and slippery area. Reasoning: This changes the traditional passive response mode of dehumidification equipment, moving the safety checkpoint forward, which is the core of improving proactive safety assurance capabilities.

[0103] (3) An energy-saving dehumidification actuator that combines high-pressure airflow injection with controllable auxiliary heating;

[0104] This device employs a composite actuator structure that uses high-pressure gas to generate a high-speed airflow to strip away moisture as the primary dehumidification method, and only activates an integrated electric heating device as needed in low-temperature environments to preheat the airflow to prevent icing. Reasoning: This structure achieves efficient physical dehumidification with low energy consumption, avoids the energy waste of traditional hot air drying which involves heating throughout the process, and also solves the risk of icing at low temperatures associated with pure cold air dehumidification.

[0105] (4) A redundant sensing architecture that combines visual perception with point sensing;

[0106] The system deploys visual detection and recognition units and physical contact / short-range sensor groups in parallel, forming a sensing system architecture that mutually verifies and complements the spatial distribution visualization and judgment of track surface status and the precise quantitative measurement of points. Reasoning: This architecture enhances the system's robustness under partial sensor failure or specific environmental interference, ensuring the reliability of status judgments. Attached Figure Description

[0107] Figure 1 This is the control flowchart of the present invention.

[0108] Figure 2 This is a schematic diagram of the system structure of the present invention.

[0109] Figure 3 This is a front structural diagram of the dehumidification execution module in Embodiment 2 of the present invention.

[0110] Figure 4 This is a side view of the dehumidification execution module in Embodiment 2 of the present invention. Detailed Implementation

[0111] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0112] Example 1

[0113] like Figure 1 As shown, this embodiment provides a rail transit track surface dehumidification method based on multimodal data fusion, including parallel data acquisition, multimodal data fusion and state prediction, generation and transmission of control commands, execution of dehumidification operation, feedback and adjustment. The specific technical implementation scheme is as follows:

[0114] 1. Collect various types of data related to the track surface condition, divided into three levels;

[0115] A. Local sensor data:

[0116] Humidity sensor data: humidity values ​​on the track surface and in the near-track space;

[0117] Temperature sensor data: ambient temperature and rail surface temperature, used to correct dehumidification parameters and provide early warning of icing risk;

[0118] Optical sensor data: Non-contact detection of the thickness of water film or thin ice on the rail surface using optical principles.

[0119] B. Visual inspection data:

[0120] It is equipped with a deep learning-based computer vision algorithm to analyze track surface images in real time, identify the track surface status of "dry track", "wet track" and "ice track", and assess the distribution range of ice and water.

[0121] c. External data:

[0122] Meteorological data: including real-time data from meteorological stations or weather forecast services along the route, obtaining information such as rainfall, snowfall, temperature, humidity, and wind speed;

[0123] Line topology data: This includes line information about the train's current location, such as bridges, curves, and other areas prone to condensation or slipperiness.

[0124] 2. Input the collected data related to the track surface condition into the pre-trained multimodal data fusion and intelligent decision-making model for prediction, and output the optimal dehumidification strategy.

[0125] The core algorithm of the multimodal data fusion and intelligent decision-making model is the multimodal data fusion and intelligent decision-making algorithm:

[0126] Data fusion layer: Using Kalman filtering, Bayesian networks or deep learning fusion models, spatiotemporal alignment and fusion processing are performed on data from sensors, visual inspection results, external meteorological data and track data to generate accurate and unified estimates of the track surface status of the current and the section ahead.

[0127] Prediction and Decision Layer: Based on the fused current status, weather forecasts, and track topology, time series analysis or prediction models are used to determine the trend and risk of track surface condition deterioration over a future period. The core control algorithm can employ fuzzy PID control or neural network control algorithms to establish a track surface humidity-adhesion coefficient model and predict optimal dehumidification parameters. This algorithm receives the fused and predicted results, dynamically calculates, and outputs the optimal dehumidification strategy, including start / stop, intensity, range, and operating mode.

[0128] The dehumidification execution module has several operating modes, including continuous high-intensity operation mode, continuous low-intensity operation mode, and intermittent operation mode, to meet the needs of emergency dehumidification, regular dehumidification, and low-pressure dehumidification.

[0129] The modes that output the optimal dehumidification strategy include:

[0130] Response mode: When the fusion determines that the current track surface humidity or water film thickness exceeds the set safety threshold, the optimal dehumidification strategy is immediately triggered.

[0131] Predictive mode: When the system predicts that the train is about to enter a high-risk area of ​​slippery conditions, such as a section that is currently experiencing rainfall or a location with a large slope, the dehumidification module can be activated in advance to achieve preventative drying.

[0132] 3. Perform dehumidification operation:

[0133] The optimal dehumidification strategy is input into the dehumidification execution module as an instruction, and the dehumidification execution module executes high-pressure drying airflow dehumidification according to the instruction.

[0134] 4. Visual inspection algorithm:

[0135] a. Image preprocessing:

[0136] First, the acquired track surface image is converted to grayscale. Then, an adaptive threshold segmentation algorithm (Formula 1) is used to remove background noise and preserve the track surface area. Next, morphological opening operation (erosion followed by dilation) is used to eliminate small interference points in the image and repair the track surface contour.

[0137] T(x,y)=μ(x,y)+k*σ(x,y) (1)

[0138] Where T(x,y) is the adaptive threshold of pixel (x,y), μ(x,y) is the gray mean of the neighborhood (3×3 window) of the pixel, σ(x,y) is the standard deviation of the gray value of the neighborhood, and k is the adjustment coefficient (range 0.8-1.2), which is dynamically adjusted according to the light intensity.

[0139] b. Feature extraction:

[0140] An improved Canny edge detection algorithm is used to extract edge features of the track surface area. Combined with Hough linear transform, the track boundary is located, and the effective track surface area is segmented. For conditions such as water accumulation and condensation on the track, color features (mean grayscale value, saturation), texture features (entropy of the grayscale co-occurrence matrix), and shape features (aspect ratio of the bounding rectangle of the water accumulation area) are extracted to construct a multi-dimensional visual feature vector F=[f1,f2,...,f n ], where f1 is the gray-level mean, f2 is the saturation, f3 is the entropy value of the gray-level co-occurrence matrix, f n This represents the percentage of the area affected by water accumulation.

[0141] c. State classification:

[0142] A track surface condition classification model is built based on a lightweight convolutional neural network (MobileNet-V2). After offline training, the model can be stored in the processor's Flash memory. Hardware-accelerated inference enables the classification and recognition of track surface conditions (dry, slight condensation, moderate water accumulation, and severe water accumulation), achieving a classification accuracy of ≥95% and an inference time of ≤20ms / frame, meeting real-time detection requirements. The classification and recognition results are output as discrete state values ​​Sv∈0,1,2,3 (0 represents dry track surface, 1 represents slight condensation, 2 represents moderate water accumulation, and 3 represents severe water accumulation).

[0143] 5. Data Fusion Layer Algorithm

[0144] a. Local filtering denoising (Kalman filtering): To suppress random noise in sensor data (especially humidity and temperature data), the Kalman filtering algorithm is used. Its state equation and observation equation are as follows:

[0145] State equation: X k =AX k-1 +Bu k-1 +w k-1

[0146] Observation equation: Z k =HX k +v k

[0147] Among them, X k =[h k ,t k ] is the state vector at time k (h k For the rail surface humidity, t k A is the rail surface temperature); A is the state transition matrix (taking the identity matrix I, because the temperature and humidity change gradually in a short time); B is the control matrix (taking the 0 matrix, no active control input); u k-1 To control the quantity; w k-1 Z represents the process noise (following a Gaussian distribution N(0,Q), where Q is the process noise covariance matrix with values ​​of text{diag}([0.01,0.005])); k v is the observation value at time k (data acquired by the sensor); H is the observation matrix (taken as the identity matrix I); k The observation noise (following a Gaussian distribution N(0,R), where R is the observation noise covariance matrix, set according to the sensor accuracy) is text{diag}([0.02,0.01])).

[0148] b. Spatiotemporal alignment processing: Due to the different acquisition frequencies of multimodal data (10Hz for physical sensors, 40Hz for visual inspection, and 0.011Hz for meteorological data), the time of acquisition of physical sensor data is used as the reference. Linear interpolation is used for time alignment of visual inspection data, and nearest neighbor interpolation is used for time matching of meteorological data. Spatially, the track surface area of ​​visual inspection is spatially associated with the installation location of physical sensors by combining the mileage markers in the track topology data, so as to ensure that data from different sources correspond to the same track surface section.

[0149] c. Global Feature Fusion (Deep Learning Fusion Model): Construct a multimodal fusion model based on an attention mechanism to integrate the Kalman-filtered physical sensor data (h... f ,t f ), the feature vector F extracted by visual detection, and external meteorological data (h m Atmospheric humidity, p m The model inputs are the probability of precipitation and the track topology features (s: track gradient, r: curve radius). The weights of each modality are dynamically allocated through an attention module, and the fused track surface state feature vector F is output. fusion ;

[0150]

[0151] Where M is the number of modes (M=4 in this embodiment), F i Let α be the eigenvector of the i-th mode. i The attention weights for the i-th modality (satisfying) The result is calculated by the attention module using the sigmoid function:

[0152]

[0153] Among them, W i b i The trainable parameters for the attention module are determined through offline training. Finally, the fused orbital state feature vector F... fusion The comprehensive state estimate H of the track surface is output through the fully connected layer. est (Normalized to [0,1], corresponding to the normalized value of track surface humidity, 0 is completely dry, 1 is completely wet).

[0154] 6. Prediction and Decision-Making Algorithms

[0155] a. Track surface condition trend prediction (time series analysis model)

[0156] An improved ARIMA model (Autoregressive Integral Moving Average model) is used to predict the trend of orbital surface status over a future period (0-60 minutes). The traditional ARIMA model is optimized to account for the influence of meteorological data. The expression of the improved ARIMA model is as follows:

[0157]

[0158] Among them, H est (t) represents the estimated state of the track surface at time t, ▽ d The operator is a d-order difference operator (d=1, to eliminate data nonstationarity), c is a constant term, p is the autoregression order (value 5), q is the moving average order (value 3), and Φ is the difference operator. i For autoregressive coefficients, θ j ε is the moving average coefficient, ε(t-j) is the random error term at time j, γ is the meteorological data influence coefficient, and h is the moving average coefficient. m (t) represents the atmospheric humidity at time t.

[0159] The model outputs the predicted orbital state H at future time t+Δt (Δt=5,10,…,60 minutes). pred (t+Δt), combined with the preset threshold H th1 (Warning threshold, 0.6), H th2 (Emergency threshold, 0.8) Determine the risk level of deterioration in the track surface condition:

[0160] When H pred (t+Δt)<H th1 At that time, the risk level was low;

[0161] When H th1 ≤H pred (t+Δt)<H th2 At that time, the risk level was medium.

[0162] When H pred (t+Δt)≥H th2 At that time, the risk level was high.

[0163] b. Rail surface humidity-wheel-rail adhesion coefficient model

[0164] A nonlinear relationship model between rail surface humidity and wheel-rail adhesion coefficient was established, namely the rail surface humidity-wheel-rail adhesion coefficient model, to provide a theoretical basis for optimizing dehumidification parameters. Based on experimental data fitting, an exponential function model was used to describe the relationship between the two:

[0165] μ(h) = μ0·exp(-k·h) + μ min

[0166] Where μ(h) is the wheel-rail adhesion coefficient, μ0 is the adhesion coefficient in the dry state (h=0) (value ranges from 0.35 to 0.45, adjusted according to the rail surface material), k is the attenuation coefficient (value ranges from 2.5 to 3.5), and h is the normalized value of rail surface humidity (compared to H). est Consistent), μ min It is the limiting adhesion coefficient (with a value of 0.08-0.12, corresponding to the state of complete water accumulation).

[0167] Target adhesion coefficient μ target The value is set to 0.25-0.30 (balancing braking performance and energy consumption), and the target rail surface humidity h is calculated by using a rail surface humidity-wheel-rail adhesion coefficient model. target :

[0168]

[0169] The core objective of dehumidification parameter optimization is to control the rail surface humidity at h target Nearby, ensure that the wheel-rail adhesion coefficient meets safety requirements.

[0170] c. Fuzzy PID control algorithm (optimal dehumidification parameter calculation)

[0171] The fuzzy PID control algorithm is used as the core control algorithm, with the current rail surface humidity H as the reference. est relative humidity h of the target rail surface target The deviation e and the rate of change of deviation ec are used as inputs to dynamically adjust the proportional coefficient K of the PID controller. p Integral coefficient K i Differential coefficient K d Then, the optimal dehumidification parameters (start / stop status, dehumidification intensity, dehumidification range, and working mode) are calculated.

[0172] Input fuzzification:

[0173] The deviation e (range [-0.5, 0.5]) is divided into 5 fuzzy subsets: negative large NB, negative small NS, zero ZO, positive small PS, and positive large PB;

[0174] The deviation change rate ec (range [-0.1, 0.1]) is divided into 5 fuzzy subsets: negative large nb, negative small ns, zero zo, positive small ps, and positive large pb;

[0175] The fuzzy universes are all set to [-2,-1,0,1,2], and fuzzification is performed using triangular membership functions.

[0176] Fuzzy rule setting:

[0177] Based on engineering experience, 25 fuzzy control rules were formulated, with the core rules as follows:

[0178] When e is PB and ec is pb, it indicates that the rail surface humidity is much higher than the target value, and the dehumidification intensity needs to be increased. K should be set accordingly. p For large, K i For small, K d For the middle;

[0179] When e is ZO and ec is zo, it indicates that the rail surface humidity is close to the target value and needs to be stabilized. K is then set. p For middle, K i For middle, K d For the middle;

[0180] When e is NB and ec is nb, it indicates that the rail surface humidity is too low and dehumidification needs to be stopped. K should be set. p For small, K i For large, K d It is small.

[0181] For the specific 25 control rules, please refer to Table 1, the Fuzzy PID Control Rule Table;

[0182] Table 1 Fuzzy PID Control Rule Table

[0183] 1 NB nb Small big Small 2 NB ns Small big Small 3 NB zo Small middle Small 4 NB ps Small middle middle 5 NB pb middle middle middle 6 NS nb Small big Small 7 NS ns middle big Small 8 NS zo middle middle middle 9 NS ps middle middle middle 10 NS pb middle Small middle 11 ZO nb middle middle middle 12 ZO ns middle middle middle 13 ZO zo middle middle middle 14 ZO ps middle middle middle 15 ZO pb middle Small middle 16 PS nb middle middle middle 17 PS ns middle middle middle 18 PS zo middle middle middle 19 PS ps big Small middle 20 PS pb big Small big 21 PB nb middle middle middle 22 PB ns big Small middle 23 PB zo big Small middle 24 PB ps big Small big 25 PB pb big Small big

[0184] The fuzzy subsets of deviation e are: NB (negative large), NS (negative small), ZO (zero), PS (positive small), and PB (positive large).

[0185] The fuzzy subsets of the rate of change of deviation (ec) are: nb (negative large), ns (negative small), zo (zero), ps (positive small), and pb (positive large).

[0186] Output quantity: Scale factor correction ΔK p Integral coefficient correction ΔK i Differential coefficient correction ΔK d ;

[0187] Table format: K p Adjust / K i Adjust / K d Adjust (the adjustment amount can be categorized as large, medium, or small).

[0188] Sharpening process:

[0189] The centroid method is used to clarify the fuzzy inference results, resulting in K. p K i K d The correction amount ΔK p ΔK i ΔK d Combined with the initial proportionality coefficient K p0 Initial integration coefficients K i0 Initial differential coefficients Kd0 To obtain the real-time PID parameters:

[0190] K p =K p0 +ΔK p

[0191] K i =K i0 +ΔK i

[0192] K d =K d0 +ΔK d

[0193] The initial parameters are preset according to the line environment, K p0 ∈[5-8], K i0 ∈[0.1-0.3], K d0 ∈[1-2].

[0194] Dehumidification parameter output:

[0195] Based on the output U of the PID controller (normalized to [0,1]), it is mapped to specific dehumidification parameters:

[0196] Start / Stop Status: Dehumidification starts when U > 0.1; dehumidification stops when U ≤ 0.1.

[0197] Dehumidification intensity: U\in(0.1,0.4] corresponds to low intensity, U\in(0.4,0.7] corresponds to medium intensity, and U\in(0.7,1] corresponds to high intensity;

[0198] Dehumidification range: Based on the location of water accumulation areas detected by visual inspection, output zoned dehumidification commands (left rail, right rail, full rail).

[0199] Work mode: Low-risk levels adopt an intermittent work mode (work for 5 minutes, then pause for 2 minutes), medium-risk levels adopt a continuous low-intensity mode, and high-risk levels adopt a continuous high-intensity mode.

[0200] Example 2

[0201] This embodiment provides a rail transit track surface dehumidification system based on multimodal data fusion, such as... Figure 2 As shown, it includes:

[0202] Humidity sensors, deployed beside the track or under the train, are connected to the intelligent control unit to detect and upload the humidity of the track surface and the near-track space; resistive or microwave humidity sensors can be used.

[0203] Temperature sensors, deployed beside the track or under the train, are connected to the intelligent control unit to detect and upload the temperature of the environment and the track surface;

[0204] Thickness detection sensors, deployed beside the track or under the train, are connected to the intelligent control unit to detect and transmit the thickness of water film or thin ice on the track surface; non-contact detection sensors based on optical principles can be used.

[0205] High-definition cameras are used to capture real-time images of the track surface; they are installed at key locations on the front of the train or beside the track and can be equipped with supplementary lighting or infrared thermal imagers to adapt to nighttime environments; visible light cameras can also be used.

[0206] An image processor, connected to a high-definition camera and an intelligent control unit, has built-in computer vision algorithms to obtain and upload track surface status and the distribution range of ice water.

[0207] An external data access interface connects to external databases and intelligent control units to acquire and upload meteorological data and line topology data.

[0208] The dehumidification execution module is connected to the intelligent control unit and is used to execute high-pressure drying airflow dehumidification according to instructions;

[0209] The intelligent control unit has a pre-set computer program / instruction that, when executed, implements the steps of the rail transit track surface dehumidification method based on multimodal data fusion as described in Example 1.

[0210] In this embodiment, all data can be transmitted to the intelligent control unit via CAN bus, Ethernet, or wireless communication. The intelligent control unit uses a high-performance embedded microcontroller, with the STM32N6 series as its core.

[0211] Specifically, such as Figure 3 and Figure 4 As shown, the dehumidification execution module includes:

[0212] The dehumidification module includes a main dehumidification unit and an auxiliary electric heating unit;

[0213] The main dehumidification device includes an air compressor, a high-pressure airflow injection device 1, and an angle adjustment mechanism 2;

[0214] The air compressor is used to generate high-pressure dry airflow;

[0215] The high-pressure airflow injection device is equipped with a nozzle array 11, which is used to spray the high-pressure drying airflow generated by the air compressor onto the rail surface to be dried.

[0216] The angle adjustment mechanism 2, on which the high-pressure airflow injection device 1 is disposed, is used to adjust the injection angle of the nozzle array 11 of the high-pressure airflow injection device;

[0217] The auxiliary electric heating device is integrated near the nozzle array or in the airflow channel of the high-pressure airflow jet device, and is used to preheat the jetted high-pressure drying airflow.

[0218] Explanation of the dehumidification execution module:

[0219] a. Main dehumidification unit: Primarily employs high-pressure airflow jet dehumidification. A high-pressure dry airflow is generated by a high-efficiency air compressor, and the jet pressure is flexibly adjusted via the nozzle array 11 on the high-pressure airflow jet device 1, allowing for directional and concentrated spraying onto the rail surface. Utilizing the Venturi effect and the shear force of the high-speed airflow, the surface tension of the water film is rapidly broken, causing it to peel off and disperse, thus achieving physical dehumidification.

[0220] b. Auxiliary electric heating device: integrated near the nozzle array or in the airflow channel, it is activated only in low-temperature environments according to the control unit command to moderately preheat the jet airflow and prevent the stripped moisture from re-icing on the rail surface.

[0221] c. Angle adjustment mechanism: The spray angle and pitch of the nozzle array can be adjusted by a servo motor or electric push rod to ensure full coverage of different positions.

[0222] d. Nozzle array airflow adjustment: The nozzle array can be controlled by multiple sets of nozzles with different diameters and independent on / off control. By combining the number of nozzles opened, the total jet area can be changed, and the pressure can be flexibly adjusted.

[0223] In some exemplary embodiments, the humidity sensor is a capacitive humidity sensor; the high-definition camera uses lidar or millimeter-wave radar for detecting track surface profiles and abnormal attachments under extreme weather conditions.

[0224] In some exemplary embodiments, the dehumidification execution module employs ultrasonic dehumidification technology as its core. A high-power ultrasonic transducer array is installed on or beside the rail surface. When excessive humidity is detected, the ultrasonic generator is activated to produce high-frequency vibrations, causing water molecules on the rail surface to atomize and accelerate evaporation / dispersion.

[0225] In some exemplary embodiments, the dehumidification execution module employs active moisture-absorbing material technology. A strip or plate-shaped device containing a composite high-efficiency moisture-absorbing material is laid beside the track. This device automatically covers the track surface to absorb moisture via mechanical means. Once saturated, the material is automatically recycled and regenerated offline using an integrated hot air or microwave device.

[0226] Verification test

[0227] Theoretical Explanation: Based on fluid mechanics, high-pressure, high-speed airflow can effectively overcome the surface tension and adhesion of the water film, achieving rapid peeling with energy consumption lower than the thermal energy required for phase change evaporation. Based on information theory and control theory, multimodal data fusion can reduce system uncertainty and improve state estimation accuracy; predictive control can control risks at their inception. Fuzzy PID or neural network algorithms can effectively handle nonlinear, time-varying control problems such as rail surface dehumidification that are influenced by multiple factors.

[0228] Verification of experimental data:

[0229] Test environment: Simulated on the test loop, with track surface humidity of 70%-100% and ambient temperature of -10°C to 40°C, simulating various working conditions such as light rain, condensation, and thin ice.

[0230] Test Methods: This system was compared with traditional hot air dehumidification devices and systems using only the hardware of this system but with simple threshold control. Measurement indicators included: time required to dry to the safe threshold, energy consumption per unit area, early warning lead time, and false alarm / missed alarm rate.

[0231] Test results: Under the same conventional dehumidification requirements, the average drying time of this system is significantly shorter than that of traditional hot air devices, and it is also significantly optimized compared to the simple threshold control mode; in terms of energy consumption, the energy consumption per unit area of ​​this system is significantly lower than that of traditional hot air devices, and it also has a significant advantage compared to the simple threshold control mode.

[0232] In terms of predictive performance, this system can provide an average advance warning of 3-5 minutes for simulated rainfall areas (depending on vehicle speed), making preventative dehumidification possible.

[0233] Reagents and equipment: deionized water (simulated moisture), SHT35 temperature and humidity sensor, FLIR infrared thermal imager, STM32N6 control board, AP100 air compressor, industrial-grade camera, and simulated meteorological data server.

[0234] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for dehumidifying rail surfaces in rail transit based on multimodal data fusion, characterized in that, include: Collect data related to the condition of the rail surface, including: Sensor data, including humidity sensor data representing the humidity of the rail surface and near-rail space, temperature sensor data representing the ambient and rail surface temperature, and thickness detection sensor data representing the thickness of the water film or thin ice on the rail surface. Visual inspection data includes the rail surface status and the distribution range of ice and water obtained by running computer vision algorithms on real-time rail surface images; the rail surface status includes dry rail status, wet rail status, and icy rail status. Meteorological data, including real-time data from meteorological stations or weather forecast services along the route; Line topology data, including line information for the current train operating section; The collected data related to the track surface condition is input into a pre-trained multimodal data fusion and intelligent decision-making model for prediction, and the optimal dehumidification strategy is output. The optimal dehumidification strategy is input into the dehumidification execution module as an instruction, and the dehumidification execution module executes high-pressure drying airflow dehumidification according to the instruction.

2. The rail transit track surface dehumidification method based on multimodal data fusion according to claim 1, characterized in that, The multimodal data fusion and intelligent decision-making model includes: Data fusion layer: Using Kalman filtering, Bayesian network or deep learning fusion model, sensor data, visual inspection data, meteorological data and track topology data are spatiotemporally aligned and fused to generate an estimate of the track surface status of the current and the section ahead. Forecasting and Decision-Making Layer: Based on the fused current status and weather forecasts and track topology, time series analysis or forecasting models are used to determine the trend and risk of track surface condition deterioration in the future. The core control algorithm of the prediction and decision layer adopts fuzzy PID control or neural network control algorithm, establishes the rail surface humidity-wheel-rail adhesion coefficient model, and predicts the optimal dehumidification parameters. The algorithm receives the fusion and prediction results, dynamically calculates and outputs the optimal dehumidification strategy, which includes the start and stop, intensity, range and working mode of the dehumidification execution module. The dehumidification module has three operating modes: continuous high-intensity operation mode, continuous low-intensity operation mode, and intermittent operation mode.

3. The rail transit track surface dehumidification method based on multimodal data fusion according to claim 2, characterized in that, The algorithms of the data fusion layer include: Noise suppression is achieved by applying a Kalman filter algorithm to the sensor data; the state equation and observation equation of the Kalman filter algorithm are as follows; State equation: X k =AX k-1 +Bu k-1 +w k-1 Observation equation: Z k =HX k +v k Among them, X k =[h k ,t k [] represents the state vector at time k, h k For the rail surface humidity, t k A is the rail surface temperature; B is the state transition matrix; C is the control matrix; D is the control matrix. k-1 To control the quantity; w k-1 For process noise; Z k v is the observation value at time k; H is the observation matrix; k To observe noise; Based on the sensor data acquisition time, linear interpolation is used to align the visual inspection data, and nearest neighbor interpolation is used to match the meteorological data. Spatially, the track surface area in the visual inspection data is spatially associated with the sensor installation location by combining the mileage markers in the line topology data, so that data from different sources correspond to the same track surface section. Construct a multimodal fusion model based on an attention mechanism to integrate Kalman-filtered sensor data (h f ,t f The feature vector F extracted by visual detection includes atmospheric humidity h. m and precipitation probability p m External meteorological data, including track gradient s and curve radius r, are used as model inputs. The weights of each modal data are dynamically allocated through an attention module, and the fused track surface state feature vector F is output. fusion ; Where M is the number of modes, F i Let α be the eigenvector of the i-th mode; i Let be the attention weights for the i-th modality, satisfying It is calculated by the attention module using the sigmoid function: Among them, W i b i These are the trainable parameters for the attention module, determined through offline training. Based on the fused track surface state feature vector F fusion The comprehensive state estimate H of the track surface is output through the fully connected layer. est H est Normalized to the track surface humidity normalized value [0,1], where 0 is completely dry and 1 is completely wet.

4. The rail transit track surface dehumidification method based on multimodal data fusion according to claim 2, characterized in that, The algorithms for prediction and decision-making include: Rail surface condition trend prediction: An improved ARIMA model is used to predict the trend of track surface conditions over a future period. The traditional ARIMA model is optimized to take into account the influence of meteorological data. The expression of the improved ARIMA model is as follows: Among them, H est (t) represents the estimated state of the track surface at time t, ▽ d Let d be the difference operator, c be the constant term, p be the autoregressive order, q be the moving average order, and Φ be the difference operator. i For autoregressive coefficients, θ j ε is the moving average coefficient, ε(t-j) is the random error term at time j, γ is the meteorological data influence coefficient, and h is the moving average coefficient. m (t) represents the atmospheric humidity at time t; The improved ARIMA model outputs the predicted orbital state H at time t+Δt. pred (t+Δt), combined with the preset early warning threshold H th1 Emergency threshold H th2 Determine the risk level of deterioration in the track surface condition: When H pred (t+Δt)<H th1 At that time, the risk level was low; When H th1 ≤H pred (t+Δt)<H th2 At that time, the risk level was medium; When H pred (t+Δt)≥H th2 At that time, the risk level was high; A rail surface humidity-wheel-rail adhesion coefficient model is established; the relationship between the two is described by an exponential function model. μ(h)=μ0·exp(-k·h)+μ min Where μ(h) is the wheel-rail adhesion coefficient, μ0 is the adhesion coefficient in the dry state, k is the attenuation coefficient, and h is the coefficient of adhesion with H. est Consistent track surface humidity normalized value, μ min The limiting adhesion coefficient; Set the target adhesion coefficient μ target : The target rail surface humidity h is obtained by inverse calculation using the rail surface humidity-wheel-rail adhesion coefficient model. target .

5. The rail transit track surface dehumidification method based on multimodal data fusion according to claim 4, characterized in that, The method for calculating the optimal dehumidification parameters using the fuzzy PID control algorithm is as follows: The fuzzy PID control algorithm is used as the core control algorithm, with the current rail surface humidity H as the reference. est relative humidity h of the target rail surface target The deviation e and the rate of change of deviation ec are used as inputs to dynamically adjust the proportional coefficient K of the PID controller. p Integral coefficient K i Differential coefficient K d The optimal dehumidification parameters were calculated and obtained: Input fuzzification: The deviation e, with a value range of [-0.5, 0.5], is divided into 5 fuzzy subsets: negative large NB, negative small NS, zero ZO, positive small PS, and positive large PB; The deviation change rate ec, with a value range of [-0.1, 0.1], is divided into 5 fuzzy subsets: negative large nb, negative small ns, zero zo, positive small ps, and positive large pb; The fuzzy universes are all set to [-2,-1,0,1,2], and fuzzification is performed using triangular membership functions. The following fuzzy control rules are established: When e is PB and ec is pb, set K. p For large, K i For small, K d For the middle; When e is ZO and ec is zo, K is set. p For middle, K i For middle, K d For the middle; When e is NB and ec is nb, set K. p For small, K i For large, K d Small; The centroid method is used to clarify the fuzzy inference results, resulting in K. p K i K d The correction amount ΔK p ΔK i ΔK d Combined with the initial proportionality coefficient K p0 Initial integration coefficients K i0 Initial differential coefficients K d0 To obtain the real-time PID parameters: K p =K p0 +ΔK p K i =K i0 +ΔK i K d =K d0 +ΔK d The initial parameters are preset according to the line environment, K p0 ∈[5-8], K i0 ∈[0.1-0.3], K d0 ∈[1-2]; The output U of the PID controller, normalized to [0,1], is mapped to specific dehumidification parameters: Start / Stop Status: Dehumidification starts when U > 0.1; dehumidification stops when U ≤ 0.

1. Dehumidification intensity: U\in(0.1,0.4] corresponds to low intensity, U\in(0.4,0.7] corresponds to medium intensity, and U\in(0.7,1] corresponds to high intensity; Dehumidification range: Based on the location of water accumulation areas detected by visual detection, output zoned dehumidification commands for the left rail, right rail, and full rail; Work mode: Low-risk levels adopt intermittent work mode, medium-risk levels adopt continuous low-intensity work mode, and high-risk levels adopt continuous high-intensity work mode.

6. The rail surface dehumidification method for rail transit based on multimodal data fusion according to claim 1, characterized in that, The computer vision algorithm for processing real-time track surface images is as follows: First, the acquired track surface image is converted to grayscale, and an adaptive threshold segmentation algorithm is used to remove background noise while preserving the track surface area. Then, morphological opening operations are used to eliminate small interference points in the image and restore the track surface contour. The adaptive threshold segmentation algorithm is: T(x,y)=μ(x,y)+k*σ(x,y); Where T(x,y) is the adaptive threshold of pixel (x,y), μ(x,y) is the gray mean of the neighborhood of the pixel, σ(x,y) is the standard deviation of the gray value of the neighborhood, and k is the adjustment coefficient, which is dynamically adjusted according to the light intensity. An improved Canny edge detection algorithm is used to extract edge features of the track surface region. Combined with Hough line transform, the track boundary is located, and the effective track surface region is segmented. Based on the track surface condition, color, texture, and shape features are extracted to construct a multi-dimensional visual feature vector F=[f1,f2,...,f...]. n ], where f1 is the gray-level mean, f2 is the saturation, f3 is the entropy value of the gray-level co-occurrence matrix, f n This represents the percentage of the area affected by water accumulation. A track surface state classification model is constructed based on the lightweight convolutional neural network MobileNet-V2. The classification and recognition results of the track surface state are inferred and output based on the constructed track surface state classification model. The classification and recognition results are output as discrete state values ​​Sv∈0,1,2,3, where 0 represents dry track surface, 1 represents slight condensation on track surface, 2 represents moderate water accumulation on track surface, and 3 represents severe water accumulation on track surface.

7. A rail transit track surface dehumidification system based on multimodal data fusion, characterized in that, include: A humidity sensor, connected to the intelligent control unit, is used to detect and upload the humidity of the track surface and the near-track space; A temperature sensor, connected to the intelligent control unit, is used to detect and upload the temperature of the environment and the rail surface; A thickness detection sensor, connected to the intelligent control unit, is used to detect and upload the thickness of the water film or thin ice on the rail surface. High-definition camera used to capture real-time images of the track surface; An image processor, connected to a high-definition camera and an intelligent control unit, has built-in computer vision algorithms to obtain and upload track surface status and the distribution range of ice water. An external data access interface connects to external databases and intelligent control units to acquire and upload meteorological data and line topology data. The dehumidification execution module is connected to the intelligent control unit and is used to execute high-pressure drying airflow dehumidification according to instructions; The intelligent control unit has a pre-set computer program / instruction that, when executed, implements the steps of the rail transit track surface dehumidification method based on multimodal data fusion as described in any one of claims 1-6.

8. The rail transit track surface dehumidification system based on multimodal data fusion according to claim 7, characterized in that, The dehumidification module includes a main dehumidification unit and an auxiliary electric heating unit; The main dehumidification device includes an air compressor, a high-pressure airflow injection device, and an angle adjustment mechanism; The air compressor is used to generate a high-pressure dry airflow; The high-pressure airflow injection device is equipped with a nozzle array, which is used to inject the high-pressure drying airflow generated by the air compressor onto the rail surface to be dried. The angle adjustment mechanism, on which the high-pressure airflow injection device is disposed, is used to adjust the injection angle of the nozzle array of the high-pressure airflow injection device; The auxiliary electric heating device is integrated near the nozzle array or in the airflow channel of the high-pressure airflow jet device, and is used to preheat the jetted high-pressure drying airflow.

Images