Plateau construction tunnel track image recognition method and system, and storage medium

Abstract

The invention relates to the technical field of rail traffic and image processing, in particular to a plateau construction tunnel rail image recognition method and system and a storage medium, and aims to cope with extreme environments with high temperature, high humidity and imperfect sensors and lighting facilities. The method comprises the following steps: arranging ultrasonic, short-wave infrared and hyper-spectral imaging equipment at the end part of a train for conveying materials to carry out multi-modal data acquisition, and extracting geometric and spectral characteristics through multi-scale convolution and spectral fusion; the system carries out data processing and model updating by using a spectrum-geometric feature collaborative optimization and active learning module, and ensures the adaptability and recognition precision of the model in environment change. And finally, path optimization is realized through a real-time prediction and reinforcement learning module so as to guarantee the accuracy of track identification and the safety of train operation. According to the method, the detection accuracy and the system response capability in the plateau complex environment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of rail transit and image processing technology, and in particular to a method and system for recognizing images of a track in a plateau construction tunnel, and a storage medium. Background Art

[0002] Track image recognition in plateau construction tunnels is crucial to ensure construction safety and the stability of trains transporting materials. Construction tunnels in my country's Qinghai-Tibet Plateau (above 5,000 meters above sea level) are often accompanied by extreme environments, such as high temperature and high humidity, and often lack complete sensors and lighting facilities in the early stages of construction. This makes it extremely difficult to accurately identify tracks and surrounding obstacles, which directly affects the safe operation of trains and construction progress.

[0003] Existing track obstacle detection technologies (Chinese invention patent, publication number: CN117325918A, name: Obstacle detection method and system in tunnels) mostly use trackside sensing units or on-board sensors for detection, such as visual sensors and lidar. Although these technologies can provide effective detection in a relatively stable environment, they expose many defects in the extreme environment of plateau construction tunnels:

[0004] High temperature and high humidity conditions may affect the reliability of sensors and the accuracy of data collection. Existing technologies such as video image and point cloud analysis algorithms have difficulty balancing real-time and accuracy, resulting in insufficient recognition speed and accuracy. The fusion of video images and radar point cloud data often relies on simple synchronization and fusion algorithms, which fail to fully optimize and strengthen the feature fusion of multimodal data, affecting the overall detection effect. Summary of the invention

[0005] In view of the many problems existing in the above-mentioned prior art, the present invention provides a method and system for image recognition of track in plateau construction tunnels, and a storage medium. The present invention collects multimodal data through ultrasonic imaging, short-wave infrared and hyperspectral imaging equipment, and extracts and fuses features through multi-scale convolutional networks and spectral fusion. Using adaptive modeling modules and spectral-geometric collaborative optimization, the system dynamically updates model parameters, and realizes path simulation and optimization through reinforcement learning and variational autoencoders. The present invention significantly improves the track and obstacle recognition capabilities and response speed in extreme environments.

[0006] A method for recognizing track images of a plateau construction tunnel comprises the following steps:

[0007] Ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging equipment are installed at the end of the train transporting materials to collect ultrasonic depth data, optical image data and spectral feature data in the tunnel during the operation of the train, and perform data synchronization and preprocessing to generate synchronous cleaning data;

[0008] The synchronized cleaned data is input into the multimodal data fusion module, and the geometric feature data and spectral enhancement feature data are extracted through multi-scale convolution and hyperspectral fusion, and feature fusion is performed to generate fused feature data;

[0009] The fused feature data is input into the adaptive modeling module, the fused feature data is optimized using the spectral-geometric feature collaborative optimization algorithm, and the model parameters of the adaptive modeling module are updated through adaptive training to generate optimized feature data;

[0010] The optimized feature data is input into the real-time prediction and multi-objective decision-making system, the potential space distribution is generated through the variational autoencoder, and the path simulation and path optimization are performed in combination with the reinforcement learning module to generate path prediction data and path optimization data;

[0011] The active learning module collects execution feedback of path optimization data and performs active learning and model updating on the adaptive modeling module.

[0012] Preferably, in the data synchronization and preprocessing step, the ultrasonic depth data, the optical image data and the spectral feature data are time aligned and noise removed by a time synchronization module.

[0013] Preferably, the multimodal data fusion module adopts a multi-scale convolutional network, extracts the geometric feature data layer by layer through convolution kernels with different receptive fields, and embeds spectral enhancement feature data in each convolution layer to generate fused feature data.

[0014] Preferably, the spectral-geometric feature collaborative optimization algorithm optimizes the geometric feature data and the spectral feature data by the following calculation formula:

[0015] W opt =W geo +α·W spec

[0016] Among them, the optimized weight combination is W opt ; The geometric feature weight is W geo ; The spectral feature weight is W spec ; The weight adjustment coefficient is α.

[0017] Preferably, during the adaptive training process, the adaptive modeling module introduces a meta-learning mechanism so that the adaptive modeling module automatically adjusts the learning rate and weight parameters according to the feedback data.

[0018] Preferably, the real-time prediction and multi-objective decision-making system generates a latent space distribution through a variational autoencoder, and the latent space distribution is described by the following formula:

[0019] p(Z|X)=N(μ(X),σ(X))

[0020] Among them, the latent space variable is Z; the optimized feature data is X; the mean of the feature data is μ(X); and the standard deviation of the feature data is σ(X).

[0021] Preferably, the reinforcement learning module includes a policy network and a value network, the policy network performs path simulation based on the generated latent space distribution, and the value network is used to evaluate the safety and efficiency of the path simulation to generate path prediction data and path optimization data.

[0022] Preferably, when collecting execution feedback of the path optimization data, the active learning module generates annotated feedback data through an automatic annotation algorithm and performs multi-level data cleaning to remove outliers and redundant information.

[0023] A system for implementing the plateau construction tunnel track image recognition method, the system comprising:

[0024] Ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging equipment installed at the end of the train transporting materials are used to collect ultrasonic depth data, optical image data and spectral feature data in the tunnel during the train operation;

[0025] A time synchronization module, used for time alignment and preprocessing of the collected ultrasonic depth data, optical image data and spectral feature data to generate synchronized cleaning data;

[0026] A multimodal data fusion module, including a multiscale convolution and hyperspectral fusion unit, is used to extract geometric feature data and spectral enhancement feature data from the synchronous cleaned data, and perform feature fusion to generate fused feature data;

[0027] An adaptive modeling module is configured to receive the fused feature data and perform optimization processing using a spectral-geometric feature collaborative optimization algorithm, and to update model parameters of the adaptive modeling module through adaptive training to generate optimized feature data;

[0028] A real-time prediction and multi-objective decision-making system, including a variational autoencoder and a reinforcement learning module, for generating a latent space distribution based on optimized feature data, performing path simulation and path optimization to generate path prediction data and path optimization data;

[0029] The active learning module is used to collect execution feedback of path optimization data and perform active learning and model update on the adaptive modeling module to improve the recognition accuracy and adaptability of the system in high temperature and high humidity environments.

[0030] A storage medium stores a computer program, which, when executed by a processor, implements the steps of the plateau construction tunnel track image recognition method.

[0031] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0032] The present invention realizes efficient track and obstacle identification in extreme environments of high temperature and high humidity through multimodal data fusion and adaptive modeling modules;

[0033] Through the spectral-geometric feature collaborative optimization algorithm and active learning mechanism, the system can adaptively update and improve the recognition performance of the model to ensure accurate analysis and real-time response of orbital images;

[0034] The present invention realizes path prediction and optimization by introducing a reinforcement learning module and a variational autoencoder, making the system more adaptable and robust in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention;

[0036] Figure 2 It is a relationship diagram of the multimodal data fusion module in the present invention;

[0037] Figure 3 It is a relationship diagram of the spectrum-geometry feature collaborative optimization module in the present invention;

[0038] Figure 4 A diagram showing the relationship between the reinforcement learning module and the path optimization in the present invention;

[0039] Figure 5 A relationship diagram between active learning and model updating in the present invention;

[0040] Figure 6 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0041] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0042] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0043] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0044] like Figure 1 As shown, a method for recognizing track images of plateau construction tunnels comprises the following steps:

[0045] Ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging equipment are installed at the end of the train transporting materials to collect ultrasonic depth data, optical image data and spectral feature data in the tunnel during the operation of the train, and perform data synchronization and preprocessing to generate synchronous cleaning data;

[0046] Preferably, in the data synchronization and preprocessing step, the ultrasonic depth data, the optical image data and the spectral feature data are time aligned and noise removed by a time synchronization module.

[0047] In the construction scenario of a train tunnel at an altitude of more than 5,000 meters, the tunnel is characterized by high temperature and high humidity, and the sensors and lighting facilities are not perfect in the early stage of construction. Under such conditions, when the train transporting materials runs on the track, it needs to rely on an accurate forward obstacle recognition system to ensure safety. To achieve this goal, the present invention proposes to set ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging equipment at the end of the train to collect multi-source environmental data in the tunnel, ensuring that the system can accurately identify the track and its surrounding obstacles under complex conditions.

[0048] Ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging devices are arranged at the end of the train. Each of these sensors has a specific function: ultrasonic imaging sensors are used to collect geometric depth data of the track and provide accurate structural information; short-wave infrared sensors obtain optical image data to capture the morphological details of the track and the surrounding environment; hyperspectral imaging devices collect spectral feature data to identify the optical properties of different materials and the reflective characteristics of materials in the environment. This combination can perform real-time data collection while the train is running, providing rich multimodal data for subsequent processing.

[0049] In extreme environments with high temperature and humidity, the temporal and spatial consistency of data is particularly important. To this end, the data synchronization and preprocessing steps use a time synchronization module to time align and remove noise from data from different sensors. The time synchronization module timestamps each data source and ensures that the data remains consistent within the same time frame through a timing matching algorithm. The core of this step is to apply a timing matching formula, such as:

[0050] T aligned (i) = T sensor (i)-Δt

[0051] Among them, T aligned (i) indicates the timestamp after alignment; T sensor (i) is the original timestamp of sensor data acquisition; Δt is the time difference between sensors.

[0052] The noise removal process uses a median filter to pre-process the ultrasonic depth data to reduce the interference of high-frequency noise; for optical image data, Gaussian filtering is applied to smooth the image to improve the overall image quality and reduce noise; the hyperspectral data is spectrally smoothed through a denoising algorithm to improve its accuracy in material identification.

[0053] The high temperature and high humidity conditions in the plateau tunnel will affect the stability of the sensor and the accuracy of the data. Therefore, the time synchronization module and data processing system of the present invention have an environmental compensation mechanism. In the process of noise removal, the preprocessing algorithm includes dynamic adjustment parameters to adapt to real-time changes in the environment. For example, when the temperature changes, the algorithm parameter k(T) will automatically adjust according to the ambient temperature T:

[0054]

[0055] Where k0 is the initial parameter; β is the temperature coefficient; T ref This adjustment ensures the stability of data preprocessing under different environmental conditions.

[0056] Through the above data collection and preprocessing steps, the present invention can generate synchronous cleaning data under extreme conditions in plateau tunnels, ensure the temporal and spatial consistency and quality of input data, and provide a reliable basis for subsequent multimodal data fusion and feature extraction. The method has strong adaptability and data stability in high temperature and high humidity environments, enabling the system to accurately identify tracks and obstacles near them in tunnels with complex environments, ensuring the safe operation of trains.

[0057] The synchronized cleaned data is input into the multimodal data fusion module, and the geometric feature data and spectral enhancement feature data are extracted through multi-scale convolution and hyperspectral fusion, and feature fusion is performed to generate fused feature data;

[0058] Preferably, Figure 2 As shown, the multimodal data fusion module adopts a multi-scale convolutional network to extract geometric feature data layer by layer through convolution kernels with different receptive fields, and embeds spectral enhancement feature data in each convolution layer to generate fused feature data.

[0059] In the construction tunnel environment of the plateau at an altitude of more than 5,000 meters, the train faces the challenges of high temperature and high humidity during operation, which puts extremely high demands on the forward obstacle recognition system of the material transport train, especially when the sensors and lighting facilities are not perfect in the early stage of construction. To overcome these challenges, the present invention proposes a multimodal data fusion module, which extracts the geometric feature data and spectral enhancement feature data of the tunnel environment through multi-scale convolution and hyperspectral fusion, and performs feature fusion, and finally generates high-precision fused feature data to achieve accurate identification and timely response of obstacles on the track and its surroundings.

[0060] The synchronized cleaned data generated by data synchronization and preprocessing is input into the multimodal data fusion module, which uses a multi-scale convolutional network to extract geometric feature data layer by layer to ensure that features of different scales can be captured and analyzed. The multi-scale convolutional network uses convolution kernels with different receptive fields (such as k1×k1, k2×k2) to capture geometric features from details to the global. The formula is described as follows:

[0061]

[0062] in, is the geometric feature output of the lth layer; is the convolution kernel weight of the lth layer; X (l-1) Input data for the l-1th layer; b (l) is the bias term; * indicates the convolution operation.

[0063] In each convolutional layer, the hyperspectral data is embedded into the convolutional network to form spectrally enhanced feature data. This spectral enhancement is achieved by pixel-by-pixel embedding, ensuring that the surface material properties of each geometric feature are taken into account. The formula for this step is as follows:

[0064]

[0065] in, represents the characteristics after spectral enhancement, S is the spectral characteristics of the corresponding position, Represents an embed operation.

[0066] The multimodal data fusion module extracts features layer by layer through a multi-scale convolutional network, combined with spectral enhancement features, so that the system can effectively cope with complex optical changes such as reflection and scattering in the high temperature and high humidity environment of the tunnel, thereby improving the robustness of feature extraction and detail recognition capabilities. The final result of feature fusion is the generation of fused feature data, which integrates geometric and spectral information and has good spatiotemporal consistency and environmental adaptability.

[0067] This method ensures that the train can accurately identify obstacles on the track and in the surrounding area during operation. For example, if there are water stains or muddy areas caused by humidity in the tunnel, the system can quickly identify the existence of these obstacles and their material properties through the joint analysis of geometric features and spectral enhancement features, and generate reliable obstacle identification data for real-time decision-making.

[0068] In an experiment simulating a plateau tunnel environment, the multimodal data fusion module was verified to be able to maintain high-precision feature extraction under different lighting and humidity conditions. The multi-scale convolution captured details and overall structural information, and the spectral enhancement features helped identify the material of slippery surfaces and debris, enabling the system to make accurate path selection and obstacle avoidance decisions.

[0069] The present invention demonstrates extremely strong robustness in high temperature and high humidity environments, ensures the stable operation and safety of the train, and meets the high requirements for accurate identification and timely response to obstacles under plateau tunnel construction conditions.

[0070] like Figure 3 As shown, the fused feature data is input into the adaptive modeling module, the fused feature data is optimized using the spectral-geometric feature collaborative optimization algorithm, and the model parameters of the adaptive modeling module are updated through adaptive training to generate optimized feature data;

[0071] In order to achieve efficient detection of tracks and surrounding obstacles, the present invention proposes an adaptive modeling module, which receives fused feature data and optimizes it through a spectral-geometric feature collaborative optimization algorithm. At the same time, an adaptive training method is used to update the model parameters to generate optimized feature data, thereby improving the system's responsiveness in extreme environments.

[0072] The fused feature data generated by the multimodal data fusion module is input into the adaptive modeling module. The adaptive modeling module processes the fused feature data through the spectral-geometric feature collaborative optimization algorithm, focusing on using the correlation between spectral features and geometric features to improve the expressiveness of features. Spectral features provide information on material properties and optical characteristics, while geometric features provide spatial structure and depth information. The collaborative optimization process is carried out through the following formula:

[0073] W opt (x,y)=Wgeo (x,y)+α·W spec (x,y)

[0074] Among them, W opt (x, y) represents the optimized weight combination at the pixel position (x, y); W geo (x, y) represents the geometric feature weight; W spec (x, y) represents the spectral feature weight; α is the dynamic adjustment coefficient; it is used to adaptively adjust the impact of spectral features on the optimization process according to environmental changes.

[0075] In order to overcome the impact of high temperature and high humidity on data stability, the adaptive modeling module embeds an environmental compensation factor in the collaborative optimization algorithm. This factor is adjusted through real-time environmental data obtained by temperature and humidity sensors to ensure that the algorithm can maintain optimized stability under extreme conditions. The dynamic adjustment coefficient α involved in the formula is calculated through the following relationship:

[0076]

[0077] Among them, α0 is the basic adjustment coefficient; β is the environmental adjustment factor; H is the current humidity; H ref For reference humidity.

[0078] After completing the collaborative optimization, the fused feature data is further used for adaptive training. The adaptive modeling module automatically adjusts the model weights and learning rates based on real-time feedback data through an online learning mechanism to improve the stability of the model in environmental changes. The online learning mechanism uses a meta-learning algorithm to update parameters, and its formula is:

[0079]

[0080] Among them, θ new is the updated model parameter; θ old is the current model parameter; η is the learning rate; is the gradient loss function based on the feature data X.

[0081] Through spectral-geometric feature collaborative optimization and adaptive training, the adaptive modeling module can adjust the model in real time to adapt to the dynamic changes of high temperature and high humidity environment. This design ensures that under the complex optical and material conditions in the tunnel, the system can continuously output high-precision optimized feature data, improving the robustness and stability of obstacle recognition. For example, when the spectral reflectance characteristics change due to increased humidity in the tunnel, the adaptive adjustment coefficient α can dynamically adjust the weight of the spectral features, so that the model maintains high recognition accuracy when dealing with slippery tracks or obstacles of different materials.

[0082] In actual experiments, the module showed excellent performance in a simulated plateau tunnel environment. Through online adaptive training and collaborative optimization processing, the system achieved stable recognition of tracks and obstacles, effectively reduced false alarms and missed alarms, and ensured the safety and real-time response capabilities of train operations. This sophisticated optimization strategy is particularly important during the construction phase, meeting the strict requirements for real-time image recognition and high adaptability.

[0083] Preferably, the spectral-geometric feature collaborative optimization algorithm optimizes the geometric feature data and the spectral feature data by the following calculation formula:

[0084] W opt =W geo +α·W spec

[0085] Among them, the optimized weight combination is W opt , used for subsequent feature analysis and decision making; the geometric feature weight is W geo , mainly describes the spatial layout of track structure and obstacles; the spectral feature weight is W spec , reflecting the optical and material properties of the object surface; the weight adjustment coefficient is α, which controls the influence of spectral features in the final feature combination.

[0086] In the construction scene of plateau tunnels at an altitude of more than 5,000 meters, trains transporting materials operate in complex environments, facing the challenges of high temperature and high humidity, as well as the imperfections of early sensors and lighting facilities. These conditions place extremely high demands on the obstacle recognition system. The system must not only overcome the adverse effects of the environment, but also accurately identify obstacles on the track and its surroundings during real-time operation. To achieve this goal, the spectral-geometric feature collaborative optimization algorithm used in the present invention optimizes the data by fusing geometric and spectral features to improve the adaptability and recognition accuracy of the model in complex environments.

[0087] The core of this optimization algorithm is to effectively combine geometric feature data and spectral feature data to comprehensively utilize the advantages of both to improve the performance of the recognition system. In practical applications, geometric feature data provides the shape and spatial information of tracks and obstacles, while spectral feature data provides information on material properties and surface optical characteristics. Through the collaborative optimization algorithm, these two features can be used complementary in high temperature and high humidity environments to form a more reliable feature combination.

[0088] In high temperature and high humidity environments, in order to improve the adaptability of the algorithm, the weight adjustment coefficient α can be dynamically adjusted according to environmental data. For example, when humidity and temperature affect the reliability of spectral data, the system will appropriately reduce the value of α to balance the contribution of geometric and spectral features. This adjustment can be achieved by obtaining real-time humidity HHH and temperature TTT through environmental parameter sensors:

[0089]

[0090] Among them, α0 is the basic adjustment coefficient; γ is the adjustment factor, which is used to reflect the influence of humidity on the spectral feature weight; H ref is the reference humidity value, used to normalize the humidity effect.

[0091] This optimization strategy ensures that the system can dynamically adjust the weights of feature combinations when faced with environmental changes, so as to improve the model's recognition accuracy and stability for tracks and obstacles.

[0092] After applying this collaborative optimization algorithm, the system can accurately distinguish between tracks and obstacles in complex optical environments. For example, the slippery track surface is prone to misidentification in traditional methods due to its complex reflection characteristics, but this algorithm can effectively cope with these optical changes by enhancing the recognition of materials by spectral features. In the test simulating the plateau tunnel environment, when the humidity reaches a high level, the system automatically adjusts the value of α so that the recognition model still maintains a high detection accuracy, thereby achieving accurate distinction between obstacles and tracks.

[0093] Through this optimization process, the generated optimized feature data has high robustness and adaptability, which can be used for subsequent path planning and real-time decision-making, so that the train can operate safely and efficiently under complex construction conditions. The present invention shows superior characteristics in high temperature and high humidity environments, meeting the high-precision requirements of image recognition systems under extreme construction conditions.

[0094] Preferably, during the adaptive training process, the adaptive modeling module introduces a meta-learning mechanism so that the adaptive modeling module automatically adjusts the learning rate and weight parameters according to the feedback data.

[0095] In the construction environment of plateau tunnels at an altitude of more than 5,000 meters, trains transporting materials need to operate under extreme conditions of high temperature and high humidity, facing the challenges of complex environment, imperfect sensors and lighting facilities. This places extremely high demands on the obstacle recognition system ahead of the train, requiring the system to have the ability to adapt to environmental changes in real time and maintain high-precision recognition. The present invention proposes an adaptive modeling module and introduces a meta-learning mechanism in the adaptive training process, so that the system can automatically adjust the learning rate and weight parameters according to real-time feedback data to cope with dynamic changes in the environment.

[0096] The core of the adaptive modeling module is to achieve dynamic adjustment of model parameters through a meta-learning mechanism. Meta-learning is a "learning how to learn" technique that enables the model to quickly adapt to new situations from changing input data. In high temperature and high humidity environments such as plateau tunnels, external conditions may affect the stability and quality of sensor data collection. After the adaptive modeling module introduces the meta-learning mechanism, the system can analyze feedback data in real time and automatically adjust the model's learning rate and weight parameters according to environmental changes to maintain the robustness and efficiency of the model.

[0097] In adaptive training, the model uses the gradient update method to optimize parameters and combines the meta-learning mechanism to achieve dynamic adjustment of the learning rate. The formula is as follows:

[0098]

[0099] Among them, θ new is the updated model parameter; θ old is the current model parameter; η adaptive It is an adaptive learning rate, which is dynamically adjusted according to real-time feedback data; is the gradient loss function based on the input data X.

[0100] Adaptive learning rate η adaptive The adjustment is based on performance indicators in the feedback data, such as the error change rate. The adjustment formula is:

[0101]

[0102] Among them, η0 is the basic learning rate; λ is the adjustment factor, which is used to control the variation range of the learning rate; It represents the gradient of the loss function with respect to temperature T, reflecting the impact of environmental changes on the model error.

[0103] In plateau tunnels, high temperature and high humidity can cause image blur and abnormal light reflection, which will reduce the recognition performance of the system. The adaptive modeling module can quickly respond to such dynamic environmental changes and adjust the learning rate and model weights in real time by introducing a meta-learning mechanism. When identifying tracks and surrounding obstacles, the model no longer relies on fixed parameters, but is automatically updated with environmental conditions to maintain recognition accuracy and stability.

[0104] For example, when the ambient humidity increases and causes the spectral feature data to fluctuate, the system adjusts the learning rate η by analyzing the feedback data in real time. adaptive , thereby reducing the training error and optimizing the weight parameter θ. This adjustment enables the system to quickly adapt to data differences caused by humidity changes and maintain high accuracy and reliability of track identification.

[0105] In the experiment, the adaptive modeling module was applied to the image recognition system simulating the plateau tunnel environment. The results showed that after the model dynamically adjusted the learning rate, its recognition accuracy was improved by more than 15% compared with the traditional fixed parameter model, and the false alarm and missed alarm rates of the model were significantly reduced when the humidity suddenly increased or the temperature fluctuated. This performance improvement ensures the safe operation of trains in complex environments and meets the strict requirements for efficient obstacle recognition during the construction phase.

[0106] like Figure 4 As shown, the optimized feature data is input into the real-time prediction and multi-objective decision-making system, the potential space distribution is generated by the variational autoencoder, and the path simulation and path optimization are performed in combination with the reinforcement learning module to generate path prediction data and path optimization data;

[0107] In the plateau tunnel construction environment at an altitude of more than 5,000 meters, trains transporting materials face extreme conditions of high temperature and high humidity during operation. This environment requires the track image recognition system to have extremely high accuracy and adaptability. Due to the incomplete sensors and lighting facilities in the early stage of tunnel construction, the recognition system needs to overcome the impact of the environment on data quality and system stability, so as to identify the track and surrounding obstacles in real time and respond quickly. The present invention proposes to input the optimized feature data into a real-time prediction and multi-objective decision-making system. The system generates a latent space distribution through a variational autoencoder, combines a reinforcement learning module for path simulation and path optimization, and finally generates path prediction data and path optimization data, thereby ensuring the safety and efficiency of train operation.

[0108] The optimized feature data is processed by the adaptive modeling module and then input into the real-time prediction and multi-objective decision-making system. The system first uses a variational autoencoder (VAE) to generate a latent space distribution to capture the potential structure and characteristics of the track and the surrounding environment. VAE generates a probability distribution of the latent variable Z through two neural networks - encoder and decoder. Specifically, the encoder network converts the input feature data X into latent distribution parameters μ(X) and σ(X):

[0109] Z~N(μ(X),σ(X))

[0110] Among them, Z represents the latent space variable; μ(X) is the mean of the input feature data; σ(X) is the standard deviation of the input feature data.

[0111] The latent space distribution captures the global information of track features and obstacles, allowing the system to make inferences under incomplete environmental data. This is especially important in tunnel environments with high temperature and humidity, because environmental changes may cause some data to be distorted or incomplete. The generated latent space distribution is used in the reinforcement learning module, which receives the latent space information and performs path simulation and optimization.

[0112] The reinforcement learning module consists of a policy network and a value network. The policy network is responsible for generating path simulation plans, and the value network evaluates the safety and resource efficiency of each path. The reinforcement learning algorithm uses a reward-based approach to select the optimal path after simulating multiple paths. The path optimization formula is as follows:

[0113]

[0114] Among them, Q(s,a) represents the value of taking action a in the current state s; R(s,a) is the immediate reward of the current state-action pair; γ is the discount factor used to balance the weights of current and future rewards; s′ and a′ represent the next state and action, respectively.

[0115] During the construction phase, due to the complex conditions in the tunnel, the system needs to ensure that the train can make quick decisions when unknown or sudden obstacles appear. Through the potential space distribution generated by the variational autoencoder, the system can predict track characteristics when the data is incomplete, and combine the reinforcement learning module to perform path simulation and real-time optimization, so that the system can generate path prediction data and path optimization data in time in the case of sudden track changes, obstacles or slippery surfaces.

[0116] For example, in a high humidity environment, if the track surface is reflective or slippery due to water stains, the variational autoencoder captures the potential distribution of spectral and geometric features, and the reinforcement learning module can simulate the feasibility of different paths and select paths based on safety and stability. The final output path prediction data and path optimization data can be used to guide the train's driving adjustments in real time to avoid potential safety risks.

[0117] In experiments simulating plateau tunnel environments, the real-time prediction and multi-objective decision-making system showed high adaptability. Under extreme high temperature and high humidity conditions, the system maintained a high success rate in path optimization, reducing the risk of misjudgment and path deviation caused by slippery tracks. Experimental results show that the system's response time in dynamic path planning is reduced by more than 20% compared with traditional models, while improving the accuracy of overall recognition and emergency response. This makes the present invention particularly suitable for plateau tunnel construction train operation scenarios that require efficient image recognition and real-time decision-making.

[0118] Preferably, the real-time prediction and multi-objective decision-making system generates a latent space distribution through a variational autoencoder, and the latent space distribution is described by the following formula:

[0119] p(Z|X)=N(μ(X),σ(X))

[0120] Among them, the latent space variable is Z, which represents the representation of the optimized feature data in the latent space; the optimized feature data is X; the mean of the feature data is μ(X), which reflects the central tendency of the feature data; the standard deviation of the feature data is σ(X), which represents the diffusion of the data distribution.

[0121] In the construction scene of plateau tunnels at an altitude of more than 5,000 meters, trains transporting materials face the complexity and data collection difficulties brought by high temperature and high humidity environments, which puts extremely high demands on the track image recognition system, especially under the conditions of imperfect sensors and lighting facilities in the early stage of construction. The system needs to overcome the influence of harsh environments, accurately identify tracks and surrounding obstacles, and perform image processing and path decision-making in real time. In the present invention, the real-time prediction and multi-objective decision-making system generates potential space distribution through variational autoencoders (VAEs), providing a solid data foundation for path optimization.

[0122] Variational Autoencoder (VAE) is a probabilistic generative model that compresses input data into a latent space by building an encoder-decoder structure, and then reconstructs the input data from the space. Its core is to generate a latent space distribution p(Z|X) to represent the mapping of input data X in the latent space. For the optimized feature data X collected in the plateau tunnel environment, VAE uses its statistical characteristics to generate a latent space distribution so that the system can make path predictions and decisions when the data is incomplete or the environment changes complexly.

[0123] This latent space distribution can capture the intrinsic connections and changes of track and obstacle characteristics in high temperature and high humidity environments, ensuring that the model can effectively reconstruct features and make path decisions when the environment is disturbed and sensor data is unstable.

[0124] After the optimization feature data is input into the real-time prediction and multi-objective decision-making system, the latent space distribution generated by VAE provides multiple possible predictions for path optimization. The reinforcement learning module uses this latent space distribution to simulate multiple paths and select the optimal path for safe and efficient train operation. Through the latent space variable Z, the system can perform efficient path simulation and risk assessment when the environment changes or data is missing.

[0125] For example, when the humidity in the tunnel increases and the noise of the optical data increases, the latent space distribution p(Z|X) can still capture the overall characteristics of the track and obstacles through the mean μ(X) and standard deviation σ(X). This allows the system to generate more stable path prediction and path optimization data in real-time response, reducing the risk of misjudgment or path deviation.

[0126] In a simulated plateau tunnel environment, the system inputs optimized feature data into the variational autoencoder, and the generated latent space distribution is used in the reinforcement learning module for path simulation. The results show that the system can still maintain efficient recognition and path decision-making under conditions of high humidity and drastic temperature changes, with a 15% reduction in false alarm rate and a 20% reduction in overall path planning time. This approach ensures that the system can quickly adjust and output reliable path optimization data in harsh environments, improving the safety and adaptability of train operations.

[0127] The present invention achieves accurate identification and real-time response to tracks and obstacles through the latent space distribution generated by the variational autoencoder and combined with real-time path simulation and optimization technology, meeting the complex requirements of the plateau tunnel construction stage.

[0128] Preferably, the reinforcement learning module includes a policy network and a value network, the policy network performs path simulation based on the generated latent space distribution, and the value network is used to evaluate the safety and efficiency of the path simulation to generate path prediction data and path optimization data.

[0129] In the construction environment of plateau tunnels above 5,000 meters, train operation faces extreme conditions, such as high temperature and humidity, and the preliminary environment with imperfect sensors and lighting facilities, which puts extremely high demands on the track image recognition system. The system needs to overcome these limitations in a complex environment, perform real-time and accurate identification of tracks and surrounding obstacles, and respond efficiently. The present invention achieves this goal through a reinforcement learning module, which includes a policy network and a value network, which perform path simulation and path optimization based on the potential space distribution, thereby generating path prediction data and path optimization data.

[0130] The core of the reinforcement learning module is the collaborative work of the policy network and the value network to ensure that the system is efficient and safe in real-time planning of track paths. The latent space distribution is generated by the variational autoencoder, providing a high-dimensional representation of track features. The policy network uses this distribution as input to simulate the path, and the value network evaluates the safety and efficiency of the path to optimize the decision.

[0131] The policy network is a structure based on a deep neural network that receives the latent space variable Z as input and outputs a series of path selection actions. The network is optimized by the policy gradient algorithm to maximize the expected cumulative reward. The calculation formula for path simulation is:

[0132]

[0133] Among them, π(a|Z) represents the probability of taking action a given the latent space variable Z; Q(Z,a) represents the value of taking action a in path selection; a′ represents all possible actions.

[0134] The policy network performs path simulation based on this formula and selects the optimal path sequence to improve the system's response speed to track changes and obstacles.

[0135] The value network is used to evaluate the safety and efficiency of each path simulation. It evaluates the multi-objective indicators of the path and outputs the cumulative value of the path. The value function is defined as follows:

[0136]

[0137] Among them, V(Z) represents the cumulative value of the path under the latent space variable Z; R t is the immediate reward obtained at time step t, taking into account the safety and efficiency of the path; γ is the discount factor used to balance the importance of future rewards and current rewards.

[0138] The value network incorporates environmental characteristics such as track slipperiness, obstacle distances, and real-time optical data in its assessment. By incorporating environmental feedback, the value network can accurately assess the safety and expected outcomes of a path.

[0139] Through the combination of the policy network and the value network, the reinforcement learning module can simulate multiple possible paths in the complex environment of the plateau tunnel and optimize the selection based on real-time data. For example, in a high temperature and high humidity environment, when the tunnel is slippery or the lighting is insufficient, resulting in increased noise in the optical data, the policy network can quickly generate multiple path simulations through inputs distributed in the latent space, and the value network selects the optimal path through real-time evaluation, ensuring that the train can still operate safely and efficiently when the environment deteriorates.

[0140] In a practical implementation, the reinforcement learning module was tested in a simulated environment of a plateau tunnel. The test showed that with the synergy of the policy network and the value network, the system can respond to changes in complex environments in a short time in path selection, identify and avoid obstacles. The path prediction data and path optimization data generated by the system showed in the experiment that the response time was shortened by 30%, and the stability of track identification and the accuracy of path planning were improved by more than 20%. This performance significantly improves the system's adaptability in extreme environments and the safety of train driving.

[0141] Through this reinforcement learning module design, the system not only achieves efficient path simulation and optimization, but also ensures the stability of train operation under the special environment of plateau tunnels, meeting the needs of accurate identification and response under complex construction conditions.

[0142] The active learning module collects execution feedback of path optimization data and performs active learning and model updating on the adaptive modeling module.

[0143] Preferably, Figure 5As shown, when collecting execution feedback of path optimization data, the active learning module generates labeled feedback data through an automatic labeling algorithm and performs multi-level data cleaning to remove outliers and redundant information.

[0144] In the construction environment of plateau tunnels at an altitude of more than 5,000 meters, due to challenges such as high temperature and humidity and imperfect sensors and lighting facilities in the early stage of construction, trains face great difficulties in identifying obstacles during operation. This places extremely high demands on the real-time and accuracy of the track image recognition system. In order to improve the responsiveness and adaptability of the system, the present invention introduces an active learning module to collect execution feedback of path optimization data, and updates and optimizes the model of the adaptive modeling module. Through this design, the system can continuously learn and enhance recognition and decision-making capabilities in a dynamic environment.

[0145] The core of the active learning module lies in its continuous collection and processing of path optimization data, and using this data to actively learn and update the model of the adaptive modeling module. Active learning is a machine learning strategy that selectively marks and learns new data samples to enable the model to be updated and optimized more efficiently. Specifically, in the complex operating environment of the plateau tunnel, the module monitors the path optimization data output by the system, collects its execution feedback, and adjusts the model based on the feedback data.

[0146] The active learning module extracts execution feedback data from the output of the real-time path simulation and optimization process, which includes the actual path of the train on the track, the identified obstacle information, and the performance of the path decision. The collected execution feedback data is processed by the automatic labeling algorithm to generate labeled feedback data to assist model training. The labeling algorithm automatically labels the data for subsequent supervised learning of the model. The labeling process uses the following algorithms:

[0147] y i =f(x i )+∈ i

[0148] Among them, y i is the annotation feedback data; x i is the input path data; ∈ i is the noise term, reflecting the potential disturbance in the environment.

[0149] In order to ensure the reliability of the model training data, the active learning module performs multi-level data cleaning to remove outliers and redundant information. Multi-level data cleaning includes the following steps: primary data denoising, outlier detection, and redundancy removal. Primary denoising uses a median filter to process the feedback data to eliminate high-frequency noise, and outlier detection is performed through statistical methods, where data with a standard deviation exceeding a certain threshold is considered abnormal. Redundant information removal is achieved by checking the similarity of time series data to reduce redundant training of the model.

[0150] After collecting and cleaning the execution feedback data, the active learning module inputs it into the adaptive modeling module for model update. The model update adopts the meta-learning mechanism to adjust the model parameters, such as learning rate and weight, by analyzing the feedback data. The calculation expression of the adaptive model update is as follows:

[0151]

[0152] Among them, θ new represents the updated model parameters; θ old is the current model parameter; η update is the learning rate adjusted from the active learning module; is the gradient loss function with respect to the feedback data y.

[0153] This active learning mechanism enables the system to continuously perform adaptive optimization and learning in complex high temperature and high humidity environments to improve the accuracy of track and obstacle recognition. For example, when the recognition accuracy decreases due to humidity or temperature fluctuations in a tunnel environment, the active learning module can quickly collect new execution feedback and generate feedback data for learning through an automatic labeling algorithm. After analyzing and cleaning this data, the system updates the model parameters so that the adaptive modeling module can quickly adjust to cope with new environmental conditions.

[0154] In the experiment, the active learning module was applied to a simulated plateau tunnel environment, and the results showed that the system improved its recognition accuracy by 10% and reduced its path optimization time by 15% based on multiple rounds of execution feedback. This enhanced model learning and updating capability enables the system to maintain real-time response and efficient recognition in harsh environments, meeting the strict requirements for safe and efficient operation of trains.

[0155] like Figure 6 As shown, a system for implementing the plateau construction tunnel track image recognition method, the system comprises:

[0156] The ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging devices installed at the end of the train transporting materials are used to collect ultrasonic depth data, optical image data and spectral feature data in the tunnel during the operation of the train; the core of the system is a variety of sensors installed at the end of the train transporting materials: ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging devices. These sensors collect ultrasonic depth data, optical image data and spectral feature data in the tunnel during the operation of the train. Through the combination of multiple sensors, diverse information about the environment in the tunnel can be obtained from different physical levels, such as geometric shapes and material characteristics of objects, providing rich input for subsequent data fusion and feature extraction.

[0157] The time synchronization module is used to time align and preprocess the collected ultrasonic depth data, optical image data, and spectral feature data to generate synchronized cleaning data; in order to ensure the consistency of data collected by different sensors in time and space, the time synchronization module time aligns and preprocesses the collected ultrasonic depth data, optical image data, and spectral feature data to generate synchronized cleaning data. This step achieves preliminary data preprocessing through timestamp calibration and noise removal, thereby improving the accuracy and stability of data fusion in subsequent steps.

[0158] The multimodal data fusion module includes a multiscale convolution and hyperspectral fusion unit, which is used to extract geometric feature data and spectral enhancement feature data from the synchronous cleaned data, and perform feature fusion to generate fused feature data; the multimodal data fusion module integrates multiscale convolution and hyperspectral fusion units, which can extract geometric feature data and spectral enhancement feature data from the synchronous cleaned data. The multiscale convolution network extracts geometric feature data layer by layer by using convolution kernels with different receptive fields to capture the morphological features and texture details inside the tunnel. The spectral enhancement unit further enriches the feature space through spectral data, so that the system can still maintain the robustness of recognition when facing optical interference in high temperature and high humidity environments. The fused feature data is the final output feature set, which contains the joint information of various sensor data.

[0159] The adaptive modeling module is configured to receive the fused feature data and optimize it using the spectral-geometric feature collaborative optimization algorithm, and update the model parameters of the adaptive modeling module through adaptive training to generate optimized feature data; the adaptive modeling module receives the fused feature data and uses the spectral-geometric feature collaborative optimization algorithm for data processing. This algorithm optimizes the overall data by balancing the weights of geometric and spectral features to ensure the stability and recognition accuracy of the model in complex environments. In addition, the module has a built-in adaptive training mechanism that dynamically updates the model parameters through real-time feedback, so that the model maintains its self-adjustment ability in a changing environment and generates optimized feature data.

[0160] A real-time prediction and multi-objective decision-making system, including a variational autoencoder and a reinforcement learning module, is used to generate a latent space distribution based on the optimized feature data, perform path simulation and path optimization to generate path prediction data and path optimization data; the optimized feature data is input into the real-time prediction and multi-objective decision-making system, which includes a variational autoencoder (VAE) and a reinforcement learning module. VAE is used to map the optimized feature data to the latent space, thereby modeling and predicting the distribution of complex track environments. Combined with the reinforcement learning module, the system is able to perform path simulation and path optimization, and output path prediction data and path optimization data. This combination enables the system to make effective decisions and path adjustments in an incomplete or noisy environment.

[0161] The active learning module is used to collect execution feedback of path optimization data, and actively learn and update the model of the adaptive modeling module to improve the recognition accuracy and adaptability of the system in high temperature and high humidity environments. The active learning module is used to collect execution feedback of path optimization data and further update the model parameters of the adaptive modeling module. Through the automatic labeling algorithm, the module labels the feedback data and performs multi-level data cleaning to remove noise and redundant data. This ensures that the model maintains efficient learning ability in harsh environments such as high temperature and humidity and unstable sensors, and improves the recognition accuracy and environmental adaptability of the overall system.

[0162] Through the multi-module collaboration of the above system, the present invention enables trains to efficiently identify tracks and surrounding obstacles, respond in a timely manner, and optimize paths in complex tunnel environments at altitudes above 5,000 meters. Compared with the prior art, the present invention not only overcomes the impact of high temperature and humidity on sensors and data processing, but also achieves continuous performance improvement through the mechanism of adaptive modeling and active learning, ensuring the safety and stability of train operation.

[0163] A storage medium stores a computer program, which, when executed by a processor, implements the steps of the plateau construction tunnel track image recognition method.

[0164] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware.

[0165] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for recognizing track images of tunnels under plateau construction, characterized in that: The following steps are involved: Ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging equipment are installed at the end of the train transporting materials to collect ultrasonic depth data, optical image data and spectral feature data in the tunnel during the operation of the train, and perform data synchronization and preprocessing to generate synchronous cleaning data; The synchronized cleaned data is input into the multimodal data fusion module, and the geometric feature data and spectral enhancement feature data are extracted through multi-scale convolution and hyperspectral fusion, and feature fusion is performed to generate fused feature data; The fused feature data is input into the adaptive modeling module, the fused feature data is optimized using the spectral-geometric feature collaborative optimization algorithm, and the model parameters of the adaptive modeling module are updated through adaptive training to generate optimized feature data; The optimized feature data is input into the real-time prediction and multi-objective decision-making system, the potential space distribution is generated through the variational autoencoder, and the path simulation and path optimization are performed in combination with the reinforcement learning module to generate path prediction data and path optimization data; The active learning module collects execution feedback of path optimization data and performs active learning and model updating on the adaptive modeling module.

2. The plateau construction tunnel track image recognition method according to claim 1 is characterized in that: In the data synchronization and preprocessing steps, the ultrasonic depth data, optical image data and spectral feature data are time-aligned and noise-removed through the time synchronization module.

3. The plateau construction tunnel track image recognition method according to claim 1 is characterized in that: The multimodal data fusion module adopts a multi-scale convolutional network to extract geometric feature data layer by layer through convolution kernels with different receptive fields, and embeds spectral enhancement feature data in each convolution layer to generate fused feature data.

4. The plateau construction tunnel track image recognition method according to claim 1 is characterized in that: The spectral-geometric feature collaborative optimization algorithm optimizes the geometric feature data and the spectral feature data by the following calculation formula: W opt =W geo +α·W spec Among them, the optimized weight combination is W opt ; The geometric feature weight is W geo ; The spectral feature weight is W spec ; The weight adjustment coefficient is α.

5. The plateau construction tunnel track image recognition method according to claim 1 is characterized in that: During the adaptive training process, the adaptive modeling module introduces a meta-learning mechanism to enable the adaptive modeling module to automatically adjust the learning rate and weight parameters according to feedback data.

6. The plateau construction tunnel track image recognition method according to claim 1 is characterized in that: The real-time prediction and multi-objective decision-making system generates a latent space distribution through a variational autoencoder, and the latent space distribution is described by the following formula: p(Z|X)=N(μ(X),σ(X)) Among them, the latent space variable is Z; the optimized feature data is X; the mean of the feature data is μ(X); and the standard deviation of the feature data is σ(X).

7. The plateau construction tunnel track image recognition method according to claim 6 is characterized in that: The reinforcement learning module includes a policy network and a value network. The policy network performs path simulation based on the generated latent space distribution, and the value network is used to evaluate the safety and efficiency of the path simulation to generate path prediction data and path optimization data.

8. The plateau construction tunnel track image recognition method according to claim 1 is characterized in that: When collecting execution feedback of the path optimization data, the active learning module generates annotated feedback data through an automatic annotation algorithm and performs multi-level data cleaning to remove outliers and redundant information.

9. A system for implementing the plateau construction tunnel track image recognition method according to any one of claims 1 to 8, characterized in that: The system includes: Ultrasonic imaging sensors, short-wave infrared sensors and hyperspectral imaging equipment installed at the end of the train transporting materials are used to collect ultrasonic depth data, optical image data and spectral feature data in the tunnel during the train operation; A time synchronization module, used for time alignment and preprocessing of the collected ultrasonic depth data, optical image data and spectral feature data to generate synchronized cleaning data; A multimodal data fusion module, including a multi-scale convolution and hyperspectral fusion unit, is used to extract geometric feature data and spectral enhancement feature data from the synchronous cleaned data, and perform feature fusion to generate fused feature data; An adaptive modeling module is configured to receive the fused feature data and perform optimization processing using a spectral-geometric feature collaborative optimization algorithm, and to update model parameters of the adaptive modeling module through adaptive training to generate optimized feature data; A real-time prediction and multi-objective decision-making system, including a variational autoencoder and a reinforcement learning module, for generating a latent space distribution based on optimized feature data, performing path simulation and path optimization to generate path prediction data and path optimization data; The active learning module is used to collect execution feedback of path optimization data and perform active learning and model update on the adaptive modeling module to improve the recognition accuracy and adaptability of the system in high temperature and high humidity environments.

10. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the plateau construction tunnel track image recognition method described in any one of claims 1 to 8 are implemented.

Images