Automatic obstacle avoidance and adaptive terrain inspection rail car for photovoltaic power station

Abstract

The invention provides an automatic obstacle avoidance and adaptive terrain inspection rail car for a photovoltaic power station, and belongs to the technical field of photovoltaic power station inspection equipment. A plurality of hardware modules such as a control chip, a rail walking device, an obstacle avoidance sensing device, a terrain adaptation device and an inspection equipment device are integrated; the cooperative work flow of the steps of system initialization, task receiving, path planning, movement starting, obstacle avoidance detection, obstacle classification, obstacle avoidance decision making, obstacle avoidance execution, terrain detection, attitude adjustment, inspection execution, data preprocessing, fault detection, data transmission, energy management, exception handling, task completion and the like is achieved. Particularly, two core algorithms of a photovoltaic rail car multi-target obstacle avoidance path planning function and a photovoltaic module multi-mode defect detection pre-training neural network model are adopted, and the technical problem that in the prior art, a photovoltaic power station cannot achieve efficient and intelligent inspection in a complex environment is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power station inspection equipment. Specifically, it relates to a photovoltaic power station automatic obstacle avoidance and adaptive terrain inspection rail vehicle. Background Art

[0002] Inspection of photovoltaic power stations is an important link to ensure the safe and stable operation of photovoltaic systems. Traditional inspections mainly rely on manual hand-held devices, and later, rail vehicle inspection systems have been gradually developed to improve efficiency. However, existing rail vehicle inspection systems mainly adopt fixed rail paths and rely on preset programs to run, equipped with single detection devices, with limited data collection and analysis capabilities, and are unable to adapt to the complex and changeable photovoltaic power station environment.

[0003] With the continuous expansion of the scale of photovoltaic power stations and the increasingly complex installation environment, traditional rail vehicle inspection systems have many defects: First, they lack intelligent obstacle avoidance capabilities and cannot autonomously identify and avoid obstacles on the track; second, it is difficult to adapt to complex terrains and cannot maintain a stable posture in areas with large slope changes; third, the detection means are single, and it is difficult to comprehensively and accurately identify various defect types of photovoltaic modules; fourth, the data processing ability is limited, and real-time fault diagnosis and early warning cannot be achieved.

[0004] These problems lead to low inspection efficiency, high missed inspection rate, and increased operation and maintenance costs in photovoltaic power stations. Especially in complex environments, traditional inspection systems are difficult to achieve efficient and intelligent inspections, seriously affecting the operation efficiency and safety of photovoltaic power stations. There is an urgent need for a rail vehicle inspection system that can automatically avoid obstacles, adapt to terrain, and has multi-modal detection and analysis capabilities. That is to say, there is a technical problem in the prior art that efficient and intelligent inspections cannot be achieved in photovoltaic power stations in complex environments. Summary of the Invention

[0005] In view of this, the present invention provides a photovoltaic power station automatic obstacle avoidance and adaptive terrain inspection rail vehicle, which can solve the technical problem in the prior art that efficient and intelligent inspections cannot be achieved in photovoltaic power stations in complex environments.

[0006] The present invention is implemented as follows: The present invention provides a photovoltaic power station automatic obstacle avoidance and adaptive terrain inspection rail vehicle, which includes a control chip, a rail walking device, an obstacle avoidance sensing device, a terrain adaptation device, an inspection equipment device, a power management device, a data storage device, a communication transmission device, an environmental monitoring device, and a human-computer interaction device; a system control module is provided in the control chip, which is used to execute system initialization, receive inspection tasks, path planning, start movement, obstacle avoidance detection, obstacle classification, obstacle avoidance decision-making, obstacle avoidance execution, terrain detection, attitude adjustment, inspection execution, data preprocessing, fault detection, data transmission, energy management, exception handling, and task completion steps; among them, the obstacle avoidance decision-making step calls the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle, comprehensively considering multi-dimensional objectives such as path length, safety distance, energy consumption assessment, time efficiency, and equipment stability; the fault detection step calls the pre-trained neural network model for multi-modal defect detection of photovoltaic modules to achieve multi-source data fusion analysis and precise detection and positioning of photovoltaic module defects.

[0007] Among them, the control chip uses an industrial-grade 64-bit quad-core processor with a main frequency of not less than 2.5 GHz, an 8GB high-speed cache built-in, a hardware acceleration unit for image processing and deep learning algorithm operations, supports multi-thread parallel processing, has a real-time operating system built-in, has a temperature compensation function, can operate stably within the temperature range of -40°C to 85°C, the voltage input range is 9V to 36V, has overcurrent protection and overheat protection functions, and supports multiple communication interfaces.

[0008] Among them, the rail walking device includes four groups of walking mechanisms, and each group of walking mechanisms includes a driving wheel, a driven wheel, a driving motor, and a reducer; the driving wheel and the driven wheel use high-strength rubber tires, the tire surface is provided with anti-slip patterns, and steel cord is embedded inside the tires; the driving motor is a brushless DC servo motor; the reducer uses a planetary gear reducer; the rail walking device also includes an independent suspension system, which uses a combined structure of a spiral spring and a pneumatic shock absorber.

[0009] Among them, the obstacle avoidance sensing device includes a lidar, ultrasonic sensors, and visual cameras; the lidar is installed on the top of the vehicle body and has a horizontal scanning angle and a vertical scanning angle; there are 8 ultrasonic sensors in total, which are respectively installed in the front, rear, left, and right directions of the vehicle body; there are 4 visual cameras in total, which are respectively installed in the front, rear, left, and right directions of the vehicle body; the data acquisition frequency of the obstacle avoidance sensing device is 50 times per second.

[0010] Among them, the terrain adaptation device includes a hydraulic lifting system, an attitude adjustment mechanism, and a terrain sensor; the hydraulic lifting system consists of four hydraulic cylinders, and each hydraulic cylinder is installed at the four corners of the vehicle body bottom; the attitude adjustment mechanism includes a pitch adjustment component and a roll adjustment component; the terrain sensor includes an inclination sensor and a height sensor; the data acquisition frequency of the terrain adaptation device is 20 times per second.

[0011] Among them, when the system control module executes the obstacle avoidance decision-making step, it first uses the Dijkstra algorithm for global path planning to determine the general obstacle avoidance direction, and then uses the A* algorithm for local path optimization to find the optimal obstacle avoidance path; when executing the attitude adjustment step, it leans forward on the uphill section, leans backward on the downhill section, leans sideways on the turning section, and compensates for the height on the uneven section to ensure the stable operation of the vehicle body.

[0012] Among them, the input of the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle includes the obstacle position data set, the obstacle type identifier, the current vehicle position coordinates, the target position coordinates, and the track terrain parameter set; the output is the optimized obstacle avoidance path point sequence, the vehicle speed parameters at each path point, the attitude adjustment parameters, and the energy consumption estimation value; the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle adopts an improved multi-objective optimization algorithm.

[0013] Among them, the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle internally implements an obstacle threat assessment mechanism, calculates the comprehensive threat index according to the obstacle type, size, distance, and movement characteristics, and dynamically adjusts the conservatism of the obstacle avoidance strategy; at the same time, it integrates an orbital topology constraint processing module, which can identify feasible paths in a complex orbital network; it also implements an experience learning mechanism based on historical obstacle avoidance data, and continuously optimizes the decision-making model by accumulating successful obstacle avoidance cases.

[0014] Among them, the pre-trained neural network model structure for multi-modal defect detection of photovoltaic modules is a multi-modal fusion network based on the Transformer architecture, including four parallel modality-specific encoders and a cross-modal fusion decoder; the cross-modal fusion decoder adopts a hierarchical attention mechanism, including a two-stage fusion strategy; the model output layer includes a defect type classification head and a defect position localization head, supporting the accurate detection and localization of 17 common photovoltaic module defects.

[0015] Among them, the pre-training process of the pre-trained neural network model for multi-modal defect detection of photovoltaic modules includes modality-specific pre-training, modality fusion pre-training, end-to-end fine-tuning, knowledge distillation, environment adaptability training, and online learning mechanism design; among them, the visible light image encoder is pre-trained on the ImageNet dataset, the infrared thermal image encoder is pre-trained on the thermal imaging dataset, the ultraviolet image encoder is pre-trained on the discharge detection dataset, and the electrical parameter encoder is pre-trained on the photovoltaic power station operation dataset.

[0016] Compared with the prior art, a photovoltaic power station automatic obstacle avoidance and adaptive terrain inspection rail vehicle provided by the present invention. The photovoltaic power station automatic obstacle avoidance and adaptive terrain inspection rail vehicle system of the present invention realizes efficient and intelligent inspection in complex environments by integrating a variety of sensing devices, intelligent control algorithms, and multi-modal data analysis technologies. The system can autonomously identify and avoid various obstacles on the track, and by adopting a multi-objective obstacle avoidance path planning function, comprehensively considering factors such as path length, safety distance, and energy consumption, it realizes the optimal obstacle avoidance strategy.

[0017] At the same time, the terrain adaptation device equipped with the system can adjust the vehicle body posture in real time according to the track slope to ensure the stable operation of the inspection equipment under various terrain conditions. The combination of the multi-modal inspection equipment and the pre-trained neural network model can simultaneously process visible light, infrared, ultraviolet, and electrical parameter data to achieve precise detection and positioning of common photovoltaic module defects, greatly improving the accuracy and efficiency of defect identification.

[0018] The present invention solves the technical problem that efficient and intelligent inspection cannot be achieved in complex environments in photovoltaic power stations, significantly improves the inspection efficiency and quality, reduces the operation and maintenance costs, ensures the safe and stable operation of the power station, provides strong technical support for the intelligent operation and maintenance of photovoltaic power stations, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of the method of the present invention;

[0020] Figure 2 is a schematic diagram of the rail vehicle of the present invention;

[0021] Figure 3 is a schematic structural diagram of the rail walking device and the terrain adaptation device of the present invention;

[0022] In the drawings, the list of components represented by each reference numeral is as follows:

[0023] 01, emergency stop button; 1, rail walking device; 11, driving wheel; 12, driven wheel; 13, driving motor; 14, reducer; 21, lidar; 31, hydraulic lifting system; 32, attitude adjustment mechanism; 41, high-definition camera; 42, infrared thermal imager; 51, touch display screen. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] As Figure 1As shown, it is a flowchart of an automatic obstacle avoidance and adaptive terrain inspection rail vehicle for a photovoltaic power station provided by the present invention. The method includes the following steps:

[0026] An automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station includes a control chip, a rail walking device 1, an obstacle avoidance sensing device, a terrain adaptation device, an inspection equipment device, a power management device, a data storage device, a communication transmission device, an environmental monitoring device, and a human-machine interaction device. The control chip is electrically connected to the rail walking device, the obstacle avoidance sensing device, the terrain adaptation device, the inspection equipment device, the power management device, the data storage device, the communication transmission device, the environmental monitoring device, and the human-machine interaction device respectively. A system control module is provided in the control chip. The rail walking device is used to realize the stable driving of the rail vehicle on the rail. The obstacle avoidance sensing device is used to collect information on surrounding environmental obstacles. The terrain adaptation device is used to collect rail slope data and adjust the vehicle body posture. The inspection equipment device is used to collect images and thermal imaging data of photovoltaic power station components. The power management device is used to monitor the battery power and manage the power supply system. The data storage device is used to store inspection data and system operation parameters. The communication transmission device is used to realize data communication between the rail vehicle and the control center. The environmental monitoring device is used to collect data on environmental temperature, humidity, and light intensity. The human-machine interaction device is used to realize the interactive operation between personnel and the rail vehicle. The data acquisition frequency of the obstacle avoidance sensing device is 50 times per second, the data acquisition frequency of the terrain adaptation device is 20 times per second, the data acquisition frequency of the inspection equipment device is 5 times per second, and the data acquisition frequency of the environmental monitoring device is 1 time per minute.

[0027] The system control module is used to execute the following steps:

[0028] S01. System initialization: The control chip performs self-check and detects the connection status of each hardware device, reads the system configuration parameters, initializes the software module, establishes the mapping relationship between the hardware device and the software module, and sets the sampling frequency and data processing parameters of each sensor.

[0029] S02. Receive inspection tasks: Receive inspection task instructions through the human-machine interaction device or the communication transmission device, parse the task content, extract the inspection path, inspection items, and inspection parameters, and configure the working parameters of the inspection equipment device according to the task requirements.

[0030] S03. Path planning: Based on the preset rail map and task requirements, plan the optimal inspection path, calculate the path length, estimated inspection time, and energy consumption, and display the planning result through the human-machine interaction device.

[0031] S04. Start moving, control the drive motor of the rail walking device, execute the start-up process according to the planned path, achieve smooth acceleration, gradually increase to the inspection speed, and at the same time monitor the motor current, speed and temperature parameters to ensure the safety and reliability of the start-up process;

[0032] S05. Obstacle avoidance detection, collect the surrounding environment data through the obstacle avoidance sensing device, fuse the data of the lidar 21, ultrasonic sensor and vision camera, construct a three-dimensional environment model, identify the obstacles on and around the track, and calculate the position, size and motion characteristics of the obstacles;

[0033] S06. Obstacle classification, classify the detected obstacles into fixed obstacles, moving obstacles and temporary obstacles, determine the obstacle avoidance strategy according to the obstacle type, adopt the path replanning strategy for fixed obstacles, the speed adjustment strategy for moving obstacles, and the waiting strategy for temporary obstacles;

[0034] S07. Obstacle avoidance decision-making, based on the obstacle classification results, call the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle to perform obstacle avoidance path planning. This function combines the Dijkstra algorithm and the A algorithm. First, use the Dijkstra algorithm for global path planning to determine the general obstacle avoidance direction, and then use the A algorithm for local path optimization to find the optimal obstacle avoidance path; this function also considers multi-dimensional objectives such as path length, safety distance, energy consumption evaluation, time efficiency and equipment stability, and realizes the balance between objectives in different scenarios through an adaptive weight adjustment mechanism;

[0035] S08. Obstacle avoidance execution, according to the obstacle avoidance decision result, control the rail walking device to perform obstacle avoidance actions, including deceleration, stop, reverse and lane change operations, and at the same time continuously monitor the obstacle state and adjust the obstacle avoidance strategy in real time to ensure the safety and reliability of the obstacle avoidance process;

[0036] S09. Terrain detection, collect the data of the track slope, height and tilt angle through the terrain sensor of the terrain adaptation device, construct a track terrain model, identify uphill, downhill, turning and uneven sections, and provide data support for attitude adjustment;

[0037] S10. Attitude adjustment, based on the terrain detection results, control the hydraulic lifting system and attitude adjustment mechanism of the terrain adaptation device, and adjust the vehicle body attitude in real time, tilt forward on the uphill section, tilt backward on the downhill section, tilt sideways on the turning section, and perform height compensation on the uneven section to ensure the smooth operation of the vehicle body;

[0038] S11. Inspection execution, control the inspection equipment device to collect the data of the photovoltaic modules, including visible light images, infrared thermal images, ultraviolet images and electrical parameters, and at the same time collect the environmental parameters, including temperature, humidity, light intensity and air pressure, and store the collected data in the data storage device in real time;

[0039] S12. Data preprocessing: Preprocess the collected inspection data, including image denoising, enhancement, registration, and segmentation, thermal image temperature correction, electrical parameter normalization, and environmental parameter compensation, to improve the data quality and provide a basis for subsequent analysis.

[0040] S13. Fault detection: Based on the preprocessed data, call the pre-trained neural network model for multi-modal defect detection of photovoltaic modules. This model uses a variety of fault detection algorithms, including image defect detection based on deep learning, hot spot detection based on temperature distribution, and performance degradation detection based on electrical parameters. It can simultaneously process visible light images, infrared thermal images, ultraviolet images, and electrical parameter data, realize the fusion analysis of multi-source data, and accurately detect and locate common photovoltaic module defects, comprehensively evaluate the health status of photovoltaic modules.

[0041] S14. Data transmission: Transmit the inspection results, fault information, and system status data to the remote control center through the communication transmission device. At the same time, adaptively adjust the transmission strategy according to the network status, transmit full-scale data when the network is good, and transmit key data when the network is unstable.

[0042] S15. Energy management: Monitor the battery power through the power management device, and dynamically adjust the inspection speed, sampling frequency, and data processing strategy according to the remaining inspection tasks and battery status to optimize energy utilization and extend the inspection time.

[0043] S16. Exception handling: Monitor the system operation status, detect hardware failures, software exceptions, and environmental changes, and execute corresponding handling strategies according to the exception type, including automatic recovery, degraded operation, and safe shutdown, to ensure the safe and reliable operation of the system.

[0044] S17. Task completion: After the inspection task is completed, generate an inspection report, including the inspection path, inspection time, fault statistics, and system status, and send it to the remote control center through the communication transmission device. At the same time, return to the starting position and enter the standby state.

[0045] Through the collaborative work of the above hardware devices and the system control module, the automatic obstacle avoidance and adaptive terrain inspection rail vehicle system of the photovoltaic power station can achieve autonomous inspection in complex environments, effectively improve the inspection efficiency and quality of the photovoltaic power station, reduce the operation and maintenance costs, and ensure the safe and stable operation of the power station.

[0046] The multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle is used to optimize the S07 obstacle avoidance decision-making step to achieve more efficient obstacle avoidance path planning. The inputs include the obstacle position data set, obstacle type identifier, current vehicle position coordinates, target position coordinates, and track terrain parameter set. The outputs are the optimized obstacle avoidance path point sequence, vehicle speed parameters at each path point, attitude adjustment parameters, and energy consumption estimation values. The multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle adopts an improved multi-objective optimization algorithm, which not only considers traditional factors such as path length and safety distance, but also integrates multi-dimensional objectives such as energy consumption assessment, time efficiency, and equipment stability, and achieves a balance between objectives in different scenarios through an adaptive weight adjustment mechanism. The multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle internally implements an obstacle threat degree assessment mechanism, calculates a comprehensive threat index according to the type, size, distance, and movement characteristics of the obstacle, and dynamically adjusts the conservativeness of the obstacle avoidance strategy. At the same time, the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle integrates a track topology constraint processing module, which can identify feasible paths in a complex track network and avoid planning path schemes that do not conform to the physical constraints of the track. The multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle also implements an experience learning mechanism based on historical obstacle avoidance data, continuously optimizes the decision-making model by accumulating successful obstacle avoidance cases, and improves the planning efficiency and safety.

[0047] The pre-trained neural network model for multi-modal defect detection of photovoltaic modules is used to optimize the S13 fault detection step and improve the accuracy and efficiency of photovoltaic module defect detection. The pre-trained neural network model for multi-modal defect detection of photovoltaic modules can simultaneously process visible light images, infrared thermal images, ultraviolet images, and electrical parameter data, realize the fusion analysis and comprehensive diagnosis of multi-source data, and its parameter configuration and sampling frequency are dynamically adjusted according to the temperature, humidity, and light intensity data collected by the environmental monitoring device.

[0048] The specific structure of the pre-trained neural network model for multi-modal defect detection of photovoltaic modules is a multi-modal fusion network based on the Transformer architecture, which includes four parallel modality-specific encoders and a cross-modal fusion decoder. The visible light image encoder adopts an improved Vision Transformer structure, which includes 12 layers of attention mechanisms and introduces a local perception enhancement module to improve the ability to extract fine defect features; the infrared thermal image encoder adopts 8 layers of attention mechanisms and designs a temperature gradient feature enhancement layer to strengthen the recognition of thermal anomaly patterns; the ultraviolet image encoder adopts 6 layers of attention mechanisms and integrates a discharge feature enhancement module; the electrical parameter encoder adopts a 4-layer MLP structure to achieve high-dimensional feature representation of electrical parameters. The cross-modal fusion decoder adopts a hierarchical attention mechanism, which includes a two-stage fusion strategy: the first stage performs coarse-grained alignment between modalities, and the second stage realizes fine-grained feature fusion through a sparse attention mechanism, where the sparse attention window size parameter is dynamically adjusted according to the data acquisition frequency of the inspection device (5 times per second), and the window step is associated with the data acquisition frequency of the terrain adaptation device (20 times per second). The model output layer includes a defect type classification head and a defect location positioning head, which support the accurate detection and positioning of various common photovoltaic module defects.

[0049] The steps for establishing the training data set in the pre-training process of the pre-trained neural network model for multi-modal defect detection of photovoltaic modules specifically include: First, collect module data from different types of photovoltaic power stations in multiple regions across the country, covering single-crystalline silicon, polycrystalline silicon, and thin-film multi-flow photovoltaic technologies, and at the same time ensure that the data covers photovoltaic modules with different climate conditions, different operation years, and different installation methods; Second, perform multi-modal data acquisition, including using a calibrated high-definition camera to collect visible light images, using an infrared thermal imager to collect thermal images, using an ultraviolet detector to collect ultraviolet images, and using an electrical parameter measuring instrument to collect electrical parameters such as voltage, current, and power; Then, photovoltaic field experts annotate the collected data, identify and mark various defects, including common defect types such as hot spots, hidden cracks, virtual soldering, PID, aging, and occlusion, and grade the severity of each defect; Next, perform data preprocessing, including image registration, noise removal, brightness normalization, temperature correction, and electrical parameter normalization, and align different modality data according to the time stamp; Finally, construct balanced training sets, validation sets, and test sets to ensure that the number of samples of various defect types is balanced, and expand the training samples through data augmentation techniques (including random cropping, rotation, flipping, and noise addition), especially for rare defect types, an oversampling strategy is adopted, and finally a large-scale training data set containing more than 200,000 groups of multi-modal samples is constructed.

[0050] The pre-training steps of the pre-trained neural network model for multi-modal defect detection of photovoltaic modules specifically include: First, perform modality-specific pre-training, and pre-train four modality-specific encoders on large-scale datasets in their respective fields. The visible light image encoder is pre-trained on the ImageNet dataset, the infrared thermal image encoder is pre-trained on the thermal imaging dataset, the ultraviolet image encoder is pre-trained on the discharge detection dataset, and the electrical parameter encoder is pre-trained on the photovoltaic power plant operation dataset. Secondly, perform modality fusion pre-training. Use the constructed multi-modal photovoltaic defect dataset and adopt a contrastive learning strategy to train the cross-modal fusion decoder to optimize the representation alignment and information complementarity between modalities. Then, perform end-to-end fine-tuning. Connect the pre-trained encoder and decoder to form a complete network, and use multi-modal data with defect annotations for end-to-end training. Adopt a multi-task learning framework to simultaneously optimize the defect classification and localization tasks. Next, perform knowledge distillation, distill knowledge from a large expert model to a lightweight model to reduce computational complexity and adapt to the limited computational resources of the rail vehicle. Subsequently, perform environmental adaptability training, use data under different environmental conditions for model fine-tuning, and improve the robustness of the model under various lighting, temperature, and humidity conditions. Finally, design an online learning mechanism, combine the rail vehicle system to construct an incremental learning framework, enable the model to continuously learn and optimize from newly collected data, and gradually improve the adaptability to the specific photovoltaic power plant environment. The pre-trained neural network model for multi-modal defect detection of photovoltaic modules can operate efficiently on the hardware acceleration unit of the control chip, achieve millisecond-level fault detection response, greatly improve the accuracy and efficiency of the S13 fault detection step, and provide strong support for the intelligent operation and maintenance of photovoltaic power plants.

[0051] The following describes the specific implementation manners of the above steps in detail.

[0052] The control chip uses an industrial-grade 64-bit quad-core processor with a main frequency of not less than 2.5 GHz, an 8GB high-speed cache built-in, a hardware acceleration unit for image processing and deep learning algorithm operations, supports multi-thread parallel processing, has a real-time operating system built-in, has a temperature compensation function, can operate stably in the temperature range of -40°C to 85°C, the voltage input range is 9V to 36V, has over-current protection and overheat protection functions, and supports multiple communication interfaces including USB, UART, SPI, I2C, CAN, and Ethernet interfaces to ensure the efficient and stable operation of the system.

[0053] The rail walking device 1 includes four groups of walking mechanisms. Each group of walking mechanisms includes a driving wheel 11, a driven wheel 12, a driving motor 13 and a reducer 14. The driving wheel and the driven wheel adopt high-strength rubber tires. The tire surface is provided with anti-slip patterns. Steel cord is embedded inside the tire. The pattern depth is not less than 8 mm. The driving motor is a brushless DC servo motor with a rated power of not less than 500 W, a rated speed of not less than 3000 rpm, and a maximum torque of not less than 20 Nm. The reducer adopts a planetary gear reducer with a reduction ratio of 50:1 and a transmission efficiency of not less than 95%. The rail walking device also includes an independent suspension system, which adopts a combined structure of a spiral spring and a pneumatic shock absorber, and has a stroke buffering capacity of 10 cm to ensure the stable driving of the rail vehicle on an uneven track.

[0054] The obstacle avoidance sensing device includes a lidar 21, ultrasonic sensors and a vision camera. The lidar is installed on the top of the vehicle body with a horizontal scanning angle of 360° and a vertical scanning angle of 30°. The ranging range is from 0.1 m to 100 m, the angular resolution is 0.1°, and the ranging accuracy is ±2 cm. There are 8 ultrasonic sensors in total, which are respectively installed in four directions of the front, rear, left and right of the vehicle body, with 2 installed in each direction. The ranging range is from 0.02 m to 5 m, and the ranging accuracy is ±1 cm. There are 4 vision cameras in total, which are respectively installed in four directions of the front, rear, left and right of the vehicle body, with a pixel of not less than 12 million and a viewing angle of not less than 120°, and support the night vision function. The obstacle avoidance sensing device is used to detect obstacle information in all directions and achieve accurate measurement of the position, shape and motion characteristics of obstacles.

[0055] The terrain adaptation device includes a hydraulic lifting system 31, an attitude adjustment mechanism 32 and a terrain sensor. The hydraulic lifting system consists of four hydraulic cylinders, and each hydraulic cylinder is installed at four corners of the bottom of the vehicle body. The maximum stroke of the cylinder is 30 cm, the lifting speed is not less than 5 cm / s, and the lifting accuracy is not higher than 0.5 mm. The attitude adjustment mechanism includes a pitch adjustment component and a roll adjustment component. The pitch adjustment component can achieve a pitch angle adjustment of ±15°, and the roll adjustment component can achieve a roll angle adjustment of ±15°. The terrain sensor includes an inclination sensor and a height sensor. The measurement range of the inclination sensor is ±30°, and the measurement accuracy is 0.1°. The measurement range of the height sensor is from 0 cm to 50 cm, and the measurement accuracy is 0.5 cm. The terrain adaptation device is used to collect real-time data on the track slope and height changes, and automatically adjust the vehicle body attitude to ensure the stable operation of the inspection equipment under different terrain conditions.

[0056] The inspection device includes a high-definition camera 41, an infrared thermal imager 42, an ultraviolet detector, and an electrical parameter measuring instrument. The high-definition camera has a 4K resolution, a 30x optical zoom function, and supports image stabilization. The resolution of the infrared thermal imager is not less than 640×480, the temperature measurement range is -20°C to 500°C, and the temperature measurement accuracy is ±2°C. The ultraviolet detector is used to detect partial discharge phenomena in photovoltaic modules, and the detection sensitivity is not less than 1500μW / cm 2 , and the electrical parameter measuring instrument is used to measure the voltage, current, and power parameters of photovoltaic modules. The voltage measurement range is 0V to 1500V, the current measurement range is 0A to 20A, and the power measurement range is 0W to 30kW. The inspection device is used to collect various operating parameters of the components in the photovoltaic power station and realize the automatic detection and analysis of photovoltaic module defects.

[0057] The power management device includes a lithium battery pack, a charging controller, and a power monitoring module. The capacity of the lithium battery pack is not less than 200Ah, the rated voltage is 48V, and the cycle life is not less than 2000 times. The charging controller supports multiple charging methods, including photovoltaic charging, AC grid charging, and regenerative braking energy recovery charging, and the charging efficiency is not less than 95%. The power monitoring module is used to monitor the battery voltage, current, temperature, and remaining power in real time, and has overcharge protection, over-discharge protection, over-current protection, and temperature protection functions. The power management device is used to provide a stable power supply for the rail vehicle and extend the endurance time of the rail vehicle.

[0058] The data storage device includes a solid-state drive, a memory, and a data cache module. The capacity of the solid-state drive is not less than 2TB, the read and write speed is not less than 500MB / s, and it supports the hot-swap function. The capacity of the memory is not less than 16GB, and the access speed is not less than 2400MHz. The data cache module is used to temporarily store high-speed collected data, and the cache capacity is not less than 128MB. The data storage device is used to store various data collected during the inspection, including image data, thermal imaging data, obstacle information, and system operating parameters.

[0059] The communication and transmission device includes a 5G communication module, a WiFi module, and a short-range wireless communication module. The 5G communication module is used to achieve high-speed data transmission between the rail vehicle and the remote control center. The upload rate is not less than 100Mbps, and the download rate is not less than 500Mbps. The WiFi module supports the 802.11ac protocol, the transmission rate is not less than 1Gbps, and the communication distance is not less than 100m. The short-range wireless communication module uses Bluetooth 5.0 technology, and the communication distance is not less than 50m. The communication and transmission device is used to achieve data exchange between the rail vehicle and external devices, and supports remote control and real-time data transmission functions.

[0060] The environmental monitoring device includes a temperature sensor, a humidity sensor, a light intensity sensor, and a barometric pressure sensor. The measurement range of the temperature sensor is from -40°C to 85°C, and the measurement accuracy is ±0.5°C. The measurement range of the humidity sensor is from 0% to 100%, and the measurement accuracy is ±3%. The measurement range of the light intensity sensor is from 0 lux to 100000 lux, and the measurement accuracy is ±5%. The measurement range of the barometric pressure sensor is from 300 hPa to 1100 hPa, and the measurement accuracy is ±1 hPa. The environmental monitoring device is used to collect various parameters of the operating environment of the rail vehicle and provide an environmental reference for the analysis of inspection data.

[0061] The human-machine interaction device includes a touch display screen 51, control buttons, and a voice interaction module. The touch display screen uses a 10.1-inch high-brightness screen with a resolution of not less than 1920×1200 and a brightness of not less than 500 nit, supports multi-touch, and has a waterproof and dustproof function. The control buttons include a power button, an emergency stop button 01, and function shortcut buttons. The voice interaction module supports voice recognition and voice synthesis functions, and the recognition accuracy is not less than 95%. The human-machine interaction device is used to realize the interaction operation between personnel and the rail vehicle and supports functions such as parameter setting, task distribution, and status query.

[0062] In the steps of the system control module, the specific implementation of step S01 is the system initialization process. First, the control chip executes a self-check program to verify the hardware integrity, including memory testing, processor performance testing, and peripheral connection status detection. The boundary scan technology is used during the testing process to ensure that all hardware interfaces work properly. Then, the system configuration parameter file is read. This file is stored in the non-volatile memory area of the data storage device and contains sensor calibration parameters, system operation thresholds, and default inspection parameters. The configuration file is stored in JSON format for easy parsing. Next, each software module is initialized, including the obstacle avoidance module, terrain adaptation module, inspection module, and communication module. The layered startup strategy is adopted during the initialization process to ensure that the underlying driver module starts first. Subsequently, the mapping relationship between the hardware device and the software module is established. The physical device is mapped to the software interface through the device abstraction layer to achieve hardware-independent programming. Finally, the sampling frequency and data processing parameters of each sensor are set. The obstacle avoidance sensing device is set to 50 times per second, the terrain adaptation device is set to 20 times per second, the inspection device is set to 5 times per second, and the environmental monitoring device is set to 1 time per minute. The sampling frequency can be dynamically adjusted according to the task requirements. The system status thresholds include triggering the energy-saving mode when the battery power is lower than 20%, and triggering the overheat protection when the environmental temperature exceeds 60°C. This step ensures that all parts of the system are in a normal working state and provides a basic guarantee for the subsequent execution of the inspection task.

[0063] The specific implementation method of step S02 is to receive the inspection task process. First, the task instruction is received through the touch screen of the human-computer interaction device or the 5G module of the communication transmission device, and two task sources are supported: direct input by the on-site operator through the touch screen and the remote control center is issued through the network. Then the task content is parsed, and the received instruction is decoded into a standard task description format. The parsing adopts a recursive descent analysis algorithm to ensure that the complex task structure can also be correctly parsed. Then the inspection path, inspection items and inspection parameters are extracted. The inspection path includes the starting point coordinates, the end point coordinates and the list of must-pass points. The inspection items include image acquisition, thermal imaging scanning and electrical parameter measurement. The inspection parameters include inspection speed, image acquisition resolution and thermal imaging acquisition frequency. Finally, the working parameters of the inspection equipment are configured according to the task requirements, including the camera focal length setting, the infrared thermal imager temperature range configuration and the electrical parameter measurement mode selection. The configuration process uses a parameter mapping table to convert the task requirements into equipment configuration. This step realizes the standardized parsing and equipment configuration of the task instructions to ensure that the inspection equipment works correctly according to the task requirements.

[0064] The specific implementation method of step S03 is the path planning process. First, path planning is performed based on the preset track map and task requirements. The preset track map is stored in a graph data structure. The nodes represent track intersections, the edges represent track segments, and each edge contains length, slope and curvature attributes. Then, the improved shortest path algorithm is used for global path planning. The algorithm is optimized based on the Dykstra algorithm, and a heuristic function is introduced to reduce the search space. The heuristic function considers the straight-line distance from the node to the target and the track slope factor. Then, the path length, expected inspection time and energy consumption are calculated. The path length is obtained by accumulating the length of each edge on the selected path. The inspection time is calculated based on the path length, inspection speed and estimated stay time. The energy consumption is estimated by the energy consumption model. The model takes the path slope, inspection speed and equipment working power as input, and outputs the expected energy consumption value. The energy consumption model is trained from historical operation data using a multivariate regression method. Finally, the planning results are displayed through a human-computer interaction device, including path visualization, time estimation and energy consumption prediction. Planning thresholds include a single inspection route length of no more than 10 kilometers, an estimated inspection time of no more than 6 hours, and an estimated energy consumption of no more than 80% of the battery capacity. This step provides optimized route planning for inspection tasks, ensuring that tasks are completed efficiently and energy is used reasonably.

[0065] The specific implementation of step S04 is to start the moving process. First, the driving motor of the track travel device is controlled to start according to the planned path. The motor control adopts the vector control algorithm to achieve precise control of torque and speed. Then, a smooth acceleration is achieved, using the S-curve acceleration profile, and the initial acceleration is set to 0.5m / s 2 , the maximum acceleration does not exceed 2m / s 2 , the acceleration is controlled at 0.3m / s3 Within this range, ensure that there is no impact during the startup process. Then gradually increase the speed to the inspection speed, which is set according to the task requirements and is usually in the range of 1 - 3 m / s. The speed control adopts the PID control algorithm, with a proportional coefficient of 3.5, an integral coefficient of 0.8, and a derivative coefficient of 0.2. The parameters can be adaptively adjusted according to the load conditions. At the same time, monitor the motor current, speed, and temperature parameters. The current sampling frequency is 100 Hz, the speed sampling frequency is 50 Hz, and the temperature sampling frequency is 10 Hz. The collected data is subjected to noise suppression through the Kalman filter algorithm. The monitoring thresholds include that the motor current does not exceed 85% of the rated current, and the motor temperature does not exceed 75°C. Once the threshold is exceeded, immediately reduce the acceleration or start the heat dissipation measures. This step ensures a smooth and controllable startup process of the rail vehicle, protects the equipment, and extends its service life.

[0066] The specific implementation of step S05 is the obstacle avoidance detection process. First, collect the surrounding environment data through the obstacle avoidance sensing device. The lidar scans 360° at a frequency of 50 Hz, the ultrasonic sensor measures the distance at a frequency of 25 Hz, and the vision camera collects images at a frequency of 20 Hz. Then fuse the multi-sensor data. The multi-sensor data fusion algorithm unifies the data from different sensors into the same coordinate system. The fusion algorithm is based on the Kalman filter and the theory of evidence, and assigns credibility weights to different sensors in different scenarios. Close-range obstacles mainly rely on ultrasonic data, and long-range obstacles mainly rely on lidar data. Then construct a three-dimensional environment model. The octree structure is used to represent the environment, with a spatial resolution of 10 cm and a model update frequency of 10 Hz. The model construction adopts the probabilistic occupancy grid method to reduce the influence of sensor noise. Then identify the obstacles on and around the track. The obstacle recognition adopts an object detection algorithm based on deep learning, which runs on the hardware acceleration unit of the rail vehicle control chip and can real-time identify common obstacle types such as people, vehicles, and animals. Finally, calculate the position, size, and motion characteristics of the obstacles. The position accuracy is controlled within ±5 cm. The obstacle motion obtains the speed and direction information through continuous frame tracking, and the Kalman prediction algorithm is used to predict the future position of the obstacles. This step provides comprehensive and accurate environmental perception data for subsequent obstacle avoidance decisions.

[0067] The specific implementation of step S06 is the obstacle classification process. First, the detected obstacles are classified according to their characteristics, including three categories: fixed obstacles, moving obstacles, and temporary obstacles. Fixed obstacles refer to immovable objects on the track, such as large plant debris and collapsed structures; moving obstacles refer to objects with the ability of autonomous movement, such as people and animals; temporary obstacles refer to objects that briefly appear on the track, such as fallen leaves or small plant debris. The classification uses a multi-feature classification algorithm based on decision trees, and the features include obstacle size, duration, position change rate, and shape features. Then, the obstacle avoidance strategy is determined according to the obstacle type. For fixed obstacles, a path replanning strategy is adopted, and a new path around the obstacle needs to be found; for moving obstacles, a speed adjustment strategy is adopted, and the vehicle decelerates or pauses to wait for the obstacle to leave; for temporary obstacles, a waiting strategy is adopted, and it stays briefly and monitors the change of the obstacle state. The strategy selection thresholds include that an obstacle with a size greater than 30 cm and a duration exceeding 10 seconds is determined as a fixed obstacle, a position change rate greater than 0.2 m / s is determined as a moving obstacle, and an obstacle with a size less than 20 cm and a duration not exceeding 5 seconds is determined as a temporary obstacle. This step realizes the intelligent classification of obstacles and strategy matching, providing a basis for subsequent obstacle avoidance decisions.

[0068] The specific implementation of step S07 is the obstacle avoidance decision-making process. First, based on the obstacle classification results, the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle is called to perform obstacle avoidance path planning. This function combines the advantages of Dijkstra's algorithm and A* algorithm. First, Dijkstra's algorithm is used for global path search to determine the general obstacle avoidance direction. The algorithm sets the inflation distance parameter to 50 cm to ensure that the path maintains a safe distance from the obstacle. Then, local path optimization is carried out through A* algorithm. The heuristic function of A* algorithm is designed as the weighted sum of the straight-line distance to the target point and the turning cost, and the turning cost coefficient is set to 0.3 to reduce unnecessary turning of the path. This function simultaneously considers multiple optimization objectives, including minimizing the path length, maximizing the safe distance from the obstacle, minimizing the energy consumption, maximizing the time efficiency, and maximizing the equipment stability. The optimization uses the weighted sum method, and the weights of each objective are adjusted through an adaptive mechanism, which changes dynamically according to the obstacle threat level, the remaining battery power, and the task urgency. For example, when the battery power is lower than 30%, the energy consumption weight is increased, and when the distance to the obstacle is less than 1 m, the safe distance weight is increased. The obstacle avoidance planning results include a sequence of path points, a speed curve, and attitude adjustment parameters. The distance between path points is not greater than 0.5 m to ensure the smoothness and continuity of the path. This step realizes intelligent obstacle avoidance decision-making in a complex environment, balancing the requirements of safety and efficiency.

[0069] The specific implementation of step S08 is the obstacle avoidance execution process. First, according to the obstacle avoidance decision result, the track walking device is controlled to perform obstacle avoidance actions. The deceleration operation is achieved by reducing the motor speed, and an S-shaped speed curve is adopted, with the deceleration not exceeding 2 m / s 2; The stop operation is achieved through the electric motor braking system, and the maximum deceleration during emergency stop can reach 3 m / s 2 ; The reverse operation is achieved by reversing the motor direction, and the reverse speed is limited within 1 m / s; The lane change operation is achieved by controlling the differential speed of the four groups of traveling mechanisms, and the steering angular velocity does not exceed 30° / second. Then continuously monitor the obstacle state, with the obstacle state update frequency being 20 Hz, and obtain real-time obstacle information through the obstacle avoidance sensing device. Next, adjust the obstacle avoidance strategy in real time. Based on the latest obstacle information, adopt the rolling horizon optimization method to dynamically adjust the obstacle avoidance plan, with the optimization period being 100 milliseconds, ensuring that the strategy is updated in a timely manner as the environment changes. The obstacle avoidance execution threshold includes a minimum safe distance of 0.8 meters from the obstacle, and an emergency stop is triggered when encountering an unavoidable obstacle with a distance less than 0.5 meters. This step realizes the precise execution of the obstacle avoidance decision, ensuring the safety and reliability of the obstacle avoidance process.

[0070] The specific implementation of step S09 is the terrain detection process. First, collect data on the track slope, height, and tilt angle through the terrain sensors of the terrain adaptation device. The sampling frequency of the tilt angle sensor is 20 Hz, and the sampling frequency of the height sensor is 10 Hz. Then construct an orbital terrain model, and use the piecewise cubic spline interpolation method to connect discrete measurement points into a continuous terrain curve. The spline function parameters are determined by the least squares method to ensure a smooth transition of the curve. Next, identify special terrain segments, including uphill segments, downhill segments, turning segments, and uneven segments. The judgment threshold for uphill and downhill is that the absolute value of the slope is greater than 5°, the judgment threshold for turning is that the radius of curvature is less than 10 meters, and the judgment threshold for unevenness is that the height difference between adjacent points is greater than 2 cm. Terrain parameter extraction includes the maximum slope, slope change rate, radius of curvature, and road surface roughness. These parameters are used for subsequent attitude adjustment. The terrain detection results are transmitted to the control chip in real time through the control bus, providing data support for attitude adjustment. This step realizes the precise perception of the orbital terrain, providing necessary information for vehicle body attitude adjustment.

[0071] The specific implementation of step S10 is the attitude adjustment process. First, based on the terrain detection results, the hydraulic lifting system and attitude adjustment mechanism of the terrain adaptation device are controlled to adjust the vehicle body attitude in real time. When adjusting the forward tilt on an uphill section, the front hydraulic cylinder contracts and the rear hydraulic cylinder extends. The forward tilt angle is calculated according to the slope, usually 75% - 85% of the slope angle, and the angle control accuracy is ±0.5°. When adjusting the rear tilt on a downhill section, the front hydraulic cylinder extends and the rear hydraulic cylinder contracts. The rear tilt angle is also 75% - 85% of the slope angle. When adjusting the roll on a turning section, the inner hydraulic cylinder extends and the outer hydraulic cylinder contracts. The roll angle is calculated according to the radius of curvature, and the angle is usually in the range of 3° - 8°. When compensating for height on an uneven section, the four hydraulic cylinders independently control the height, the response time is less than 200 milliseconds, and the maximum compensation height is 15 cm. The attitude adjustment uses a fuzzy PID control algorithm to improve the system response speed and stability. The fuzzy rule base contains 25 rules, covering common terrain scenarios. The attitude adjustment goal is to keep the levelness of the inspection equipment platform within ±3° to ensure the quality of inspection data. This step realizes the intelligent adjustment of the vehicle body attitude and ensures the stable operation of the rail vehicle on complex terrains.

[0072] The specific implementation of step S11 is the inspection execution process. First, the inspection equipment device is controlled to collect data of the photovoltaic modules. The high-definition camera collects visible light images at a frequency of 5 Hz, the resolution is set to 4K, and the field of view covers the entire photovoltaic module. The infrared thermal imager collects thermal imaging maps at a frequency of 2 Hz, the temperature range is set to 20 - 100 °C, and the sensitivity is 0.05 °C. The ultraviolet detector collects discharge data at a frequency of 1 Hz. The electrical parameter measuring instrument measures voltage, current, and power data at a frequency of 0.5 Hz. Then, the environmental parameters are collected. The temperature, humidity, light intensity, and air pressure data are recorded through the environmental monitoring device, and the collection frequency is 1 time per minute. The environmental parameters are used for data calibration and anomaly judgment. Next, the collected data is stored in the data storage device in real time. The data is indexed and stored according to the timestamp and geographical location, and a hierarchical storage strategy is adopted. The hot data is stored in the cache, and the historical data is stored in the solid-state drive. The storage format uses a compressed binary format to reduce the storage space requirement. During the inspection execution process, the system monitors the device status in real time to ensure the quality of data collection. When the light intensity is lower than 100 lux or the temperature exceeds 60 °C, the camera parameters are automatically adjusted or the protection mode is entered. This step realizes the comprehensive collection of photovoltaic module data and provides the original data for fault detection.

[0073] The specific implementation of step S12 is the data preprocessing process. First, the collected inspection data is preprocessed. For image denoising, the wavelet transform method is used. Appropriate wavelet bases are selected according to different noise types. For Gaussian noise, the DB4 wavelet is used, and for salt-and-pepper noise, the Haar wavelet is used. The denoising threshold is adaptively determined by the empirical Bayesian method. For image enhancement, a combination method of histogram equalization and contrast-limited adaptive histogram equalization is used, and the enhancement parameters are automatically adjusted according to the image histogram features. For image registration, a feature point matching algorithm is used. SIFT feature points are extracted, and abnormal matching points are removed through the RANSAC algorithm. The registration accuracy is controlled within 1 pixel. For image segmentation, an improved maximum inter-class variance method is used, and the segmentation accuracy is improved by combining the edge detection results. Then, the thermal image temperature correction is carried out. The infrared image is corrected based on the ambient temperature and the reflected temperature. A correction model is established using the blackbody radiation theory, and the correction accuracy is ±1°C. Next, the electrical parameters are normalized. The measured voltage and current are standardized to the standard values under unit light intensity for easy comparison of data at different time points. Finally, the environmental parameter compensation is carried out. A compensation model is established based on the ambient temperature, humidity, and light intensity to reduce the influence of environmental factors on the detection results. This step improves the quality of the inspection data and provides an accurate data basis for subsequent fault detection.

[0074] The specific implementation of step S13 is the fault detection process. First, based on the preprocessed data, a pre-trained neural network model for multi-modal defect detection of photovoltaic modules is called for fault detection. This model is a multi-modal fusion network based on the transformer architecture, including four parallel modality-specific encoders and a cross-modal fusion decoder. The visible light image encoder uses an improved vision transformer structure with 12 layers of attention mechanisms to extract defect visual features. The infrared thermal image encoder has 8 layers of attention mechanisms, focusing on the recognition of thermal anomaly patterns. The ultraviolet image encoder has 6 layers of attention mechanisms to enhance the extraction of discharge features. The electrical parameter encoder uses a 4-layer MLP structure to extract electrical features. Then, multi-source data fusion is achieved through the cross-modal fusion decoder. A hierarchical attention mechanism is used. First, coarse-grained alignment between modalities is carried out, and then fine-grained feature fusion is achieved through a sparse attention mechanism. Next, various common photovoltaic module defects are identified, including hot spots, hidden cracks, virtual soldering, PID, aging, and occlusion, etc. The fault classification accuracy reaches over 95%, and the position localization accuracy is less than 5 cm. Then, the fault severity assessment is carried out. The faults are divided into three levels: minor, medium, and severe. The assessment basis includes the degree of temperature anomaly, the proportion of the defect area, and the decline range of electrical performance. For example, if the hot spot temperature exceeds the surrounding normal temperature by 15°C, it is determined as a severe fault; if the area proportion exceeds 5%, it is determined as a severe fault; if the power loss exceeds 10%, it is determined as a severe fault. Finally, a fault report is generated, including the fault type, location, severity, and repair suggestions. This step realizes the intelligent detection of photovoltaic module faults and provides a decision-making basis for power station maintenance.

[0075] The specific implementation of step S14 is the data transmission process. First, the inspection results, fault information, and system status data are transmitted to the remote control center through the communication transmission device. The transmission is based on the 5G communication network and uses the TCP / IP protocol to ensure reliable data transmission. Then, the transmission strategy is adaptively adjusted according to the network status. The network status monitoring includes signal strength, delay time, and packet loss rate measurement. The signal strength is represented by the RSSI value, and the good threshold is above -85dBm. The good threshold for the delay time is below 100 milliseconds, and the good threshold for the packet loss rate is below 1%. Next, when the network is good, all data is transmitted, including the original image, thermal imaging map, fault detection results, and system status data, and the data transmission rate can reach 100Mbps. When the network is unstable, key data is transmitted, only the fault detection results and low-resolution images are transmitted, and the data volume is reduced to 10% - 20% of the original to ensure the timely transmission of key information. The AES-256 encryption algorithm is used to ensure data security during the data transmission process, and the HEVC algorithm is used for compression to reduce bandwidth occupancy. This step realizes the reliable transmission of inspection data and ensures that the control center can grasp the power station situation in real time.

[0076] The specific implementation of step S15 is the energy management process. First, the battery power is monitored through the power management device with a sampling frequency of once per second, and the monitored parameters include battery voltage, current, temperature, and state of charge. Then, according to the remaining inspection tasks and battery status, the system operation parameters are dynamically adjusted. The remaining task volume is estimated by calculating the length of the uncompleted path and the estimated inspection time, and the battery status assessment is based on the remaining battery percentage and battery health status. Next, the energy utilization strategy is optimized. When the battery power is sufficient (greater than 70%), the normal operation parameters are maintained. When the battery power is medium (30% - 70%), the energy-saving mode is started, the inspection speed is reduced to 80% of the normal speed, the sampling frequency of non-critical sensors is reduced, and some auxiliary functions are turned off. When the battery power is low (below 30%), the deep energy-saving mode is started, the inspection speed is reduced to 60% of the normal speed, only the core functions are retained, and the key inspection points are preferentially completed. Energy management also includes regenerative braking energy recovery. During the downhill braking process, the kinetic energy is converted into electrical energy and stored in the battery, and the recovery efficiency is not less than 60%. This step realizes the intelligent management of energy, extends the inspection time, and ensures the completion of key tasks.

[0077] The specific implementation of step S16 is an exception handling process. First, the system operating status is monitored, including hardware status, software operating status and environmental status monitoring. Hardware monitoring parameters include physical parameters such as temperature, current, and speed. Software monitoring includes memory usage, CPU load and program operating status. Environmental monitoring includes temperature, humidity and light intensity. Then detect abnormal conditions. The hardware failure judgment criteria include temperature exceeding the threshold (CPU temperature exceeds 85°C, motor temperature exceeds 80°C), current abnormality (exceeding the rated value by 120%), and communication interruption (no response for more than 500 milliseconds); software abnormality judgment criteria include program crash, memory leak (memory usage continues to grow for more than 10 minutes), response timeout (processing time exceeds expected 200%); environmental abnormality judgment criteria include temperature exceeding the working range (below -20°C or above 60°C), and light intensity is too low (below 50lux). Then, the corresponding processing strategy is executed according to the exception type. The automatic recovery strategy is for temporary software exceptions, and the normal function is restored by restarting the abnormal module; the downgrade operation strategy is for some hardware failures, shutting down the faulty module and keeping the core functions running; the safe shutdown strategy is for serious failures, and the control system stops safely and sends an emergency notification. Detailed logs are recorded during the exception handling process, including the exception time, type, on-site parameters and processing measures, to provide a basis for subsequent analysis. This step improves the robustness of the system and ensures safe and reliable operation even under abnormal conditions.

[0078] The specific implementation method of step S17 is the task completion process. First, after the inspection task is completed, an inspection report is generated. The report content includes inspection path trajectory, inspection time statistics, fault distribution statistics and system status records. The inspection path trajectory is visualized through a map, marking the inspection point location and fault point location; inspection time statistics include total inspection time, inspection time proportion of each area and residence time analysis; fault statistics include fault type distribution, fault severity distribution and spatial distribution heat map; system status records include energy consumption curves, equipment operation status and abnormal event records. Then the report is sent to the remote control center through a communication transmission device. The report adopts a standardized format to facilitate automated processing and historical comparative analysis. Then the track walking device is controlled to return to the starting position. The return path is generated by a path planning algorithm, considering the shortest path and energy efficiency. The core monitoring function is maintained during the return process to reduce the power consumption of non-essential functions. Finally, the system enters the standby state, turns off the high-energy consumption module, maintains the low-power monitoring and communication functions, and waits for the next task instruction. This step completes the closed loop of the inspection task, provides comprehensive inspection results and ensures that the system returns to the ready state.

[0079] The mathematical model or calculation process involved in the present invention is described in detail below.

[0080] In step S03, the path planning process involves an improved shortest path algorithm and energy consumption model calculation. The path planning uses an improved Dijkstra's algorithm, which introduces a heuristic function to reduce the search space. Specifically, it is expressed as follows:

[0081] f(n) = g(n) + h(n);

[0082] In the formula, f(n) is the evaluation function value of node n; g(n) is the actual path cost from the starting point to node n; h(n) is the estimated cost from node n to the target point (heuristic function).

[0083] The design of the heuristic function takes into account the straight-line distance and slope factors, and is specifically expressed as:

[0084] h(n) = α·d(n, goal) + β·s(n, goal);

[0085] In the formula, d(n, goal) is the Euclidean distance from node n to the target point; s(n, goal) is the estimated slope cost from node n to the target point; α and β are weight coefficients, which are adjusted according to the actual application scenario. Usually, the value range of α is 0.7 - 0.9, and the value range of β is 0.1 - 0.3.

[0086] The energy consumption model is obtained by training from historical operation data using the multiple regression method, and is specifically expressed as:

[0087]

[0088] In the formula, E is the estimated total energy consumption value (unit: Wh); l i is the length of path segment i (unit: m); s i is the slope of path segment i (unit: degree); v i is the average speed of path segment i (unit: m / s); p i is the device power of path segment i (unit: W); a1, a2, a3, a4, a5 are regression coefficients obtained by fitting historical data; ε is the error term, considering the influence of environmental factors and equipment state fluctuations, and usually ranges from ±5% of the total energy consumption.

[0089] The method for obtaining the parameters in the energy consumption model is: l i Obtained through the track map database; s i Obtained by measuring through the inclination sensor of the terrain adaptation device; v i Set by the mission planning, usually in the range of 1 - 3 m / s; p iReal-time measurement is carried out through the power monitoring system of the power management device; the regression coefficient is obtained by fitting the historical operation data using the least squares method. The reason for using the square term of the slope in this formula is that the energy consumption has a non-linear relationship with the slope. When going uphill, the energy consumption increases sharply with the increase of the slope, while when going downhill, part of the energy can be recovered through regenerative braking.

[0090] In step S04, starting the movement process involves an S-curve acceleration profile and a PID control algorithm. The S-curve acceleration profile is expressed as:

[0091]

[0092] where a(t) is the acceleration at time t (unit: m / s 2 ); a max is the maximum acceleration, set to 2 m / s 2 ; t1 is the end time point of the positive acceleration phase; t2 is the end time point of the constant acceleration phase; t3 is the end time point of the negative acceleration phase. The parameters t1, t2, and t3 are calculated based on the target speed and the jerk limit, and the jerk limit is 0.3 m / s 3 . The sine-squared function is used to ensure a smooth transition of the acceleration curve at the junctions of each phase and avoid shocks caused by sudden changes in acceleration.

[0093] The PID control algorithm is used for speed control and is expressed as:

[0094]

[0095] where u(t) is the control output signal; e(t) is the speed error, that is, the difference between the target speed and the actual speed; K p is the proportional coefficient, set to 3.5; K i is the integral coefficient, set to 0.8; K d is the derivative coefficient, set to 0.2. The PID parameters can be adaptively adjusted according to the load conditions, and the adjustment method is based on fuzzy logic control:

[0096] K p ′ = K p ·(1 + δ p ·μ(e, Δe));

[0097] K i ′ = K i ·(1 + δ i ·μ(e, Δe));

[0098] K d ′ = K d ·(1 + δ d ·μ(e, Δe));

[0099] Where K p ′, K i ′, K d ′ are the adjusted PID parameters; δ p , δ i , δ d are the adjustment amplitude coefficients, with a range of ±0.5; μ(e, Δe) is the fuzzy membership function, with the input being the error e and the error change rate Δe, and the output range being [-1, 1].

[0100] In step S05, the obstacle avoidance detection process involves a multi-sensor data fusion algorithm, using Kalman filtering and evidence theory. The Kalman filter state equation is expressed as:

[0101] x k = A·x k-1 + B·u k + w k ;

[0102] The measurement equation is expressed as:

[0103] z k = H·x k + v k ;

[0104] Where x k is the state vector at time k, including the obstacle position and velocity, x k = [x, y, z, v x , v y , v z T ; A is the state transition matrix; B is the control input matrix; u k is the control vector; w k is the process noise, assumed to be Gaussian white noise with a mean of zero and a covariance of Q; z k is the measurement vector, including the measurement results of each sensor; H is the measurement matrix; v k is the measurement noise, assumed to be Gaussian white noise with a mean of zero and a covariance of R.

[0105] For the uniform motion assumption, the state transition matrix A is expressed as:

[0106]

[0107] Where Δt is the sampling time interval, related to the sensor sampling frequency, and is usually set to 0.02 seconds.

[0108] Evidence theory is used to fuse data from different sensors, and the measurement credibility assignment function is:

[0109] ​

[0110] where m i (A) is the belief assignment of sensor i for the presence of an obstacle; α i is the basic credibility of sensor i, 0.8 for lidar, 0.7 for ultrasonic, and 0.6 for vision; β i is the attenuation coefficient; d i is the measured distance in meters.

[0111] Multi-sensor fusion uses the Dempster combination rule:

[0112]

[0113] where m 1，2 (A) is the belief assignment for the presence of an obstacle after the information fusion of sensor 1 and sensor 2; B and C are the recognition results of sensor 1 and sensor 2 respectively; the denominator term is the conflict factor, indicating the degree of inconsistency between sensors. The evidence theory is adopted because it can effectively handle the uncertainty and conflict of sensor data and is particularly suitable for multi-sensor fusion scenarios.

[0114] In step S07, the obstacle avoidance decision-making process involves a multi-objective obstacle avoidance path planning function. This function combines the Dijkstra algorithm and the A* algorithm, where the heuristic function of the A* algorithm is designed as:

[0115] h(n) = γ1·d euc (n, goal) + γ2·d turn (n);

[0116] where h(n) is the heuristic function value of node n; d euc (n, goal) is the Euclidean distance from node n to the target point; d turn (n) is the turning cost, related to the angle between the current path direction and the direction from node n to the target point; γ1 and γ2 are weight coefficients, γ1 is usually set to 1.0, and γ2 is set to 0.3.

[0117] Multi-objective optimization uses the weighted sum method, and the objective function is:

[0118] F(p) = w1·f len (p) + w2·f safe (p) + w3·f energy (p) + w4·f time (p) + w5·f stab (p);

[0119] where F(p) is the comprehensive evaluation function of path p; f len (p) is the path length objective function, calculating the total path length; f safe(p) is the safety distance objective function, calculating the minimum distance between the path and the obstacle; f energy (p) is the energy consumption objective function, calculating the path energy consumption based on the energy consumption model; f time (p) is the time efficiency objective function, calculating the total path time; f stab (p) is the equipment stability objective function, related to the path curvature and slope changes; w1, w2, w3, w4, w5 are weight coefficients, adjusted through an adaptive mechanism.

[0120] The weight adaptive adjustment mechanism is expressed as:

[0121]

[0122] In the formula, w i is the adjusted weight; is the initial weight; λ i is the adjustment coefficient; φ i is the influence factor. Different objectives correspond to different influence factors, such as battery power, obstacle proximity, and task urgency, etc. For example, φ3 (energy consumption influence factor) is related to the remaining battery charge SOC:

[0123]

[0124] This adaptive weight adjustment mechanism can dynamically adjust the importance of the optimization objectives according to the system state and environment, achieving more flexible decision-making.

[0125] In step S10, the attitude adjustment process adopts the fuzzy PID control algorithm. The fuzzy PID control error input is expressed as:

[0126] e(t) = θ target -θ actual ;

[0127] Δe(t) = e(t) - e(t - 1);

[0128] In the formula, e(t) is the angle error; θ target is the target angle; θ actual is the actual angle; Δe(t) is the error change rate.

[0129] The fuzzy control output calculation is:

[0130]

[0131] In the formula, ΔK p 、ΔK i 、ΔK d are the adjustment amounts of the PID parameters; μ i (e, Δe) is the activation degree of the i-th fuzzy rule; ΔK pi, ΔK ii , ΔK di is the PID parameter adjustment amount corresponding to the i-th rule. The fuzzy rule base contains 25 rules, covering common terrain scenarios.

[0132] The final PID parameters are:

[0133] K p ' = K p + ΔK p ;

[0134] K i ' = K i + ΔK i ;

[0135] K d ' = K d + ΔK d ;

[0136] In the formula, K p ', K i ', K d ' are the adjusted PID parameters; K p , K i , K d are the initial PID parameters. The fuzzy PID control algorithm combines the accuracy of traditional PID control and the intelligence of fuzzy control, and can better handle nonlinear control systems.

[0137] The attitude adjustment calculates the target angle according to the terrain type:

[0138]

[0139] In the formula, θ target is the target angle; θ slope is the slope angle; k up is the uphill compensation coefficient, with a range of 0.75 to 0.85; k down is the downhill compensation coefficient, with a range of 0.75 to 0.85; k turn is the turning compensation coefficient; v is the current speed; r is the radius of curvature. This formula takes into account the optimal attitude adjustment strategy under different terrain conditions, partially compensates for the slope angle during uphill and downhill to keep the platform relatively level; during turning, the roll angle is proportional to the centrifugal force to balance the lateral force during turning.

[0140] In step S12, the data preprocessing process involves using the wavelet transform method for image denoising, which is specifically expressed as follows:

[0141] f(x, y) = ∑ j，k α j，k ψ j，k (x, y);

[0142] where \(f(x, y)\) is the image function; \(\psi\) j，k (x, y) is the wavelet basis function; \(\alpha\) j，k is the wavelet coefficient; \(j\) and \(k\) are the scale and position parameters respectively.

[0143] The threshold selection for wavelet denoising adopts the empirical Bayes method:

[0144]

[0145] where \(T\) j is the threshold at the \(j\)-th scale; \(\sigma\) is the estimated value of the noise standard deviation; \(N\) is the number of image pixels; is the variance of the wavelet coefficients at the \(j\)-th scale. This method can adaptively select an appropriate threshold to balance the denoising effect and detail retention.

[0146] In step S13, the decision function of the pre-trained neural network model for multi-modal defect detection of photovoltaic modules can be expressed as:

[0147] \(P(c|X)=\text{softmax}(W\) c \cdot F(X)+b\) c );

[0148] where \(P(c|X)\) is the probability that the input data \(X\) belongs to the defect category \(c\); \(F(X)\) is the multi-modal fusion feature; \(W\) c and \(b\) c are the weights and biases of the classification layer; \(\text{softmax}\) is the normalization function. The defect location is achieved through region prediction:

[0149] \(B(X)=W\) b \cdot F(X)+b\) b ;

[0150] where \(B(X)\) is the predicted value of the bounding box coordinates, including the center point coordinates and width and height; \(W\) b and \(b\) b are the weights and biases of the localization layer.

[0151] In step S15, the energy management process involves the estimation of the remaining battery usage time:

[0152]

[0153] where \(T\) rmain is the estimated remaining usage time (unit: hour); SOC is the state of charge of the battery (range 0 - 1); \(C\) batt is the rated capacity of the battery (unit: Ah); \(V\) batt is the rated voltage of the battery (unit: V); \(\eta\) is the energy conversion efficiency, usually 0.85 - 0.95; \(P\) avgis the system average power consumption (unit: W).

[0154] The average power consumption calculation is based on historical data and the current operating state:

[0155] P avg = α·P hist +(1 - α)·P curr ;

[0156] In the formula, P hist is the historical average power consumption; P curr is the current power consumption; α is the smoothing coefficient, usually taking a value of 0.7 - 0.8. This calculation method takes into account historical data and the current state, improving the accuracy of the remaining time estimation.

[0157] The regenerative braking energy recovery efficiency calculation is:

[0158] E recovery = η recovery ·m·h·sinθ;

[0159] In the formula, E recovery is the recoverable energy (unit: J); η recovery is the recovery efficiency, usually 0.6 - 0.7; m is the system mass (unit: kg); g is the acceleration due to gravity (unit: m / s 2 ); h is the height change (unit: m); θ is the downhill angle. This formula is based on the principle of gravitational potential energy conversion to calculate the recoverable energy during downhill.

[0160] Specifically, the principle of the present invention is:

[0161] The core technical principle of the present invention lies in constructing an intelligent inspection framework for multi - system collaborative work, and realizing efficient inspection in complex environments through unified scheduling of each subsystem by a control chip. In terms of obstacle handling, the system adopts the closed - loop control principle of "perception - decision - execution". A multi - dimensional perception network is constructed through lidar, ultrasonic sensors and visual cameras to form a comprehensive perception of the environment; the multi - objective obstacle avoidance path planning function combining improved Dijkstra and A* algorithms is used to achieve accurate classification of obstacles and optimized selection of obstacle avoidance strategies; finally, the obstacle avoidance action is executed through the track walking device to ensure the continuity and safety of the inspection process.

[0162] In terms of terrain adaptation, the system is based on the principle of hydraulic drive and closed - loop feedback control. The track slope is detected in real time through a terrain sensor, and the hydraulic lifting system and attitude adjustment mechanism are driven to dynamically adjust the vehicle body attitude to ensure the stability of the inspection equipment on changing terrains. Among them, the precise control of the pressure and displacement of the hydraulic system is the key to realizing attitude adjustment.

[0163] In terms of defect detection, the system adopts the principle of combining multi-modal data fusion and deep learning. Through a pre-trained neural network model based on the Transformer architecture, it realizes the multi-source fusion analysis of visible light, infrared, ultraviolet, and electrical parameter data, significantly improving the accuracy of defect recognition. The model uses a hierarchical attention mechanism for multi-modal feature extraction and fusion, and realizes information complementation and enhancement between modalities through contrastive learning, overcoming the limitations of a single modality.

[0164] In addition, the system adopts the principle of modular design and intelligent decision-making to achieve the decomposition and collaborative processing of complex tasks, forming a complete technical system in aspects such as energy management, data processing, and communication transmission, enabling the rail vehicle to achieve efficient and stable autonomous inspection in complex environments.

[0165] The following provides a specific Embodiment 1 of the present invention. The specific implementation manners of each step in this Embodiment 1 are described in detail as follows.

[0166] The specific implementation manners of Steps S01 - S02 are the same as those described above and will not be elaborated here.

[0167] The specific implementation manner of Step S03 is the path planning process. First, path planning is carried out based on a preset track map and task requirements. The preset track map is stored in a graph data structure, where nodes represent track intersections and edges represent track segments. Each edge contains attributes such as length, slope, and curvature. Then, an improved shortest path algorithm is used for global path planning. This algorithm is optimized based on Dijkstra's algorithm and introduces a heuristic function to reduce the search space. The evaluation function is expressed as:

[0168] f(n) = g(n) + h(n); where f(n) is the evaluation function value of node n; g(n) is the actual path cost from the starting point to node n; h(n) is the estimated cost from node n to the target point (heuristic function). The heuristic function design takes into account the straight-line distance and slope factors and is specifically expressed as: h(n) = α·d(n, goal) + β·s(n, goal); where d(n, goal) is the Euclidean distance from node n to the target point; s(n, goal) is the estimated slope cost from node n to the target point; α and β are weight coefficients, which are adjusted according to the actual application scenario. Usually, the value range of α is 0.7 - 0.9, and the value range of β is 0.1 - 0.3. Then, the path length, estimated inspection time, and energy consumption are calculated. The path length is obtained by accumulating the lengths of each edge on the selected path. The inspection time is calculated based on the path length, inspection speed, and estimated stay time. The energy consumption is estimated through an energy consumption model, which takes the path slope, inspection speed, and equipment working power as inputs and outputs the estimated energy consumption value. The energy consumption model is expressed as: where E is the estimated total energy consumption value (unit: Wh); l iis the length of path segment i (unit: m); s i is the slope of path segment i (unit: degree); v i is the average speed of path segment i (unit: m / s); p i is the equipment power of path segment i (unit: W); a1, a2, a3, a4, a5 are regression coefficients obtained by fitting historical data; ε is the error term, considering the influence of environmental factors and equipment state fluctuations, usually in the range of ±5% of the total energy consumption. The reason for using the square term of the slope in the energy consumption model is that the energy consumption has a non-linear relationship with the slope. When going uphill, the energy consumption increases sharply with the increase of the slope, and when going downhill, part of the energy can be recovered through regenerative braking. Finally, the planning results are displayed through a man-machine interaction device, including path visualization, time estimation, and energy consumption prediction. The planning thresholds include that the length of a single inspection path does not exceed 10 kilometers, the expected inspection time does not exceed 6 hours, and the expected energy consumption does not exceed 80% of the battery capacity. This step provides an optimized path planning for the inspection task to ensure the efficient completion of the task and the reasonable use of energy.

[0169] The specific implementation of step S04 is to start the moving process. First, control the drive motor of the track walking device to start according to the planned path. The motor control adopts a vector control algorithm to achieve precise control of torque and speed. Then, achieve smooth acceleration. Use an S-curve acceleration profile, with the initial acceleration set to 0.5 m / s 2 , and the maximum acceleration does not exceed 2 m / s 2 , and the jerk is controlled within 0.3 m / s 3 to ensure no impact during the start-up process. The S-curve acceleration profile is expressed as: In the formula, a(t) is the acceleration at time t (unit: m / s 2 ); a max is the maximum acceleration, set to 2 m / s 2 ; t1 is the end time point of the jerk-up phase; t2 is the end time point of the constant acceleration phase; t3 is the end time point of the jerk-down phase. The parameters t1, t2, and t3 are calculated according to the target speed and jerk limit, and the jerk limit is 0.3 m / s 3 . The sine-squared function is used to ensure a smooth transition of the acceleration curve at the boundaries of each phase and avoid impacts caused by sudden changes in acceleration. Then, gradually increase to the inspection speed. The inspection speed is set according to the task requirements, usually in the range of 1 - 3 m / s. The speed control adopts a PID control algorithm, with a proportional coefficient of 3.5, an integral coefficient of 0.8, and a derivative coefficient of 0.2. The parameters can be adaptively adjusted according to the load conditions. The PID control algorithm is expressed as: In the formula, u(t) is the control output signal; e(t) is the speed error, that is, the difference between the target speed and the actual speed; K pis the proportionality coefficient, set to 3.5; K i is the integral coefficient, set to 0.8; K d is the derivative coefficient, set to 0.2. The PID parameters can be adaptively adjusted according to the load conditions, and the adjustment method is based on fuzzy logic control: K p ′ = K p ·(1 + δ p ·μ(e, Δe)); K i ′ = K i ·(1 + δ i ·μ(e, Δe)); K d ′ = K d ·(1 + δ d ·μ(e, Δe)); In the formula, K p ′, K i ′, K d ′ are the adjusted PID parameters; δ p , δ i , δ d are the adjustment amplitude coefficients, with a range of ±0.5; μ(e, Δe) is the fuzzy membership function, with the input being the error e and the error change rate Δe, and the output range being [-1, 1]. At the same time, monitor the motor current, speed, and temperature parameters. The current sampling frequency is 100 Hz, the speed sampling frequency is 50 Hz, and the temperature sampling frequency is 10 Hz. The collected data is subjected to noise suppression through the Kalman filter algorithm. The monitoring thresholds include that the motor current does not exceed 85% of the rated current, and the motor temperature does not exceed 75 °C. Once the threshold is exceeded, immediately reduce the acceleration or start the heat dissipation measures. This step ensures a smooth and controllable start-up process of the rail vehicle, protects the equipment, and extends its service life.

[0170] The specific implementation of step S05 is the obstacle avoidance detection process. First, collect the surrounding environment data through the obstacle avoidance sensing device. The lidar scans 360° at a frequency of 50 Hz, the ultrasonic sensor measures the distance at a frequency of 25 Hz, and the vision camera collects images at a frequency of 20 Hz. Then, fuse the multi-sensor data. Use the multi-sensor data fusion algorithm to unify the data from different sensors into the same coordinate system. The fusion algorithm is based on the Kalman filter and the evidence theory. The Kalman filter state equation is expressed as: x k = A·x k-1 + B·u k + w k ; The measurement equation is expressed as: z k = H·x k + v k ; In the formula, x k is the state vector at time k, including the obstacle position and speed, x k = [x, y, z, v x , v y , vz T ; A is the state transition matrix; B is the control input matrix; u k is the control vector; w k is the process noise, assumed to be Gaussian white noise with zero mean and covariance Q; z k is the measurement vector, containing the measurement results of each sensor; H is the measurement matrix; v k is the measurement noise, assumed to be Gaussian white noise with zero mean and covariance R. For the uniform motion assumption, the state transition matrix A is expressed as: where Δt is the sampling time interval, related to the sensor sampling frequency, usually set to 0.02 seconds. The evidence theory is used to fuse data from different sensors, and the measurement credibility assignment function is: where m i (A) is the credibility assignment of sensor i to the existence of an obstacle; α i is the basic credibility of sensor i, 0.8 for lidar, 0.7 for ultrasonic, and 0.6 for vision; β i is the attenuation coefficient; d i is the measured distance, in meters. The multi-sensor fusion adopts the Dempster combination rule: where m 1，2 (A) is the credibility assignment of the fused information of sensor 1 and sensor 2 to the existence of an obstacle; B and C are the recognition results of sensor 1 and sensor 2 respectively; the denominator term is the conflict factor, indicating the degree of inconsistency between sensors. Then a three-dimensional environment model is constructed, using an octree structure to represent the environment, with a spatial resolution of 10 cm and a model update frequency of 10 Hz. The model construction adopts the probabilistic occupancy grid method to reduce the influence of sensor noise. Then, obstacles on and around the track are identified. The obstacle identification adopts an object detection algorithm based on deep learning, which runs on the hardware acceleration unit of the track vehicle control chip and can identify common obstacle types such as people, vehicles, and animals in real time. Finally, the position, size, and motion characteristics of the obstacles are calculated. The position accuracy is controlled within ±5 cm. The speed and direction information of the obstacle motion are obtained by tracking consecutive frames, and the Kalman prediction algorithm is used to predict the future position of the obstacle. This step provides comprehensive and accurate environmental perception data for subsequent obstacle avoidance decisions.

[0171] The specific implementation of step S06 is the same as the foregoing and will not be elaborated here.

[0172] ​The specific implementation of step S07 is an obstacle avoidance decision-making process. First, based on the obstacle classification result, a multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle is called for obstacle avoidance path planning. This function combines the advantages of Dijkstra's algorithm and the A* algorithm. First, Dijkstra's algorithm is used for global path search to determine the general obstacle avoidance direction. The algorithm sets the inflation distance parameter to 50 cm to ensure that the path maintains a safe distance from the obstacles. Then, local path optimization is carried out through the A* algorithm. The heuristic function of the A* algorithm is designed as: h(n) = γ1·d euc (n, goal) + γ2·d turn (n); where h(n) is the heuristic function value of node n; d euc (n, goal) is the Euclidean distance from node n to the target point; d turn (n) is the turning cost, which is related to the angle between the current path direction and the direction from node n to the target point; γ1 and γ2 are weight coefficients, γ1 is usually set to 1.0, and γ2 is set to 0.3. This function simultaneously considers multiple optimization goals, including minimizing the path length, maximizing the safe distance from obstacles, minimizing energy consumption, maximizing time efficiency, and maximizing equipment stability. Multi-objective optimization uses the weighted sum method, and the objective function is: F(p) = w1·f len (p) + w2·f safe (p) + w3·f energy (p) + w4·f time (p) + w5·f stab (p); where F(p) is the comprehensive evaluation function of path p; f len (p) is the path length objective function, which calculates the total path length; f safe (p) is the safe distance objective function, which calculates the minimum distance between the path and the obstacles; f energy (p) is the energy consumption objective function, which calculates the path energy consumption based on the energy consumption model; f time (p) is the time efficiency objective function, which calculates the total path time; f stab (p) is the equipment stability objective function, which is related to the path curvature and slope change; w1, w2, w3, w4, w5 are weight coefficients, which are adjusted through an adaptive mechanism. The weight adaptive adjustment mechanism is expressed as: where w i is the adjusted weight; is the initial weight; λ i is the adjustment coefficient; φ i is the influence factor. Different goals correspond to different influence factors, such as battery power, obstacle proximity, and task urgency. For example, φ3 (energy consumption influence factor) is related to the remaining battery charge SOC: This adaptive weight adjustment mechanism can dynamically adjust the importance of the optimization objective according to the system state and environment, achieving more flexible decision-making. The obstacle avoidance planning results include a sequence of path points, a speed curve, and attitude adjustment parameters. The distance between path points is no more than 0.5 meters to ensure a smooth and continuous path. This step realizes intelligent obstacle avoidance decision-making in complex environments, balancing the requirements of safety and efficiency.

[0173] The specific implementation manners of steps S08 - S09 are the same as those described above and will not be elaborated here.

[0174] The specific implementation manner of step S10 is the attitude adjustment process. First, based on the terrain detection results, the hydraulic lifting system and the attitude adjustment mechanism of the terrain adaptation device are controlled to adjust the vehicle body attitude in real time. The attitude adjustment calculates the target angle according to the terrain type: In the formula, θ target is the target angle; θ slope is the slope angle; k up is the uphill compensation coefficient, with a range of 0.75 - 0.85; k down is the downhill compensation coefficient, with a range of 0.75 - 0.85; k turn is the turning compensation coefficient; v is the current speed; r is the radius of curvature. This formula considers the optimal attitude adjustment strategy under different terrain conditions. When going uphill or downhill, it partially compensates for the slope angle to keep the platform relatively horizontal; when turning, the roll angle is proportional to the centrifugal force to balance the lateral force during turning. When adjusting the forward tilt on the uphill section, the front hydraulic cylinder contracts and the rear hydraulic cylinder extends. The forward tilt angle is calculated according to the slope, usually 75% - 85% of the slope angle, and the angle control accuracy is ±0.5°; when adjusting the rear tilt on the downhill section, the front hydraulic cylinder extends and the rear hydraulic cylinder contracts, and the rear tilt angle is also 75% - 85% of the slope angle; when adjusting the roll on the turning section, the inner hydraulic cylinder extends and the outer hydraulic cylinder contracts, and the roll angle is calculated according to the radius of curvature, and the angle is usually in the range of 3° - 8°; when compensating for the height on the uneven section, the four hydraulic cylinders independently control the height, with a response time of less than 200 milliseconds and a maximum compensation height of 15 centimeters. The attitude adjustment adopts a fuzzy PID control algorithm to improve the system response speed and stability. The fuzzy rule base contains 25 rules, covering common terrain scenarios. The fuzzy PID control error input is expressed as: e(t) = θ target -θ actual ; Δe(t) = e(t) - e(t - 1); In the formula, e(t) is the angle error; θ target is the target angle; θ actual is the actual angle; Δe(t) is the error change rate. The fuzzy control output is calculated as: In the formula, ΔK p 、ΔK i 、ΔK dis the adjustment amount of PID parameters; μ i (e, Δe) is the activation degree of the i-th fuzzy rule; ΔK pi 、ΔK ii 、ΔK di is the PID parameter adjustment amount corresponding to the i-th rule. The final PID parameters are: K p ′ = K p + ΔK p ; K i ′ = K i + ΔK i ; K d ′ = K d + ΔK d ; In the formula, K p ′, K i ′, K d ′ are the adjusted PID parameters; K p , K i , K d are the initial PID parameters. The attitude adjustment target is to keep the leveling of the inspection equipment platform within ±3°, ensuring the quality of inspection data. This step realizes the intelligent adjustment of the vehicle body attitude, ensuring the stable operation of the rail vehicle on complex terrains.

[0175] The specific implementation manner of step S11 is the same as the foregoing and will not be elaborated herein.

[0176] The specific implementation manner of step S12 is the data preprocessing process. First, the collected inspection data is preprocessed. For image denoising, the wavelet transform method is used. Appropriate wavelet bases are selected according to different noise types. The DB4 wavelet is used for Gaussian noise, and the Haar wavelet is used for salt-and-pepper noise. The denoising threshold is adaptively determined by the empirical Bayes method. The image can be expressed as a linear combination of wavelet coefficients: f(x, y) = ∑ j，k α j，k ψ j，k (x, y); In the formula, f(x, y) is the image function; ψ j，k (x, y) is the wavelet basis function; α j，k is the wavelet coefficient; j and k are the scale and position parameters respectively. The threshold selection for wavelet denoising adopts the empirical Bayes method: In the formula, T j is the threshold of the j-th scale; σ is the estimated value of the noise standard deviation; N is the number of image pixels; is the variance of the wavelet coefficients at the j-th scale. This method can adaptively select an appropriate threshold to balance the denoising effect and detail preservation. Image enhancement uses a combination of histogram equalization and contrast-limited adaptive histogram equalization, and the enhancement parameters are automatically adjusted according to the characteristics of the image histogram; image registration uses a feature point matching algorithm to extract SIFT feature points, and the abnormal matching points are removed through the RANSAC algorithm, and the registration accuracy is controlled within 1 pixel; image segmentation uses an improved maximum inter-class variance method, combined with the edge detection results to improve the segmentation accuracy. Then, thermal image temperature correction is performed. Based on the ambient temperature and reflected temperature, the infrared image is corrected, and a correction model is established using the blackbody radiation theory, and the correction accuracy is ±1°C. Next, electrical parameter normalization is performed to standardize the measured voltage and current to the standard values under unit light intensity, which is convenient for comparing data at different time points. Finally, environmental parameter compensation is performed. Based on the ambient temperature, humidity, and light intensity, a compensation model is established to reduce the influence of environmental factors on the detection results. This step improves the quality of the inspection data and provides an accurate data basis for subsequent fault detection.

[0177] The specific implementation manners of steps S13 - S14 are the same as those described above and will not be elaborated in detail here.

[0178] The specific implementation manner of step S15 is the energy management process. First, the battery power is monitored by the power management device with a sampling frequency of 1 time per second, and the monitored parameters include battery voltage, current, temperature, and state of charge. Then, according to the remaining inspection tasks and battery status, the system operation parameters are dynamically adjusted. Estimation of the remaining battery usage time: In the formula, T remain is the estimated remaining usage time (unit: hour); SOC is the state of charge of the battery (range 0 - 1); C batt is the rated capacity of the battery (unit: Ah); V batt is the rated voltage of the battery (unit: V); η is the energy conversion efficiency, usually 0.85 - 0.95; P avg is the average power consumption of the system (unit: W). The average power consumption is calculated based on historical data and the current operating state: P avg = α·P hist +(1 - α)·P curr ; in the formula, P hist is the historical average power consumption; P curris the current power consumption; α is the smoothing coefficient, usually taking values from 0.7 to 0.8. The remaining task volume is estimated by calculating the length of the uncompleted path and the estimated inspection time, and the battery state assessment is based on the remaining battery percentage and the battery health status. Then, the energy utilization strategy is optimized. When the battery is fully charged (greater than 70%), normal operating parameters are maintained; when the battery level is medium (30% - 70%), the energy-saving mode is activated, the inspection speed is reduced to 80% of the normal speed, the sampling frequency of non-critical sensors is reduced, and some auxiliary functions are turned off; when the battery level is low (less than 30%), the deep energy-saving mode is activated, the inspection speed is reduced to 60% of the normal speed, only the core functions are kept running, and key inspection points are given priority to be completed. Energy management also includes regenerative braking energy recovery, converting kinetic energy into electrical energy during downhill braking and storing it in the battery. The regenerative braking energy recovery efficiency is calculated as: E recovery = η recovery ·m·g·h·sinθ; where E recovery is the recoverable energy (unit: J); η recovery is the recovery efficiency, usually 0.6 - 0.7; m is the system mass (unit: kg); g is the acceleration due to gravity (unit: m / s 2 ); h is the height change (unit: m); θ is the downhill angle. The recovery efficiency is not less than 60%. This step realizes the intelligent management of energy, extends the inspection time and ensures the completion of key tasks.

[0179] The specific implementation manners of steps S16 - S17 are the same as those described above, and will not be elaborated in detail here.

[0180] In this Embodiment 1, the automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station realizes functions of autonomous inspection, automatic obstacle avoidance and adaptive terrain, and fully utilizes advanced control algorithms, data processing methods and artificial intelligence technologies. Path planning adopts an improved shortest path algorithm, fuses heuristic functions to reduce the search space, and optimizes path selection through an energy consumption model; the starting movement process adopts an S-curve acceleration profile and an adaptive PID control algorithm to ensure smooth start and precise speed control; the obstacle avoidance detection process adopts a multi-sensor data fusion algorithm, combines Kalman filtering and evidence theory to accurately perceive the environment; the obstacle avoidance decision-making process adopts a multi-objective optimization algorithm, comprehensively considering path length, safety distance, energy consumption, time efficiency and equipment stability; the attitude adjustment process adopts a fuzzy PID control algorithm to adjust the vehicle body attitude in real time according to terrain characteristics; the data preprocessing process adopts wavelet transform denoising and various image processing technologies to improve data quality; the fault detection process uses a deep learning model to realize multi-modal data fusion analysis; the energy management process extends the battery life through an adaptive strategy and regenerative braking energy recovery. The organic combination of these technologies enables the inspection of the photovoltaic power station to achieve high efficiency, high precision and high reliability, and greatly improves the operation and maintenance level and safety of the photovoltaic power station.

[0181] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 1 and Table 2 below.

[0182] Table 1 Variable Explanation Table

[0183]

[0184]

[0185] Table 2 Variable Explanation Table

[0186]

[0187]

[0188] The above are only the specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. An automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station, characterized in that It includes a control chip, an orbital walking device, an obstacle avoidance sensing device, a terrain adaptation device, an inspection equipment device, a power management device, a data storage device, a communication transmission device, an environmental monitoring device, and a human-machine interaction device; a system control module is provided in the control chip for performing steps of system initialization, receiving inspection tasks, path planning, starting movement, obstacle avoidance detection, obstacle classification, obstacle avoidance decision-making, obstacle avoidance execution, terrain detection, attitude adjustment, inspection execution, data preprocessing, fault detection, data transmission, energy management, exception handling, and task completion; Among them, the obstacle avoidance decision-making step calls the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle, comprehensively considering multi-dimensional objectives such as path length, safety distance, energy consumption assessment, time efficiency, and equipment stability; the fault detection step calls the pre-trained neural network model for multi-modal defect detection of photovoltaic modules to achieve multi-source data fusion analysis and accurate detection and positioning of photovoltaic module defects.

2. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 1, wherein The control chip uses an industrial-grade 64-bit quad-core processor with a main frequency of not less than 2.5 GHz, an 8GB high-speed cache built-in, a hardware acceleration unit for image processing and deep learning algorithm operations, supports multi-thread parallel processing, has a real-time operating system built-in, has a temperature compensation function, can operate stably within the temperature range of -40°C to 85°C, the voltage input range is 9V to 36V, has over-current protection and over-heat protection functions, and supports multiple communication interfaces.

3. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 2, wherein, The orbital walking device includes four groups of walking mechanisms, and each group of walking mechanisms includes a driving wheel, a driven wheel, a driving motor, and a reducer; the driving wheel and the driven wheel use high-strength rubber tires, the tire surface is provided with anti-slip patterns, and steel cord is embedded inside the tire; the driving motor is a brushless DC servo motor; the reducer uses a planetary gear reducer; the orbital walking device also includes an independent suspension system, which adopts a combined structure of a spiral spring and a pneumatic shock absorber.

4. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 3, wherein The obstacle avoidance sensing device includes a lidar, ultrasonic sensors, and vision cameras; the lidar is installed on the top of the vehicle body and has a horizontal scanning angle and a vertical scanning angle; there are 8 ultrasonic sensors in total, which are respectively installed in the front, rear, left, and right directions of the vehicle body; there are 4 vision cameras in total, which are respectively installed in the front, rear, left, and right directions of the vehicle body; the data acquisition frequency of the obstacle avoidance sensing device is 50 times per second.

5. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 4, wherein, The terrain adaptation device includes a hydraulic lifting system, an attitude adjustment mechanism, and a terrain sensor; the hydraulic lifting system consists of four hydraulic cylinders, and each hydraulic cylinder is installed at the four corners of the vehicle body bottom; the attitude adjustment mechanism includes a pitch adjustment component and a roll adjustment component; the terrain sensor includes an inclination sensor and a height sensor; the data acquisition frequency of the terrain adaptation device is 20 times per second.

6. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 5, characterized in that, When the system control module executes the obstacle avoidance decision-making step, it first uses the Dijkstra algorithm for global path planning to determine the general obstacle avoidance direction, and then uses the A* algorithm for local path optimization to find the optimal obstacle avoidance path; when executing the attitude adjustment step, it inclines forward on the uphill section, inclines backward on the downhill section, inclines laterally on the turning section, and compensates for the height on the uneven section to ensure the stable operation of the vehicle body.

7. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 6, wherein The input of the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle includes the obstacle position data set, the obstacle type identifier, the current vehicle position coordinates, the target position coordinates, and the track terrain parameter set; the output is the optimized obstacle avoidance path point sequence, the vehicle speed parameters at each path point, the attitude adjustment parameters, and the energy consumption estimation value; the multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle adopts an improved multi-objective optimization algorithm.

8. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 7, wherein, The multi-objective obstacle avoidance path planning function of the photovoltaic rail vehicle internally implements an obstacle threat assessment mechanism, calculates the comprehensive threat index according to the obstacle type, size, distance, and movement characteristics, and dynamically adjusts the conservatism of the obstacle avoidance strategy; at the same time, it integrates a track topology constraint processing module for identifying feasible paths in complex track networks; it also implements an experience learning mechanism based on historical obstacle avoidance data, and continuously optimizes the decision-making model by accumulating successful obstacle avoidance cases.

9. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 8, wherein The pre-trained neural network model structure for multi-modal defect detection of photovoltaic modules is a multi-modal fusion network based on the Transformer architecture, which includes four parallel modality-specific encoders and a cross-modal fusion decoder; the cross-modal fusion decoder adopts a hierarchical attention mechanism and includes a two-stage fusion strategy. The model output layer includes a defect type classification head and a defect position localization head, supporting the accurate detection and localization of 17 common photovoltaic module defects.

10. The automatic obstacle avoidance and adaptive terrain inspection rail vehicle system for a photovoltaic power station according to claim 9, characterized in that, The pre-training process of the pre-trained neural network model for multi-modal defect detection of photovoltaic modules includes modality-specific pre-training, modality fusion pre-training, end-to-end fine-tuning, knowledge distillation, environment adaptability training, and online learning mechanism design; among them, the visible light image encoder is pre-trained on the ImageNet dataset, the infrared thermal image encoder is pre-trained on the thermal imaging dataset, the ultraviolet image encoder is pre-trained on the discharge detection dataset, and the electrical parameter encoder is pre-trained on the photovoltaic power station operation dataset.

Images