Gas inspection robot hanging rail type fixed point detection system and method

Abstract

The invention relates to a gas inspection robot rail-hanging type fixed point detection system, which belongs to the technical field of gas inspection, and comprises a multi-source fusion positioning module for positioning a gas inspection robot through a rail encoder, a UWB positioning system and a visual identification system, and performing fusion by using an extended Kalman filtering algorithm to realize source fusion positioning; the multi-parameter gas detection array comprises various sensors arranged on the gas inspection robot, and realizes multi-parameter parallel sampling based on a bus synchronous acquisition mechanism; the environment self-adaptive compensation algorithm module is used for correcting the gas concentration in real time through a deep neural network model; the on-line self-calibration device is internally provided with a standard gas microcavity, is connected with the outside and various gas sensor gas chambers through electromagnetic valves, and regularly and automatically calibrates the zero point and gain of the sensor; and the intelligent inspection strategy module dynamically adjusts the inspection plan of the gas inspection robot according to real-time and historical data based on a reinforcement learning algorithm.

Description

Technical Field

[0001] This invention belongs to the field of gas inspection technology, and relates to a gas inspection robot rail-mounted fixed-point detection system and method. Background Technology

[0002] Traditional track-mounted robots rely solely on track encoders for positioning, which is susceptible to track deformation and vibration, resulting in insufficient accuracy in point detection and significant errors in the detection data. Multiple gas sensors use a serial acquisition method, leading to large acquisition delays and difficulty in ensuring synchronous measurement of multiple parameters. Changes in downhole temperature, humidity, dust concentration, and airflow affect the accuracy of sensor readings, and there is a lack of effective compensation for sensor drift and zero-point drift. Sensor calibration requires manual downhole operation, resulting in frequent and costly maintenance, and the sensors lack predictive lifespan capabilities. Furthermore, the inability to intelligently adjust inspection frequency and dwell time based on real-time gas concentration and historical data impacts inspection efficiency. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a rail-mounted fixed-point detection system and method for gas inspection robots.

[0004] To achieve the above objectives, the present invention provides the following technical solution: On one hand, the present invention provides a rail-mounted fixed-point detection system for a gas inspection robot, comprising: The multi-source fusion positioning module locates the gas inspection robot through the track encoder, UWB positioning system and visual recognition system respectively, and uses the extended Kalman filter algorithm to fuse them to achieve multi-source fusion positioning of the gas inspection robot. The multi-parameter gas detection array includes various gas sensors, dust sensors, and temperature and humidity sensors installed on the gas inspection robot. Based on a bus synchronous acquisition mechanism, it realizes parallel sampling of multiple parameters. The environmental adaptive compensation algorithm module uses a deep neural network model to combine temperature, humidity, airflow and dust data to correct gas concentration in real time. The online self-calibration device has a built-in standard gas microcavity filled with zero-point gas and range gas of known concentration. It is connected to the outside world and the gas chambers of various gas sensors through a solenoid valve, and automatically calibrates the zero point and gain of the sensor periodically. The intelligent inspection strategy module, based on reinforcement learning algorithms, dynamically adjusts the inspection plan of the gas inspection robot according to real-time and historical data.

[0005] Furthermore, the track encoder is mounted on the drive wheel of the gas inspection robot to provide continuous mileage information.

[0006] Furthermore, the UWB positioning system includes UWB anchor points deployed at certain intervals in the inspection roadway, and UWB tags set on the gas inspection robot. By measuring the time of flight (TOF) of the wireless signals between the tags and multiple anchor points, the absolute position of the gas inspection robot is calculated.

[0007] Furthermore, the visual recognition system includes an industrial camera installed at the front end of the gas inspection robot, and visual beacons with unique IDs distributed in the inspection lanes. The absolute coordinates of each visual beacon are pre-stored. The industrial camera identifies the visual beacons, calculates their pose in the image, and obtains the absolute position reference point.

[0008] Furthermore, the method of using the extended Kalman filter algorithm for fusion to achieve multi-source fusion localization of the gas inspection robot specifically includes: The extended Kalman filter algorithm is adopted. The state variables are the position and velocity of the gas inspection robot, and the observations are the encoder odometer increment, the absolute position of the UWB, and the absolute position of the visual beacon. Using high-frequency track encoder data as control input, the position and velocity of the gas inspection robot are estimated, as follows:

[0009] in, It is a state variable. It is the encoder reading; When the UWB positioning system calculates new position coordinates, it uses them as observations to update the state estimate and correct the cumulative error of the track encoder. When the visual recognition system identifies a visual beacon, it obtains the pre-stored absolute position of the visual beacon, performs a strong correction, and eliminates the long-term drift that may exist in the UWB positioning system and track encoder.

[0010] Furthermore, the multi-parameter gas detection array generates a global hardware trigger signal based on the FPGA controller, which is simultaneously sent to all sensor interface circuits via a hardware synchronization trigger line, specifically including: During each sampling cycle, the main control computer on the gas inspection robot sends a start sampling command to the FPGA controller. After receiving the instruction, the controller sends a pulse signal to all sensors through the hardware synchronization trigger line; all sensors simultaneously start their internal analog-to-digital conversion on the same edge of the pulse. After sampling is completed, the FPGA controller polls each sensor in a preset order and reads the converted data; each data packet is accompanied by the same synchronization timestamp. The FPGA controller packages all sensor data into a single frame and sends it to the host computer via the RS-485 bus.

[0011] Furthermore, the environmental adaptive compensation algorithm module is based on a multilayer perceptron (MLP) structure, which corrects the gas sensor readings according to real-time environmental parameters to eliminate measurement errors; the multilayer perceptron (MLP) specifically includes: Input layer: The number of nodes is the same as the dimension of the input vector, and the input vector is the raw readings of all sensors; Hidden layers: 3 to 5 fully connected layers, each containing 64 to 256 neurons, using ReLU as the activation function to enhance non-linear fitting ability; Output layer: The number of nodes is the number of gas types that need to be corrected, and the corrected concentration values ​​are output. The training data consisted of standard gases of different concentrations introduced into the laboratory under different temperature, humidity, and dust conditions. Pairs of raw readings and actual concentrations were collected, and the network weights were trained using the backpropagation algorithm until the error between the model output and the actual value was minimized.

[0012] Furthermore, the online self-calibration device includes a sealed cavity with a controllable volume, one side of which has a movable sealing structure for completely covering and sealing the gas sensor array; it includes one or more miniature high-pressure gas cylinders for storing standard gases and high-purity nitrogen of known concentrations; it also includes a gas pump for purging and exhausting residual gas in the cavity; the self-calibration process is as follows: The gas inspection robot automatically navigates to a safe calibration point downhole according to a preset cycle; A movable sealing structure covers the sensor array and forms a sealed space; Start the air pump and blow high-purity nitrogen into the microcavity to completely replace the air inside; after the sensor reading stabilizes, record the reading at this moment as the new zero point. Open the standard gas solenoid valve and inject standard gas into the microcavity; after the reading stabilizes, record the reading at this moment. Based on this reading and the known concentration of the standard gas, the new gain coefficient is calculated; Store the new zero point and gain coefficient for subsequent measurement calculations; Start the air pump and blow fresh air through the microcavity to reset the sealing structure.

[0013] Furthermore, the intelligent inspection strategy module is based on the Q-Learning algorithm and specifically includes: Status: Defined as the historical gas concentration sequence, most recent alarm status, equipment health status, and time for each monitoring point; Actions: The set of actions that the gas inspection robot can perform, including selecting the next inspection point, the dwell time at a point, and the speed at which it moves to the next point; Rewards: Design reward functions to guide the agent in learning the optimal policy, including: Positive rewards: rapid passage through low-risk areas; long-term monitoring in high-risk areas or concentration mutation points; accurate prediction and detection of potential hazards; Negative rewards: missed alarms; excessive energy consumption; excessively long inspection cycles; The gas inspection robot performs inspections by continuously and randomly selecting actions and choosing the action with the highest current Q value; after each action is completed, the Q table is updated based on the reward obtained; after long-term operation, it learns an optimal strategy.

[0014] Furthermore, it also includes an anti-interference data fusion algorithm module, which is used to preprocess the raw data output by the sensor in real time, filter out random noise and sudden abnormal interference, specifically by using a combination of outlier removal and Kalman filtering. The outlier removal process includes: maintaining a sliding window containing the N most recent samples using a 3σ criterion based on a moving window; for each new sample, calculating the mean μ and standard deviation σ of the data within the window; if the new sample value exceeds... If the value falls outside the range, it is considered an outlier and is either discarded or replaced with the window mean. The Kalman filtering includes: applying Kalman filtering to smooth the data stream after outlier removal; treating the true value of the gas concentration as a system state that changes slowly over time; predicting the value at the current time based on the optimal estimate at the previous time; and taking a weighted average of the actual measured value and the predicted value at the current time to obtain the optimal estimate at the current time.

[0015] On the other hand, the present invention provides a rail-mounted fixed-point detection method for a gas inspection robot, comprising the following steps: The gas inspection robot is activated, and the intelligent inspection strategy module generates the inspection plan for the day. The gas inspection robot travels along the track to the inspection point according to the inspection plan. The track encoder continuously provides odometer data; the UWB positioning system provides global coordinates at regular intervals; the extended Kalman filter algorithm fuses the information from the track encoder and the UWB positioning system in real time to output a smooth trajectory. When the gas inspection robot passes the first visual beacon, the camera captures and decodes the image to obtain an absolutely accurate position calibration. The covariance matrix of the extended Kalman filter algorithm converges, and the positioning error is reset. Finally, the gas inspection robot stops at the detection point. During the time period at the detection point, the FPGA controller sends a synchronous trigger pulse every second, and data is sampled simultaneously on the rising edge of each pulse. The system performs environmental adaptive compensation and anti-interference data fusion on the collected sensor data, outputs the corrected concentration value, and uploads it to the monitoring center; at the same time, it uses the data to update the strategy of the intelligent inspection strategy module. After running for a period of time, the gas inspection robot travels to the nearest calibration point, performs a self-calibration process, and uploads the calibration results to the monitoring center.

[0016] The innovation of this invention lies in: 1. Multi-source fusion positioning technology combining track encoders, UWB, and visual recognition achieves centimeter-level positioning accuracy.

[0017] 2. A parallel acquisition method for multi-parameter gas sensor arrays based on bus synchronous triggering.

[0018] 3. An environment adaptive compensation algorithm based on deep neural networks is used to correct gas concentration measurements in real time.

[0019] 4. Built-in automatic online calibration system with standard gas microcavity.

[0020] 5. A dynamic optimization method for intelligent inspection strategies using reinforcement learning is adopted.

[0021] 6. Anti-interference algorithm for multi-sensor data fusion using Kalman filtering and outlier removal.

[0022] The beneficial effects of this invention are as follows: (1) Improved positioning accuracy: The multi-source fusion positioning algorithm makes the robot's positioning error less than 5cm, ensuring accurate detection data.

[0023] (2) Multi-parameter synchronous detection: Parallel bus synchronous acquisition technology enables real-time synchronization of gas and temperature parameters, reducing acquisition delay by more than 50%.

[0024] (3) Strong environmental adaptability: The deep learning-based environmental compensation algorithm effectively eliminates interference from temperature, humidity, dust and other factors, improving the stability and accuracy of detection data.

[0025] (4) Reduced operation and maintenance costs: The online self-calibration function reduces the frequency of manual maintenance and improves the ability to predict the lifespan of sensors.

[0026] (5) Intelligent inspection optimization: The enhanced learning strategy dynamically adjusts the inspection plan, which improves inspection efficiency and ensures high-density monitoring of risk areas.

[0027] (6) High data reliability: Kalman filtering and outlier removal improve the anti-interference ability and reliability of sensor data.

[0028] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the workflow for a rail-mounted fixed-point detection system for gas inspection robots. Detailed Implementation

[0030] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0031] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0032] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0033] Example 1: This invention provides a rail-mounted fixed-point detection system for gas inspection robots. The overall architecture of this system is as follows: Robot Body Layer: Includes the core controller, employing a high-performance ARM+FPGA heterogeneous architecture. The ARM handles overall control, communication, and upper-level algorithms, while the FPGA handles parallel acquisition, synchronization, and preprocessing of signals from multiple sensors. Drive and Communication Unit: Includes motor drivers, brakes, explosion-proof housing, etc. High-speed data exchange with the surface platform is achieved via a networked wireless access point (AP) downhole.

[0034] In-ground monitoring platform layer: Includes data storage and analysis center: Deploys time-series databases to store historical data, and relational databases (such as MySQL) to store device information, alarm events, etc. Human-machine interface: Displayed in web format, providing real-time data visualization, historical curve query, alarm management, system configuration, and other functions.

[0035] The robot body layer includes the following modules: Multi-source fusion localization module: Provides the robot with centimeter-level absolute position information, overcoming the cumulative error and slippage problems of single encoder localization, and ensuring accurate correspondence between detection data and position. Specifically, it includes: Track encoder: Mounted on the robot's drive wheels, it measures the robot's displacement relative to the track, providing high-frequency odometer information.

[0036] UWB positioning unit: UWB tags are installed on the robot body, and UWB anchor points are installed every 50-100 meters along the track in the tunnel. By measuring the flight time of the tag to multiple anchor points, the real-time three-dimensional coordinates of the tag are calculated.

[0037] Visual recognition unit: An industrial camera facing the track, used to identify visually unique QR codes pre-placed on the track or tunnel walls. The absolute coordinates of these tags are precisely measured in advance and stored in the system.

[0038] This module employs the Extended Kalman Filter (EKF) algorithm to fuse the three data sources. High-frequency track encoder data is used as the control input to predict the robot's position and velocity. The prediction model considers possible wheel slippage, introducing process noise.

[0039]

[0040] in, It is a state variable, containing position and velocity. It is the encoder reading; When the UWB unit calculates new position coordinates, it uses them as observations to update the state estimate of the EKF and correct the cumulative error of the encoder.

[0041] When the visual recognition unit successfully identifies a QR code, the system acquires the precise coordinates pre-stored in that code. This is a high-precision, discontinuous observation that significantly corrects the current position estimate and eliminates potential long-term drift from UWB and the encoder. Through this fusion approach, the encoder ensures trajectory smoothness and short-term accuracy, UWB provides global coordinates without cumulative error, and visual recognition provides a high-precision absolute coordinate reference point. The combination of these three ultimately achieves a positioning accuracy of less than 5cm.

[0042] Multi-parameter gas detection array: Synchronously and parallelly acquires multiple gas and environmental parameters, avoiding inconsistencies in timestamps caused by serial acquisition. Specifically, it includes: Sensor array: Includes electrochemical or infrared sensors for methane, carbon monoxide, carbon dioxide, and oxygen, as well as digital temperature and humidity, dust concentration, and thermal airflow sensors.

[0043] Bus Synchronous Acquisition Controller: A controller based on an FPGA or high-performance MCU. All sensor data outputs are connected to this controller. The controller is connected to the sampling control pin of each sensor via a separate hardware synchronization trigger line.

[0044] Communication bus: It adopts an industrial-grade CAN bus or RS-485 bus, which is responsible for sending the collected data packets to the vehicle's main control computer.

[0045] The workflow is as follows: In each sampling cycle, the onboard main control computer sends a "start sampling" command to the bus synchronization acquisition controller. Upon receiving the command, the controller sends a precise pulse signal to all sensors via a hardware synchronization trigger line. All sensors simultaneously initiate internal analog-to-digital conversion or data latching on the rising edge of the pulse. After sampling is complete, the controller polls each sensor sequentially according to a preset order or via a token ring protocol, reading its converted data. Each data packet carries the same synchronization timestamp. The controller packages all sensor data into a single frame and sends it to the main control computer via CAN or RS-485 bus. This method physically ensures the consistency of sampling time for all parameters, resulting in lower acquisition latency compared to traditional serial polling methods.

[0046] Environmental Adaptive Compensation Algorithm Module: This module uses an algorithm model to correct gas sensor readings based on real-time environmental parameters, eliminating measurement errors caused by environmental factors such as temperature, humidity, dust, and airflow. This module is a purely software-based module running on the vehicle's main control computer. It employs a pre-trained deep neural network model, specifically a multilayer perceptron structure.

[0047] The model input is a vector containing the raw readings from all sensors: [C CH4 CCO [,T,H,P,V], where C is the concentration, T is the temperature, H is the humidity, P is the dust concentration, and V is the airflow velocity.

[0048] The model structure is as follows: Input layer: The number of nodes is the same as the dimension of the input vector.

[0049] Hidden layers: 3 to 5 fully connected layers, each containing 64 to 256 neurons, using ReLU as the activation function to enhance non-linear fitting capabilities.

[0050] Output layer: The number of nodes is the number of gas types that need correction, and the output is the corrected concentration value: [C CH4corrected C COcorrected ].

[0051] Training Process: This model requires offline training in a laboratory environment. In a standard gas chamber, standard gases of varying concentrations are configured, and environmental conditions such as temperature, humidity, and dust concentration are systematically varied. Simultaneously, the raw readings from each sensor and the actual gas concentrations are recorded. A massive dataset is used as the training set, and the network weights are trained using a backpropagation algorithm until the error between the model output and the true value is minimized.

[0052] Online self-calibration device: Eliminates the need for manual downhole operation, automatically performing periodic zero-point and gain calibration on gas sensors to compensate for long-term sensor drift. Includes: Standard Gas Microcavity: A 50mL sealed chamber with a movable mechanical structure on one side to completely enclose and seal the gas sensor array. Standard Gas Cylinder and Valve Assembly: Contains one or more miniature high-pressure gas cylinders storing 500ppm CH4 standard gas and high-purity nitrogen for zero-point calibration. A precision solenoid valve accurately controls the gas flow. Miniature Gas Pump: A miniature vacuum pump or gas pump used to purge and remove residual gas from the chamber. The workflow is as follows: Once a month, or when sensor drift exceeds a threshold, the system automatically navigates to a safe calibration point downhole. A mechanical structure drives a microcavity, enclosing the sensor array and forming a sealed space. An air pump is activated, blowing high-purity nitrogen into the microcavity to completely replace the air inside. Once the sensor reading stabilizes, this reading is recorded as the new zero point. The standard gas solenoid valve is opened, injecting standard gas into the microcavity. Once the reading stabilizes, this reading is recorded. Based on this reading and the known concentration of the standard gas, the system calculates a new sensitivity or gain coefficient. The new zero point and gain coefficient are stored for subsequent measurement calculations. The air pump is activated again to purge the microcavity with fresh air, and then the mechanical structure resets. The entire process is fully automated.

[0053] The intelligent inspection strategy module dynamically adjusts the inspection path, inspection points, and dwell time at each point based on historical and real-time data. It employs Q-Learning or DQN algorithms from reinforcement learning, specifically including: State: Defined as a vector describing the current risk status of the entire inspection area, including: [L current ,T last_visit,1 ,...,T last_visit,n G avg,1, ...,G avg,n G var,1, ...,G var,n Where L is the robot's current position, T is the time since the last inspection at each detection point, and G is the time since the last inspection. avg and G var It represents the average and variance of the historical gas concentration at each point.

[0054] Action: The set of actions that the robot can perform, namely "go to the next detection point X and stay to detect for Y seconds".

[0055] Rewards: Design a reward function to guide the agent in learning the optimal strategy. This includes: awarding a high positive reward for detecting a significant increase or fluctuation in gas concentration; awarding a small positive reward after a long period of neglect is inspected; and awarding a small negative reward for the time and energy consumed by the robot's movement to improve efficiency.

[0056] The robot performs inspections by continuously selecting actions randomly and choosing the action with the highest current Q value. After each action is completed, the weights of the Q-table or DQN network are updated based on the reward received. Through long-term operation, the system learns an optimal strategy: high-frequency, long-term focused monitoring of high-risk areas with previously abnormal or highly fluctuating concentrations, while reducing the inspection frequency of areas with long-term stability, thereby improving inspection efficiency while ensuring safety.

[0057] Anti-interference data fusion algorithm: This algorithm performs real-time preprocessing on the raw data output from sensors to filter out random noise and sudden abnormal interference, improving data stability and reliability. It employs a combination of outlier removal and Kalman filtering.

[0058] Outlier removal: A 3σ criterion based on a moving window is used. The system maintains a sliding window containing the most recent N=10 samples. For each new sample, the mean μ and standard deviation σ of the data within the window are calculated. If the new sample value exceeds... If the value falls within the specified range, it is considered an outlier and is either discarded or replaced with the window mean.

[0059] Kalman filtering: For the data stream after outlier removal, Kalman filtering is applied for smoothing. This algorithm treats the true gas concentration as a system state that changes slowly over time.

[0060] The model predicts the current value based on the optimal estimate from the previous time step. It typically assumes that gas concentration changes are stationary or subject to minor random fluctuations. The optimal estimate for the current time step is obtained by weighted averaging the actual measurement and the predicted value. The Kalman gain depends on the magnitude of both the prediction and measurement errors. Higher measurement noise leads to greater confidence in the prediction, and vice versa. This method effectively filters out high-frequency random noise from the sensor itself, outputting a smooth concentration change curve that more closely approximates the true value.

[0061] Example 2: This embodiment describes the specific application of the present invention in a typical coal mine ventilation roadway.

[0062] A suspended track is deployed at the top of a 500-meter-long ventilation tunnel. UWB anchor points are installed every 50 meters along the track inside the tunnel, totaling 11. Below the track, a QR code visual label containing a unique ID and coordinate information is affixed every 10 meters. A gas inspection robot is deployed on the track. The robot's hardware configuration is as follows: Main control computer: NVIDIA Jetson AGX Orin, providing powerful parallel computing capabilities to run DNN and RL models. Positioning hardware: E6B2-CWZ6C incremental encoder, Decawave DWM1001 UWB module, Hikvision MV-CE060-10UC industrial camera.

[0063] The sensor array includes: Methane: 0-5% CH4 using an NDIR infrared sensor.

[0064] Carbon monoxide: 0-1000 ppm CO was measured using an electrochemical sensor.

[0065] Temperature and humidity: SHT31 digital sensor.

[0066] Dust: PM2.5 laser dust sensor.

[0067] Bus controller: Developed based on Xilinx Artix-7 FPGA, achieving nanosecond-level synchronous triggering accuracy.

[0068] Self-calibration device: It has a built-in 30mL standard gas cylinder and a 30mL high-purity N2 gas cylinder. The standard gas cylinder contains a mixture of 500ppm CH4 and 100ppm CO.

[0069] like Figure 1 As shown, the workflow is as follows: Startup and initialization: The robot is awakened, and the intelligent inspection strategy module generates the inspection plan for the day, prioritizing the inspection of the three inspection points P1, P2, and P3 that are closest to the upstream working face.

[0070] Heading to point P1: The robot departs from the docking point, and the main control computer commands the motors to start.

[0071] Localization process: The encoder provides odometry data at a frequency of 100Hz. The UWB module provides global coordinates at a frequency of 10Hz. The EKF algorithm fuses the two in real time, outputting a smooth trajectory. When the robot passes the first QR code (ID: V-001), the camera captures and decodes it, the system obtains an absolutely accurate position calibration, the EKF covariance matrix converges, and the localization error is reset. Finally, the robot stops at point P1 (coordinates: 100.00m), with an actual position of 100.02m and a localization error of 2cm.

[0072] Detection at point P1: The policy module instructs to stay at point P1 for 60 seconds for detection.

[0073] Data Acquisition: During these 60 seconds, the FPGA controller sends a synchronous trigger pulse every second. The CH4, CO, temperature and humidity, and dust sensors simultaneously sample data on the rising edge of each pulse.

[0074] Data processing: The 60 sets of raw data collected were sent into the processing flow: Anti-interference: The moving window 3σ algorithm detected a spike pulse in the CO reading at the 32nd second that far exceeded normal fluctuations, which was identified as interference and removed. The subsequent Kalman filter smoothed the 60 data points, outputting a stable average value.

[0075] Environmental compensation: The temperature inside the tunnel is 28℃, and the humidity is 90%RH. The pre-processed gas readings, along with the temperature and humidity data, are input into the pre-trained DNN model. The model outputs the corrected concentrations: CH4 is 0.45%, and CO is 15ppm.

[0076] Data Reporting and Decision Making: The final results (location: 100.02m, CH4: 0.45%, CO: 15ppm, ...) were packaged and sent to the surface monitoring center via the downhole wireless network. Simultaneously, this data updated the state vector of the RL model. Since the CH4 concentration of 0.45% was slightly higher than the historical average, the reward function assigned a small positive reward, which will slightly increase the probability of future inspections of point P1.

[0077] Automatic Calibration: After one month of operation, the system detected that the zero-point reading of the CH4 sensor was slowly and continuously deviating from 0 in clean air, with the cumulative drift reaching a threshold. The system automatically generated a calibration task. During nighttime production shutdowns, the robot automatically traveled to a dedicated calibration dock at the end of the aisle. The self-calibration device activated, a sealed cover enveloped the sensor, nitrogen was first introduced to complete zero-point calibration, and then a standard mixed gas was introduced to complete gain calibration. The entire process took approximately 15 minutes and was performed without human intervention. After calibration, the robot reported the new calibration coefficients for record-keeping and returned to standby mode.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A rail-mounted fixed-point detection system for gas inspection robots, characterized in that: include: The multi-source fusion positioning module locates the gas inspection robot through the track encoder, UWB positioning system and visual recognition system respectively, and uses the extended Kalman filter algorithm to fuse them to achieve multi-source fusion positioning of the gas inspection robot. The multi-parameter gas detection array includes various gas sensors, dust sensors, and temperature and humidity sensors installed on the gas inspection robot. Based on a bus synchronous acquisition mechanism, it realizes parallel sampling of multiple parameters. The environmental adaptive compensation algorithm module uses a deep neural network model to combine temperature, humidity, airflow and dust data to correct gas concentration in real time. The online self-calibration device has a built-in standard gas microcavity filled with zero-point gas and range gas of known concentration. It is connected to the outside world and the gas chambers of various gas sensors through a solenoid valve, and automatically calibrates the zero point and gain of the sensor periodically. The intelligent inspection strategy module, based on reinforcement learning algorithms, dynamically adjusts the inspection plan of the gas inspection robot according to real-time and historical data.

2. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The track encoder is mounted on the drive wheel of the gas inspection robot, providing continuous mileage information.

3. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The UWB positioning system includes UWB anchor points deployed at certain intervals in the inspection roadway, and UWB tags set on the gas inspection robot. The absolute position of the gas inspection robot is calculated by measuring the time of flight (TOF) of the wireless signals between the tag and multiple anchor points.

4. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The visual recognition system includes an industrial camera installed at the front end of the gas inspection robot and visual beacons with unique IDs distributed in the inspection lanes. The absolute coordinates of each visual beacon are pre-stored. The industrial camera identifies the visual beacons, calculates their pose in the image, and obtains the absolute position reference point.

5. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The method of using the extended Kalman filter algorithm for fusion to achieve multi-source fusion localization of the gas inspection robot specifically includes: The extended Kalman filter algorithm is adopted. The state variables are the position and velocity of the gas inspection robot, and the observations are the encoder odometer increment, the absolute position of the UWB, and the absolute position of the visual beacon. Using high-frequency track encoder data as control input, the position and velocity of the gas inspection robot are estimated, as follows: in, It is a state variable. It is the encoder reading; When the UWB positioning system calculates new position coordinates, it uses them as observations to update the state estimate and correct the cumulative error of the track encoder. When the visual recognition system identifies a visual beacon, it obtains the pre-stored absolute position of the visual beacon, performs a strong correction, and eliminates the long-term drift that may exist in the UWB positioning system and track encoder.

6. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The multi-parameter gas detection array generates a global hardware trigger signal based on an FPGA controller, which is simultaneously sent to all sensor interface circuits via a hardware synchronization trigger line. Specifically, it includes: During each sampling cycle, the main control computer on the gas inspection robot sends a start sampling command to the FPGA controller. After receiving the instruction, the controller sends a pulse signal to all sensors via the hardware synchronization trigger line; all sensors simultaneously start their internal analog-to-digital conversion on the same edge of the pulse. After sampling is completed, the FPGA controller polls each sensor in a preset order and reads the converted data; each data packet is accompanied by the same synchronization timestamp. The FPGA controller packages all sensor data into a single frame and sends it to the host computer via the RS-485 bus.

7. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The environmental adaptive compensation algorithm module is based on a multilayer sensor (MLP) structure. It corrects the gas sensor readings according to real-time environmental parameters to eliminate measurement errors. The multilayer sensor (MLP) specifically includes: Input layer: The number of nodes is the same as the dimension of the input vector, and the input vector is the raw readings of all sensors; Hidden layers: 3 to 5 fully connected layers, each containing 64 to 256 neurons, using ReLU as the activation function to enhance non-linear fitting ability; Output layer: The number of nodes is the number of gas types that need to be corrected, and the corrected concentration values ​​are output. The training data consisted of standard gases of different concentrations introduced into the laboratory under different temperature, humidity, and dust conditions. Pairs of raw readings and actual concentrations were collected, and the network weights were trained using the backpropagation algorithm until the error between the model output and the actual value was minimized.

8. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The online self-calibration device includes a sealed cavity with a controllable volume, one side of which has a movable sealing structure for completely covering and sealing the gas sensor array; it includes one or more miniature high-pressure gas cylinders for storing standard gases and high-purity nitrogen of known concentrations; it also includes a gas pump for purging and removing residual gas from the cavity; the self-calibration process is as follows: The gas inspection robot automatically navigates to a safe calibration point downhole according to a preset cycle; A movable sealing structure covers the sensor array and forms a sealed space; Start the air pump and blow high-purity nitrogen into the microcavity to completely replace the air inside; after the sensor reading stabilizes, record the reading at this moment as the new zero point. Open the standard gas solenoid valve and inject standard gas into the microcavity; after the reading stabilizes, record the reading at this moment. Based on this reading and the known concentration of the standard gas, the new gain coefficient is calculated; Store the new zero point and gain coefficient for subsequent measurement calculations; Start the air pump and blow fresh air through the microcavity to reset the sealing structure.

9. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: The intelligent inspection strategy module is based on the Q-Learning algorithm and specifically includes: Status: Defined as the historical gas concentration sequence, most recent alarm status, equipment health status, and time for each monitoring point; Actions: The set of actions that the gas inspection robot can perform, including selecting the next inspection point, the dwell time at a point, and the speed at which it moves to the next point; Rewards: Design reward functions to guide the agent in learning the optimal policy, including: Positive rewards: rapid passage through low-risk areas; long-term monitoring in high-risk areas or concentration mutation points; accurate prediction and detection of potential hazards; Negative rewards: missed alarms; excessive energy consumption; excessively long inspection cycles; The gas inspection robot performs inspections by continuously and randomly selecting actions and choosing the action with the highest current Q value; after each action is completed, the Q table is updated based on the reward obtained; after long-term operation, it learns an optimal strategy.

10. The gas inspection robot rail-mounted fixed-point detection system according to claim 1, characterized in that: It also includes an anti-interference data fusion algorithm module, which is used to preprocess the raw data output by the sensor in real time, filter out random noise and sudden abnormal interference, specifically by using a combination of outlier removal and Kalman filtering. The outlier removal process includes: maintaining a sliding window containing the N most recent samples using a 3σ criterion based on a moving window; for each new sample, calculating the mean μ and standard deviation σ of the data within the window; if the new sample value exceeds... If the value falls outside the range, it is considered an outlier and is either discarded or replaced with the window mean. The Kalman filtering includes: applying Kalman filtering to smooth the data stream after outlier removal; treating the true value of the gas concentration as a system state that changes slowly over time; predicting the value at the current time based on the optimal estimate at the previous time; and taking a weighted average of the actual measured value and the predicted value at the current time to obtain the optimal estimate at the current time.

11. A method for fixed-point detection of gas inspection robots using a rail-mounted mechanism, characterized in that: Includes the following steps: The gas inspection robot is activated, and the intelligent inspection strategy module generates the inspection plan for the day. The gas inspection robot travels along the track to the inspection point according to the inspection plan. The track encoder continuously provides odometer data; the UWB positioning system provides global coordinates at regular intervals; the extended Kalman filter algorithm fuses the information from the track encoder and the UWB positioning system in real time to output a smooth trajectory. When the gas inspection robot passes the first visual beacon, the camera captures and decodes the image to obtain an absolutely accurate position calibration. The covariance matrix of the extended Kalman filter algorithm converges, and the positioning error is reset. Finally, the gas inspection robot stops at the detection point. During the time the device stays at the detection point, the FPGA controller sends a synchronous trigger pulse every second, and data is sampled simultaneously on the rising edge of each pulse. The system performs environmental adaptive compensation and anti-interference data fusion on the collected sensor data, outputs the corrected concentration value, and uploads it to the monitoring center; at the same time, it uses the data to update the strategy of the intelligent inspection strategy module. After running for a period of time, the gas inspection robot travels to the nearest calibration point, performs a self-calibration process, and uploads the calibration results to the monitoring center.

Images