Intelligent inspection robot for underground plant environment

Abstract

The invention discloses an intelligent inspection robot for an underground plant environment, which comprises a wheel-foot robot body (1), environment sensing modules (2) are arranged at the front and rear ends and two sides of the body of the wheel-foot robot body (1), and an equipment state detection module (3) is arranged at the back of the wheel-foot robot body (1); the wheel-foot robot further comprises a communication module (5) connected with a remote control center, and the communication module (5) is connected with a data processing and control module (4) for controlling the wheel-foot robot body (1), the environment sensing module (2) and the equipment state detection module (3). The method has the characteristics that manual inspection can be replaced (the safety risk in the manual inspection process is eliminated) and the inspection efficiency is improved.

Description

TECHNICAL FIELD

[0001] The application relates to a pumped storage power plant inspection equipment, in particular to an intelligent inspection robot for an underground power house environment. BACKGROUND

[0002] The pumped storage power station is a core facility for peak shaving, valley filling, frequency regulation, phase modulation and accident backup in the power grid. The internal structure of the underground power house is complex, the space is huge, and key equipment such as generator sets, transformers, GIS equipment, various pipeline valves and cable corridors is arranged. At present, the inspection work of the underground power house mainly depends on manual work, which has the following obvious shortcomings: 1) high environmental risk: there may be noise, humidity, SF6 gas leakage, local high temperature and high pressure in the power house, which threatens the safety of the inspection personnel.

[0003] 2) low inspection efficiency: manual inspection takes a long time and has limited frequency, and it is difficult to realize all-weather and full-coverage monitoring.

[0004] Therefore, if a robot that can meet the inspection of the underground space of the pumped storage power station can be developed, the above problems can be well solved. SUMMARY

[0005] The purpose of the application is to provide an intelligent inspection robot for an underground power house environment. The application has the characteristics of being able to replace manual inspection (eliminating the safety risks in the process of manual inspection) and improving the inspection efficiency.

[0006] The technical scheme of the application: the intelligent inspection robot for the underground power house environment comprises a wheel-foot robot body, environment perception modules are arranged at the front and rear ends and both sides of the trunk of the wheel-foot robot body, and a device state detection module is arranged on the back of the wheel-foot robot body; further comprising a communication module connected with a remote control center, the communication module is connected with a data processing and control module for controlling the wheel-foot robot body, the environment perception module and the device state detection module.

[0007] In the foregoing intelligent inspection robot for the underground power house environment, the environment perception module comprises a laser radar and a visual camera located at the front and rear ends of the wheel-foot robot body, and further comprises a surround view camera located at both sides of the trunk of the wheel-foot robot body.

[0008] In the foregoing intelligent inspection robot for the underground power house environment, the device state detection module comprises a rotary mechanical arm located at the middle position of the back of the wheel-foot robot body, a high-definition visible light camera and an infrared thermal imager are arranged on the front side of the rotary mechanical arm; an acoustic sensor, a gas sensor and a partial discharge sensor are arranged on the rear side of the rotary mechanical arm, and a vibration sensor is arranged on the rotary mechanical arm.

[0009] In the foregoing intelligent inspection robot for underground powerhouse environment, the specific inspection control process of the data processing and control module adopts the following steps: S1, a complete environment map of the underground powerhouse is constructed, and an operator labels key equipment points and inspection path points on the map to generate a global inspection route; S2, the inspection robot receives an inspection task issued by a remote control center, and after receiving the task, the inspection robot autonomously moves along the predetermined path for inspection; S3, after reaching a preset inspection detection point, the inspection robot stops and uses the equipment state detection module to collect equipment state and environmental parameter data; S4, the collected data is analyzed, and based on the analysis result, an inspection result is generated and uploaded to the remote control center, and an alarm is issued when an exception occurs.

[0010] In the foregoing intelligent inspection robot for underground powerhouse environment, the analysis and processing process of the collected data in step S4 is: S41, the data collected by each sensor is feature extracted to generate a feature vector of each sensor modality; S42, the feature vectors of each modality are preprocessed and time-aligned; S43, the feature vectors after preprocessing and alignment are feature spliced to obtain a joint feature vector; S44, the joint feature vector is input into an XGBoost model for inference to output an analysis result.

[0011] In the foregoing intelligent inspection robot for underground powerhouse environment, in step S41, for the visible light image collected by the high-definition visible light camera, YOLOv8 is used for analysis to output a visible light feature vector [feat_img1, feat_img2,...]; For the infrared thermal image collected by the infrared thermal imager, U-Net is used to segment the corresponding region, calculate the max_temp, avg_temp and temp_std of the region, and output an infrared feature vector: [feat_ir1, feat_ir2, feat_ir3]; For the sound signal collected by the acoustic sensor, the first five coefficients mfcc1, mfcc2, mfcc3, mfcc4, and mfcc5 of the MFCC are calculated, and a sound feature vector: [feat_audio1,..., feat_audio5] is outputted; For the vibration signal collected by the vibration sensor: the vibration signal is subjected to FFT, the amplitude vib_1x of the rotation frequency (1x) is extracted, the amplitude vib_2x of the double frequency (2x) is extracted, the high-frequency band RMS value vib_hf_rms is extracted, and the vibration feature vector [feat_vib1, feat_vib2, feat_vib3] is outputted; For the gas signal collected by the gas sensor, the response values of the entire sensor array at a certain moment are combined to form a vector, PCA dimension reduction is performed, the first two principal component scores PC1 and PC2 are taken as features, and the gas feature vector [feat_gas] is outputted. A 7-dimensional feature vector feat_pd containing [Skewness, Kurtosis, Cross_correlation, Rise_time, Dominant_frequency, Spectral_energy, Discharge_trend] is extracted from the partial discharge signal collected by the partial discharge sensor; wherein Skewness is skewness, Kurtosis is kurtosis, Cross_correlation is cross-correlation coefficient, Rise_time is rise time, Dominant_frequency is dominant frequency, Spectral_energy is energy in a specific frequency band, and Discharge_trend is the change slope of the maximum discharge amount in a set time period.

[0012] In the foregoing intelligent inspection robot for underground powerhouse environment, in step S42, the StandardScaler is used for standardization processing, and the time sequence alignment is that all sensor data are marked with a uniform high-precision timestamp.

[0013] Compared with the prior art, the environment perception module is arranged at the front and rear ends of the wheel-foot robot body and the two sides of the trunk, the equipment state detection module is arranged on the back of the wheel-foot robot body, and the data processing and control module is arranged, so as to realize the inspection of the underground powerhouse and the equipment detection of each equipment inspection point, thereby replacing manual inspection, eliminating the safety risks in the manual inspection process, and effectively improving the inspection efficiency. In addition, the data processing and control module of the present application also analyzes and processes the collected data, effectively fuses the collected data of various modal sensors, and outputs the analysis result, which is helpful for the control center personnel to intuitively understand the on-site equipment operation condition and whether there is a risk. In summary, the present application has the characteristics of being able to replace manual inspection (eliminate the safety risks in the manual inspection process) and improve the inspection efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 It is a structural schematic diagram of the present application.

[0015] The marks in the drawings are: 1- wheel-legged robot body, 2- environment perception module, 3- equipment state detection module, 4- data processing and control module, 5- communication module, 21- laser radar, 22- visual camera, 23- surround view camera, 31- rotary robot arm, 32- vibration sensor, 33- high-definition visible light camera, 34- infrared thermal imager, 35- acoustic sensor, 36- gas sensor, 37- partial discharge sensor. DETAILED DESCRIPTION

[0016] The application will be further described below in conjunction with the drawings and examples, but not as the basis for limiting the application.

[0017] Example. Intelligent inspection robot for underground powerhouse environment, as shown in Figure 1 The figure shows that it includes a wheel-legged robot body 1, environment perception modules 2 are arranged at the front and rear ends and both sides of the trunk of the wheel-legged robot body 1, and an equipment state detection module 3 is arranged at the back of the wheel-legged robot body 1; it also includes a communication module 5 connected with a remote control center, and a data processing and control module 4 connected with the wheel-legged robot body 1, the environment perception module 2 and the equipment state detection module 3.

[0018] The environment perception module 2 includes laser radars 21 and visual cameras 22 arranged at the front and rear ends of the wheel-legged robot body 1, and surround view cameras 23 arranged at both sides of the trunk of the wheel-legged robot body 1.

[0019] The equipment state detection module 3 includes a rotary robot arm 31 arranged at the middle position of the back of the wheel-legged robot body 1, a high-definition visible light camera 33 and an infrared thermal imager 34 arranged at the front side of the rotary robot arm 31; an acoustic sensor 35, a gas sensor 36 and a partial discharge sensor 37 arranged at the back side of the rotary robot arm 31, and a vibration sensor 32 arranged on the rotary robot arm 31.

[0020] The specific inspection control process of the data processing and control module adopts the following steps: S1, build a complete environment map of the underground powerhouse, and mark key equipment points and inspection path points on the map by the operator to generate a global inspection route; S2, the inspection robot receives the inspection task issued by the remote control center, and after receiving the task, the inspection robot moves along the predetermined path for inspection; S3, after reaching the preset inspection detection point, the inspection robot stops and uses the equipment state detection module to collect equipment state and environmental parameter data; S4, analyze the collected data, generate an inspection result based on the analysis result, upload the inspection result to the remote control center, and issue an alarm when an abnormality occurs.

[0021] The analysis and processing of the data collected in step S4 are as follows: S41, feature extraction is performed on the data collected by each sensor to generate a feature vector of each sensor modality; S42, pre-processing and time sequence alignment are performed on the feature vectors of each modality; S43, feature splicing is performed on the pre-processed and aligned feature vectors to obtain a joint feature vector; S44, the joint feature vector is input into the XGBoost model for inference, and an analysis result is output.

[0022] In step S41, for the visible light image collected by the high-definition visible light camera, YOLOv8 is used for analysis, and a visible light feature vector [feat_img1, feat_img2,...] is output. Image recognition (visible light) Task: Instrument reading, state indication, appearance anomaly (rust, oil leakage, crack), infrared hot spot analysis.

[0023] Algorithm: Target detection and image segmentation model based on deep learning.

[0024] YOLOv7 / v8 (You Only Look Once): Single-stage detection algorithm is used. The image is divided into grids, and each grid predicts multiple bounding boxes and class probabilities. Extremely fast, meets the real-time inspection requirements.

[0025] For the infrared thermal image collected by the infrared thermal imager, U-Net is used to segment the corresponding region, calculate the max_temp, avg_temp and temp_std of the region, and output the infrared feature vector: [feat_ir1, feat_ir2, feat_ir3]; Image recognition (infrared) U-Net: Principle: Fully convolutional network with encoder-decoder structure. The encoder down-samples to extract features, and the decoder up-samples to restore the spatial dimension and preserves the detail information through skip connection. The output is a class label for each pixel (pixel-level classification). It is used to accurately segment the overheat area of transformer bushings, circuit breaker contacts and other components, calculate the accurate temperature distribution, area and shape, far beyond simple box selection.

[0026] For the sound signal collected by the acoustic sensor, the first 5 coefficients mfcc1, mfcc2, mfcc3, mfcc4, mfcc5 of MFCC are calculated, and the sound feature vector: [feat_audio1,..., feat_audio5] is output. For the vibration signal collected by the vibration sensor: perform FFT on the vibration signal, extract the amplitude vib_1x of the rotation frequency 1x, the amplitude vib_2x of the double frequency 2x, and the high-frequency band RMS value vib_hf_rms, and output the vibration feature vector: [feat_vib1, feat_vib2, feat_vib3]; For the gas signal collected by the gas sensor, the response values of the entire sensor array at a certain time are combined into a vector, PCA dimensionality reduction is performed, the first two principal component scores PC1 and PC2 are taken as features, and the gas feature vector [feat_gas] is output. From the partial discharge signal collected by the partial discharge sensor, a 7-dimensional feature vector feat_pd containing [Skewness, Kurtosis, Cross_correlation, Rise_time, Dominant_frequency, Spectral_energy, Discharge_trend] is extracted; wherein Skewness is skewness, Kurtosis is kurtosis, Cross_correlation is cross-correlation coefficient, Rise_time is rise time, Dominant_frequency is dominant frequency, Spectral_energy is energy in a specific frequency band, and Discharge_trend is the change slope of the maximum discharge amount in a set time period.

[0027] In step S42, StandardScaler is used for standardization processing, and time series alignment is to apply uniform high-precision timestamps to all sensor data.

[0028] The setting of the rotary mechanical arm can make the vibration sensor better fit the equipment to be measured and collect vibration data.

[0029] Laser radar: used to build high-precision two-dimensional / three-dimensional point cloud maps of underground workshops, and realize SLAM (simultaneous localization and mapping), providing global positioning and obstacle avoidance functions for robots.

[0030] Depth vision camera: auxiliary three-dimensional environment perception and close-range obstacle avoidance.

[0031] High-definition visible light camera: used to collect equipment appearance images, perform instrument reading recognition, oil level meter recognition, indicator light state recognition, and equipment surface rust / oil leakage appearance anomaly detection. Visible light camera: 20 million pixels or more industrial camera, global shutter, supporting wide-angle and fixed-focus lenses and fill light.

[0032] Infrared thermal imager: used to detect the temperature distribution of transformer windings, circuit breaker contacts, cable joints and other equipment, and find overheating defects. Infrared thermal imager: resolution not less than 640x480, support temperature flow output, integrated laser range finder for automatic correction of distance parameters.

[0033] Acoustic sensor: used to collect noise and abnormal sound during equipment operation, and judge equipment state through voiceprint analysis (such as bearing damage). Acoustic sensor: industrial microphone array with directional noise reduction function.

[0034] Vibration sensor: three-axis IEPE acceleration sensor, frequency range at least 0.5Hz-10kHz.

[0035] Gas sensor: used to detect SF6 gas concentration, ozone concentration, smoke and other environmental factors, prevent gas leakage and fire risk. Gas sensor: high-performance electrochemical SF6 sensor and MOS multi-gas sensor module.

[0036] Partial discharge sensor: used to detect partial discharge phenomenon of high-voltage electrical equipment.

[0037] Communication module, using industrial-grade wireless AP, supporting Wi-Fi 6 / 5G network, realizing high-speed and low-latency data interaction with station-side monitoring platform. In areas with weak wireless signals, it has data caching and breakpoint resume functions.

[0038] Training process of XGBoost model: Prepare a large historical data set, where each sample is a "joint feature vector" and corresponds to a "label" (such as "normal", "bearing failure", "winding overheating", etc.).

[0039] Divide the data set into training set, validation set and test set.

[0040] Train the selected model (such as XGBoost) with the training set data.

[0041] Adjust the model hyperparameters with the validation set to prevent overfitting.

[0042] Evaluate the final performance of the model (accuracy, F1-score, etc.) with the test set.

[0043] Inference process of trained XGBoost model: 1) The inspection robot collects data online;

[0044] 2) Each module extracts features and pre-processes in real time;

[0045] 3) Joint vector is spliced;

[0046] 4) Input the joint vector into the trained XGBoost model;

[0047] 5) The model outputs the final diagnosis and probability.

Claims

1. An intelligent inspection robot for an underground plant environment, characterized by: The wheel-legged robot body (1) is provided with an environment perception module (2) at both ends and both sides of the trunk, and a device state detection module (3) is arranged on the back of the wheel-legged robot body (1); the communication module (5) connected with the remote control center is further included, and the data processing and control module (4) controlling the wheel-legged robot body (1), the environment perception module (2) and the device state detection module (3) is connected to the communication module (5). 2.The intelligent inspection robot for underground plant environment of claim 1, wherein: The environment perception module (2) includes a laser radar (21) and a visual camera (22) located at both ends of the wheel-legged robot body (1), and further includes a surround-view camera (23) located at both sides of the trunk of the wheel-legged robot body (1). 3.The intelligent inspection robot for underground plant environment of claim 1, wherein: The device state detection module (3) includes a rotary robot arm (31) located at the middle position of the back of the wheel-legged robot body (1), a high-definition visible light camera (33) and an infrared thermal imager (34) are arranged on the front side of the rotary robot arm (31); an acoustic sensor (35), a gas sensor (36) and a partial discharge sensor (37) are arranged on the rear side of the rotary robot arm (31), and a vibration sensor (32) is arranged on the rotary robot arm (31).

4. The intelligent inspection robot for underground powerhouse environment according to claim 1, characterized in that: The specific inspection control process of the data processing and control module adopts the following steps: S1, a complete environment map of the underground powerhouse is constructed, the operator marks key equipment points and inspection path points on the map, and generates a global inspection route; S2, the inspection robot receives the inspection task issued by the remote control center, and after receiving the task, the inspection robot moves along the predetermined path for inspection; S3, after reaching the preset inspection detection point, the inspection robot stops, and the device state detection module is used to collect device state and environmental parameter data; S4, analyze the collected data, generate an inspection result based on the analysis result, upload the remote control center, and issue an alarm when an exception occurs.

5. The intelligent inspection robot for underground powerhouse environment according to claim 1, characterized in that: The analysis and processing process of the collected data in step S4 is: S41, feature extraction is performed on the data collected by each sensor to generate a feature vector of each sensor modality; S42, the feature vectors of each modality are preprocessed and time-aligned; S43, the feature vectors after preprocessing and alignment are spliced to obtain a joint feature vector; S44, input the joint feature vector into the XGBoost model for inference, and output the analysis result. 6.The intelligent inspection robot for underground plant environment of claim 5, wherein: In step S41, for the visible light image collected by the high-definition visible light camera, YOLOv8 is used for analysis, and a visible light feature vector [feat_img1, feat_img2,...] is outputted; For the infrared thermal image collected by the infrared thermal imager, the corresponding region is segmented using U-Net, the max_temp, avg_temp and temp_std of the region are calculated, and an infrared feature vector [feat_ir1, feat_ir2, feat_ir3] is outputted; For the sound signal collected by the acoustic sensor, the first 5 coefficients mfcc1, mfcc2, mfcc3, mfcc4, mfcc5 of MFCC are calculated, and the sound feature vector [feat_audio1,..., feat_audio5] is outputted; For the vibration signal collected by the vibration sensor: the vibration signal is subjected to FFT, the 1x amplitude vib_1x, the 2x amplitude vib_2x and the high-frequency band RMS value vib_hf_rms are extracted, and the vibration feature vector [feat_vib1, feat_vib2, feat_vib3] is outputted; For the gas signal collected by the gas sensor, the response values of the entire sensor array at a certain moment are combined to form a vector, PCA dimension reduction is performed, the first two principal component scores PC1 and PC2 are taken as features, and the gas feature vector [feat_gas] is outputted; A 7-dimensional feature vector feat_pd containing [Skewness, Kurtosis, Cross_correlation, Rise_time, Dominant_frequency, Spectral_energy, Discharge_trend] is extracted from the partial discharge signal collected by the partial discharge sensor; wherein Skewness is skewness, Kurtosis is kurtosis, Cross_correlation is cross-correlation coefficient, Rise_time is rise time, Dominant_frequency is dominant frequency, Spectral_energy is energy in a specific frequency band, and Discharge_trend is the change slope of the maximum discharge amount in a set time period.

7. The intelligent inspection robot for underground powerhouse environment according to claim 5, characterized in that: In step S42, the StandardScaler is used for standardization processing, and the time sequence alignment is to mark all sensor data with a uniform high-precision timestamp.

Images