A wastewater plant aeration identification method based on underwater robot

Abstract

The application discloses a sewage plant aeration identification method based on an underwater robot, and comprises the following steps: step S1, inspection path planning and multi-source data collection; step S2, underwater image preprocessing and target segmentation identification; step S3, aeration state quantitative calculation and working condition threshold comparison; step S4, bubble motion trajectory analysis and abnormal precise discrimination; step S5, multi-source data cross verification and extreme working condition adaptive processing; and step S6, result visualization output and aeration energy consumption intelligent optimization. The application realizes unmanned intelligent inspection of the whole domain of a sewage plant aeration tank, and finally realizes the purposes of intelligent diagnosis of an aeration system fault, precise evaluation of a state, energy-saving optimization of energy consumption, and automatic operation and management control.

Description

Technical Field

[0001] This invention relates to smart water management, intelligent monitoring of wastewater treatment, and the cross-technology of underwater robot inspection and machine vision image recognition. More specifically, it relates to a wastewater treatment plant aeration identification method based on an underwater robot. Background Technology

[0002] Currently, monitoring of aeration systems in wastewater treatment plants still mainly relies on manual surface visual observation, single-point dissolved oxygen monitoring, and overall blower airflow monitoring.

[0003] Humans can only observe the aeration on the water surface and cannot penetrate the water to see the actual working status of the underwater aeration heads. Early blockages and local damage to the aeration heads are difficult to detect, and many early hidden dangers cannot be identified in time.

[0004] Traditional dissolved oxygen monitoring is a single-point, small-scale method that can only represent the water conditions in a small area around the measuring point. It cannot achieve full coverage of the aeration tank and cannot reflect the differences in aeration intensity distribution within the tank.

[0005] Blower air volume monitoring can only count the total aeration volume, but cannot pinpoint the specific location of insufficient aeration, excessive aeration, or local blockage, thus lacking fault location capability.

[0006] In summary, existing methods lack underwater high-definition visual inspection and intelligent image recognition capabilities, making it impossible to quantitatively count the number, size, and distribution density of bubbles, and thus unable to provide a refined and quantitative assessment of aeration status. Furthermore, they fail to achieve multi-sensor data fusion and adaptive control under various operating conditions, resulting in significant monitoring errors when facing conditions such as high water turbidity, low winter temperatures, and fluctuations in influent water quality. They also lack a mechanism for optimizing aeration energy consumption based on historical data, leading to only extensive control methods, resulting in high energy consumption and high maintenance costs. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a wastewater treatment plant aeration identification method based on an underwater robot, enabling unmanned intelligent inspection of the entire underwater aeration tank of the wastewater treatment plant, and ultimately achieving the invention's objectives of intelligent fault diagnosis, accurate status assessment, energy consumption optimization, and automated operation and maintenance management of the aeration system.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A wastewater treatment plant aeration identification method based on an underwater robot includes the following steps:

[0010] Step S1: Inspection route planning and multi-source data collection;

[0011] Step S2: Underwater image preprocessing and target segmentation and recognition;

[0012] Step S3: Quantitative calculation of aeration status and comparison with operating condition thresholds;

[0013] Step S4: Bubble motion trajectory analysis and accurate anomaly identification;

[0014] Step S5: Cross-validation of multi-source data and adaptive processing of extreme conditions;

[0015] Step S6: Results visualization output and intelligent optimization of aeration energy consumption.

[0016] Preferably, step S1 specifically includes:

[0017] Based on the structure of the aeration tank, a pre-planned grid inspection path is established, and the grid size, inspection speed, and sampling parameters are set.

[0018] The underwater robot uses inertial navigation and visual SLAM fusion positioning to autonomously cruise along a preset path, simultaneously collecting high-definition underwater images, water depth data and dissolved oxygen data in real time, and marking the three-dimensional coordinates of each sampling point.

[0019] The preferred grid-based pre-defined inspection path planning is as follows:

[0020] The grid size is set according to the installation spacing of the aeration heads, ensuring that each grid contains at least one complete aeration head;

[0021] For inspection speed, the underwater robot's cruising speed is set to 0.2~0.5m / s; 0.3m / s for normal inspection; and reduced to 0.2m / s for abnormal verification or high turbidity conditions.

[0022] Sampling parameters: image acquisition frame rate 5~10 fps; depth data sampling frequency 20Hz; dissolved oxygen sensor sampling frequency 1Hz;

[0023] The path arrangement adopts a "bow" shaped coverage strategy, starting from one end of the aeration tank and scanning line by line along the length direction, with each line spacing equal to the grid width; the inter-row overlap rate is ≥10%; after the inspection is completed, automatic resampling around the edge of the tank wall is performed.

[0024] Preferably, a fusion localization framework is used to fuse inertial navigation and visual SLAM for localization. The specific process is as follows:

[0025] (1) Inertial navigation provides three-axis acceleration and angular velocity, and integrates to obtain position, velocity and attitude;

[0026] (2) Visual SLAM: ORB-SLAM3 improved version is used to extract ORB feature points of underwater images, match features of previous and next frames, and estimate the relative pose of the underwater robot by bundle adjustment.

[0027] (3) Fusion state quantity, define the EKF state vector as ; wherein, is a three-dimensional position coordinate, is a movement speed in three coordinate axis directions, is an Euler angle, is an accelerometer and gyroscope zero offset;

[0028] (4) Prediction step, drive state prediction with IMU measurement value, update covariance matrix;

[0029] (5) Update step, when obtaining the key frame pose estimation of visual SLAM, take it as an observation value, calculate the Kalman gain, correct the IMU integral drift and re-estimate the zero offset;

[0030] (6) Underwater adaptation, introduce a depth sensor as an absolute constraint on the z axis to suppress vertical drift.

[0031] Preferably, the step S2 specifically comprises:

[0032] The collected underwater images are sequentially subjected to dark channel defogging, adaptive histogram equalization contrast enhancement, and water turbidity interference is weakened;

[0033] The enhanced underwater images are input into an improved Mask R-CNN segmentation model to complete pixel-level segmentation of three types of targets, i.e., bubbles, aeration heads, and sludge flocs, and output target mask regions and class labels.

[0034] Preferably, the process of the improved Mask R-CNN segmentation model is as follows:

[0035] (1) Input the enhanced image, and normalize the size to 1024x1024 pixels;

[0036] (2) Feature extraction network, use ResNet-101-FPN as the backbone network, and output multi-scale feature maps;

[0037] (3) Region proposal network, generate anchor boxes on each feature map pixel, and output candidate target regions;

[0038] (4) ROI Align, accurately map each candidate region to a feature map, and output a fixed-size feature block;

[0039] (5) Parallel branch, including:

[0040] Classification branch, predicting the class, i.e., three types of bubbles, aeration heads, and sludge flocs;

[0041] Bounding box regression, refining the position of the candidate box;

[0042] The mask branch generates a 28×28 pixel binary mask through a fully convolutional network;

[0043] (6) Mask upsampling: The 28×28 mask is restored to the original ROI size through bilinear interpolation to achieve the category assignment of each pixel.

[0044] Preferably, step S3 specifically includes:

[0045] The characteristic parameters of bubble quantity, average size, and area ratio in each sampling area are statistically analyzed, and the bubble density calculation results are corrected by combining real-time depth data.

[0046] Based on a preset aeration status feature library, it matches differentiated aeration intensity thresholds according to different process sections and water depths. At the same time, it is compatible with various types of aeration heads and has built-in exclusive identification templates to quantitatively determine the aeration status of normal aeration, insufficient aeration, excessive aeration, and aeration head blockage.

[0047] Preferably, the calculation of bubble density and depth correction includes:

[0048] (1) Initial bubble density:

[0049] Within the image mask area, the total area of ​​the bubble pixels is counted. Divide by the corresponding pool bottom area to obtain the apparent bubble coverage rate. ;

[0050] (2) Depth correction factor:

[0051] Let the water depth at the sampling point be... As water depth increases, the volume of bubbles shrinks due to pressure, resulting in a lower actual escape frequency, but a visually increased density in the image; therefore, reverse correction is required.

[0052] ;

[0053] in, , which is standard atmospheric pressure; , where is the density of water; , is the acceleration due to gravity; For reference water depth;

[0054] (3) Actual implementation:

[0055] The robot records each sampling point ;

[0056] Original bubble coverage Multiply by the pressure ratio to obtain the true coverage after depth correction. ;

[0057] Then, the bubble number density is calculated based on the average area of ​​a single bubble.

[0058] (4) Threshold correction: when When the depth is greater than 5m, the pressure ratio is less than 0.67.

[0059] Preferably, step S4 specifically includes:

[0060] Bubble tracking and matching are performed by selecting multiple consecutive frames of images to analyze the rising speed and diffusion range of bubbles, identify hidden faults in the aeration area, quantify the degree of blockage of the aeration head, and complete high-precision positioning.

[0061] Preferably, step S5 specifically includes:

[0062] The image recognition results, dissolved oxygen measurement data, and SCADA system blower operation data are compared and correlated three times. If the deviation exceeds the preset threshold, a second verification is automatically triggered.

[0063] For extreme conditions with high turbidity in water, a multi-frame overlay enhancement and noise reduction algorithm is automatically activated to ensure the accuracy of target recognition and the reliability of detection in complex water environments.

[0064] Preferably, step S6 specifically includes:

[0065] Automatically generate thermal maps of aeration tanks, anomaly point markings, and structured inspection reports, and push them to the operation and maintenance platform in a standardized format;

[0066] An aeration energy consumption optimization model is constructed by linking long-term aeration status, treated water volume, and effluent water quality data. The aeration volume of each area is dynamically adjusted according to the influent water quality, and the aeration control strategy is adaptively adjusted for low-temperature conditions in winter. This ensures that the effluent water quality meets the standards while achieving energy-saving and intelligent operation of the aeration system.

[0067] Preferably, the input to the aeration energy consumption optimization model is historical data and real-time data;

[0068] The aeration energy consumption optimization model defines the zone aeration efficiency as follows:

[0069] ;

[0070] in, For the first The oxygen mass transfer coefficient of the partition was calibrated by the DO rise rate experiment; The aeration power allocated to this zone;

[0071] The objective function of the aeration energy consumption optimization model is:

[0072] ;

[0073] The constraints are: , And the effluent water quality meets the standards;

[0074] in, This represents the total number of zones in the aeration tank.

[0075] The wastewater treatment plant aeration identification method based on an underwater robot provided by this invention has the following beneficial effects:

[0076] (1) Overcome the shortcomings of only being able to observe the surface and not being able to see the actual underwater working conditions.

[0077] Existing technologies rely solely on visual inspection from the water surface, making it impossible to detect hidden faults such as early blockages and localized damage to underwater aerators. This invention employs an underwater robot to dive to the bottom of the aeration tank for comprehensive inspection, directly capturing close-up underwater images of the aerators to intuitively understand their actual working status and identify early blockage faults that traditional methods may miss.

[0078] (2) Overcome the shortcomings of single-point monitoring, incomplete coverage, and inability to quantify the entire area.

[0079] Traditional dissolved oxygen monitoring relies on single-point measurement, and blower airflow only reflects the total amount, failing to capture localized differences in aeration within the tank. This invention utilizes robotic grid path sampling for full coverage, combined with machine vision pixel-level segmentation, to quantitatively analyze the number, size, and distribution density of bubbles. This enables a refined and quantitative assessment of the aeration status across the entire tank, significantly improving data granularity and coverage.

[0080] (3) Overcome the shortcomings of lack of precise positioning and inability to determine the degree of blockage.

[0081] Existing technologies can only qualitatively determine the quality of aeration, but cannot locate abnormal points or quantify the degree of blockage. This invention integrates inertial navigation and visual SLAM high-precision positioning, combined with continuous frame bubble trajectory analysis, to accurately locate abnormal locations and determine the degree of blockage in the aeration head, achieving fault location and quantification.

[0082] (4) Overcome the shortcomings of identification failure under water depth difference, high turbidity, and low temperature conditions.

[0083] Traditional methods do not consider the impact of water depth on bubble detection, resulting in severe monitoring distortion under high turbidity and low temperature environments. This invention sets differentiated aeration thresholds according to process sections and water depths, and uses depth data to adaptively correct the calculation results; it activates a multi-frame overlay enhancement algorithm under high turbidity conditions, and automatically adjusts the control threshold in low winter temperatures, significantly improving the accuracy and stability of identification under complex and extreme conditions.

[0084] (5) Overcome the shortcomings of isolated data, lack of cross-validation, and low reliability of results.

[0085] Existing image, dissolved oxygen, and blower data are independent and lack a verification mechanism. This invention achieves cross-comparison of multi-source data from image recognition, dissolved oxygen measurement, and SCADA blower data, automatically performs secondary verification when thresholds are exceeded, eliminates environmental interference, and improves the reliability of test results.

[0086] (6) Overcome the shortcomings of extensive control, lack of intelligent optimization, and high aeration energy consumption.

[0087] Traditional methods only monitor aeration and cannot dynamically adjust aeration volume based on water quality conditions, resulting in significant energy waste. This invention establishes an aeration energy consumption optimization model that dynamically allocates aeration volume to each area by linking long-term water volume and quality data. It can also adaptively optimize control even at low temperatures in winter. While ensuring that the effluent water quality meets standards, it significantly reduces aeration power consumption and operation and maintenance costs, achieving an integrated technology effect of intelligent monitoring, fault diagnosis, and energy-saving optimization. Attached Figure Description

[0088] Figure 1 This is a schematic flowchart of the wastewater treatment plant aeration identification method of the present invention. Detailed Implementation

[0089] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0090] Combination Figure 1 As shown, this invention provides a wastewater treatment plant aeration identification method based on an underwater robot, comprising the following steps:

[0091] Step S1, inspection path planning and multi-source data collection, specifically includes:

[0092] Based on the structure of the aeration tank, a pre-planned grid pattern and preset inspection path are established, and the grid size, inspection speed, and sampling parameters are set.

[0093] The grid-based preset inspection path planning is as follows:

[0094] The grid size should be set according to the aeration head installation spacing (usually 0.5~1.5m). A 0.8m×0.8m square grid is recommended, ensuring that each grid contains at least one complete aeration head. For the edge area of ​​the pool wall, the grid can be reduced to 0.5m×0.5m to improve boundary coverage.

[0095] For inspection speed, the underwater robot's cruising speed is set to 0.2~0.5m / s; 0.3m / s for normal inspection; and reduced to 0.2m / s for abnormal verification or high turbidity conditions.

[0096] Sampling parameters: image acquisition frame rate is 5~10 fps (automatically adjusted according to lighting conditions, reduced to 3 fps in low light); depth data sampling frequency is 20Hz; dissolved oxygen (DO) sensor sampling frequency is 1Hz (matched with spatial coordinate interpolation).

[0097] The path arrangement adopts a "bow" shaped coverage strategy, starting from one end of the aeration tank and scanning line by line along the length direction, with each line spacing equal to the grid width; the inter-row overlap rate is ≥10%; after the inspection is completed, automatic resampling around the edge of the tank wall is performed.

[0098] The underwater robot uses inertial navigation and visual SLAM fusion positioning, and is equipped with an underwater high-definition camera, depth sensor, and optical dissolved oxygen probe. It autonomously cruises and moves at a constant speed along a preset path, simultaneously collecting underwater high-definition images, water depth data and dissolved oxygen data in real time, and marking the three-dimensional coordinates of each sampling point to achieve multi-dimensional and full-area three-dimensional perception of the aeration area.

[0099] A fusion localization framework is used to fuse inertial navigation and visual SLAM for localization. The specific process is as follows:

[0100] (1) Inertial navigation (IMU): Provides three-axis acceleration (±2g) and angular velocity (±250° / s), and integrates to obtain position, velocity, and attitude (frequency 100 Hz). It has high short-term accuracy, but drift exists.

[0101] (2) Visual SLAM: An improved version of ORB-SLAM3 is used to extract ORB feature points from underwater images (approximately 300 per frame), match features from previous and subsequent frames, and estimate the robot's relative pose (frequency 5 Hz) using bundle adjustment (BA). Absolute drift correction is provided.

[0102] (3) Fusion state variables: Define the EKF state vector as ,in, For three-dimensional position coordinates, The velocity of motion in the three coordinate axes. Euler angles, Zero bias for accelerometers and gyroscopes.

[0103] (4) Prediction steps: Use IMU measurements to drive state prediction and update the covariance matrix (frequency 100 Hz).

[0104] (5) Update steps: When the keyframe pose estimation of visual SLAM is obtained, it is used as the observation value, the Kalman gain is calculated, the IMU integral drift is corrected and the zero bias is re-estimated.

[0105] (6) Underwater adaptation: A depth sensor (accuracy ±0.01m) is introduced as an absolute constraint on the z-axis to suppress vertical drift. The final positioning error is ≤ ±0.05m (relative to the starting point).

[0106] Step S2, underwater image preprocessing and target segmentation and recognition, specifically includes:

[0107] The acquired underwater images were sequentially subjected to dark channel defogging, adaptive histogram equalization, and contrast enhancement to reduce the interference of water turbidity.

[0108] The enhanced underwater image is input into the improved Mask R-CNN segmentation model to complete pixel-level segmentation of three types of targets: bubbles, aeration heads, and sludge flocs, and outputs the target mask region and category label.

[0109] The improved pixel-level segmentation process of the Mask R-CNN segmentation model is as follows:

[0110] (1) Input enhanced image: size normalized to 1024×1024 pixels.

[0111] (2) Feature extraction network: ResNet-101-FPN is used as the backbone network to output multi-scale feature maps (P2~P5).

[0112] (3) Region Proposal Network (RPN): Generates anchor boxes (9 scales and aspect ratios) on each feature map pixel and outputs candidate target regions (about 2000).

[0113] (4) ROI Align: Maps each candidate region precisely to the feature map and outputs a fixed-size (14×14) feature block.

[0114] (5) Parallel branching:

[0115] Classification branches: predicted categories (bubbles, aeration heads, sludge flocs, three categories).

[0116] Bounding box regression: Refine the position of candidate boxes.

[0117] Masking branch: Generates a 28×28 pixel binary mask through a fully convolutional network (FCN).

[0118] (6) Mask upsampling: The 28×28 mask is restored to the original ROI size (e.g., 112×112) through bilinear interpolation to achieve the category assignment of each pixel.

[0119] The specific numerical definition of "pixel level":

[0120] Pixel-level segmentation refers to assigning a category label to each pixel in an image (e.g., pixel coordinates (i,j) are classified as "bubble").

[0121] There is no uniform absolute value; instead, it depends on the segmentation accuracy metric.

[0122] Mask cross-union ratio (mIoU): ≥0.85 (for small bubble-type targets, ≥0.75 is required).

[0123] Pixel accuracy (PA): The percentage of correctly classified pixels out of the total number of pixels, with a target of ≥0.95.

[0124] Boundary error (average contour distance): ≤2 pixels.

[0125] Step S3, quantitative calculation of aeration status and comparison with operating condition thresholds, specifically includes:

[0126] The characteristic parameters of bubble quantity, average size, and area ratio in each sampling area are statistically analyzed, and the bubble density calculation results are corrected by combining real-time depth data.

[0127] Based on a preset aeration status feature library, it matches differentiated aeration intensity thresholds according to different process sections and water depths. At the same time, it is compatible with various types of aeration heads and has built-in exclusive identification templates to quantitatively determine the aeration status of normal aeration, insufficient aeration, excessive aeration, and aeration head blockage.

[0128] A feature library of aeration states was constructed, and differentiated aeration intensity thresholds were set for different process sections and water depths. The judgment criteria were adaptively corrected by combining depth sensor data to eliminate the interference of water depth differences on the statistical results of bubble density.

[0129] Step S4, Bubble Motion Trajectory Analysis and Precise Anomaly Identification, specifically includes:

[0130] By introducing bubble motion trajectory analysis, selecting multiple consecutive frames of images for bubble tracking and matching, analyzing the bubble rising speed and diffusion range, identifying hidden faults in the aeration area, quantifying the degree of blockage of the aeration head, and completing high-precision location positioning.

[0131] Bubble density calculation and depth correction formula:

[0132] (1) Initial bubble density:

[0133] Within the image mask area, the total area of ​​the bubble pixels is counted. Divide by the corresponding pool bottom area (obtained from camera calibration, unit: m²) to obtain the apparent bubble coverage. (unit:%).

[0134] (2) Depth correction factor:

[0135] Let the water depth at the sampling point be... (Unit: m). As water depth increases, the volume of bubbles shrinks due to pressure, resulting in a lower actual escape frequency, but a visually increased density in the image. Therefore, reverse correction is required.

[0136] ;

[0137] in:

[0138] (Standard atmosphere)

[0139] (Density of water)

[0140] (Acceleration due to gravity)

[0141] For reference water depth (usually the installation depth of the aeration head, such as 3m),

[0142] (3) Actual implementation:

[0143] The robot records h (from the depth sensor) for each sampling point.

[0144] Original bubble coverage Multiply by the pressure ratio to obtain the true coverage after depth correction. .

[0145] Then, convert the average area of ​​a single bubble (calibrated value, such as 2 mm²) into the bubble number density (bubbles / m²).

[0146] (4) Threshold correction: when When the pressure ratio is greater than 5m and the pressure ratio is less than 0.67, the density can be reduced by more than 30% after correction, thus avoiding misjudgment of "overexposure".

[0147] Step S5, multi-source data cross-validation and extreme condition adaptive processing, specifically includes:

[0148] The image recognition results, dissolved oxygen measurement data, and SCADA system blower operation data are compared and correlated three times. If the deviation exceeds the preset threshold, a second verification is automatically triggered.

[0149] Judgment index: Bubble coverage after depth correction (Unit: %) and dissolved oxygen concentration (DO) (mg / L).

[0150] Examples of differentiated thresholds for different process sections (taking aerobic zone 1 as an example):

[0151]

[0152] For extreme conditions with high turbidity in water, a multi-frame overlay enhancement and noise reduction algorithm is automatically activated to ensure the accuracy of target recognition and the reliability of detection in complex water environments.

[0153] Step S6, result visualization output and intelligent optimization of aeration energy consumption, specifically includes:

[0154] Automatically generate thermal maps of aeration tanks, anomaly point markings, and structured inspection reports, and push them to the operation and maintenance platform in a standardized format;

[0155] An aeration energy consumption optimization model is constructed by linking long-term aeration status, treated water volume, and effluent water quality data. The aeration volume of each area is dynamically adjusted according to the influent water quality, and the aeration control strategy is adaptively adjusted for low-temperature conditions in winter. This ensures that the effluent water quality meets the standards while achieving energy-saving and intelligent operation of the aeration system.

[0156] The aeration energy consumption optimization model is a dynamic programming model based on long-term correlation data.

[0157] Input to the aeration energy consumption optimization model:

[0158] Historical data: Aeration tank zoning status (bubble coverage, DO), treatment volume Qin (m³ / h), influent COD, ammonia nitrogen NH3−N, effluent water quality (COD, TN, NH3−N), and total blower power Pblower (kW) for the past 30 days.

[0159] Real-time data: Aeration status of each grid and DO value of each zone obtained from the robot's current inspection.

[0160] Model structure:

[0161] (1) Aeration demand prediction sub-model:

[0162] Random forest regression is used, with input (influent water quality + water quantity + water temperature), and output the target DO value and minimum required bubble coverage for each process section. Training labels are drawn from historical operating conditions of "compliant effluent + lowest energy consumption".

[0163] (2) Characterization of aeration efficiency:

[0164] Define zone aeration efficiency ,in:

[0165] For the first The oxygen mass transfer coefficient of the zone is calibrated by the DO rise rate experiment (or estimated by correlating the DO sensor with the bubble coverage).

[0166] The aeration power allocated to this zone (calculated based on air volume × pressure × efficiency).

[0167] (3) Optimize the objective function:

[0168] ;

[0169] The constraints are: , Moreover, the quality of the effluent meets the standards.

[0170] in: The number of partitions (e.g., merging into 20-50 control zones by grid).

[0171] (4) Optimization algorithm: Genetic algorithm or particle swarm optimization (PSO) is adopted, and the decision variable is the valve opening degree of each partition (0~100% continuous). The penalty function method is used for constraint handling.

[0172] (5) Dynamic adjustment strategy:

[0173] Routine adjustment: Optimization is performed every 15 minutes, and valve opening commands are output to the SCADA system.

[0174] Winter low temperature adaptation:

[0175] At low temperatures (water temperature < 10℃), oxygen mass transfer efficiency decreases by 30-50%. The model automatically increases the target DO by 0.5-1.0 mg / L, increases the target bubble coverage by 20%, and increases the lower limit of blower frequency to prevent sludge bulking.

[0176] At the same time, the "low temperature pulse aeration" mode is activated: normal aeration for 15 minutes, stop for 5 minutes, and run alternately, which saves energy and prevents the pool surface from freezing.

[0177] Online model learning: The random forest is retrained weekly using new data, and efficiency parameters are updated. Bayesian optimization is used to tune hyperparameters (number of trees). (e.g., depth, etc.)

[0178] In summary, this invention offers a comprehensive detection range, relying on an underwater robot for full-area underwater inspection, overcoming the limitations of traditional surface observation and single-point monitoring. It can detect hidden faults such as early blockage and localized damage to underwater aerators; it also boasts high identification accuracy, employing an improved mask... R-CNN pixel-level segmentation enables quantitative statistics on bubble quantity, size, and density, achieving precise quantitative assessment of aeration status and high fault location accuracy. It boasts strong environmental adaptability, allowing for differentiated threshold settings based on different process sections and water depths, and adaptive correction for water depth interference. It maintains high recognition accuracy even under extreme conditions such as high turbidity and low winter temperatures, demonstrating strong anti-interference capabilities. Data reliability is high, with multi-source data fusion and cross-comparison, and automatic secondary verification, avoiding misjudgments from single monitoring methods and ensuring reliable detection results. It offers good versatility, with built-in multiple aeration head recognition templates, adapting to various aeration devices such as microporous, disc, and tubular aeration equipment without secondary development, making it suitable for a wide range of scenarios. Maintenance costs are low, with fully automated inspection, intelligent analysis, and automatic report generation, reducing the workload of manual on-site inspections. Energy savings are significant, dynamically optimizing aeration volume, reducing aeration energy consumption by 10%~15% under normal operating conditions and an additional reduction of over 8% under low-temperature conditions, resulting in substantial electricity savings and stable effluent quality meeting standards.

[0179] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A wastewater treatment plant aeration identification method based on an underwater robot, characterized in that, Includes the following steps: Step S1: Inspection route planning and multi-source data collection; Step S2: Underwater image preprocessing and target segmentation and recognition; Step S3: Quantitative calculation of aeration status and comparison with operating condition thresholds; Step S4: Bubble motion trajectory analysis and accurate anomaly identification; Step S5: Cross-validation of multi-source data and adaptive processing of extreme conditions; Step S6: Results visualization output and intelligent optimization of aeration energy consumption.

2. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 1, characterized in that, Step S1 specifically includes: Based on the structure of the aeration tank, a pre-planned grid inspection path is established, and the grid size, inspection speed, and sampling parameters are set. The underwater robot uses inertial navigation and visual SLAM fusion positioning to autonomously cruise along a preset path, simultaneously collecting high-definition underwater images, water depth data and dissolved oxygen data in real time, and marking the three-dimensional coordinates of each sampling point.

3. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 2, characterized in that, The grid-based preset inspection path planning is as follows: The grid size is set according to the installation spacing of the aeration heads, ensuring that each grid contains at least one complete aeration head; The inspection speed of the underwater robot is set to 0.2~0.5 m / s. For normal inspections, the velocity is 0.3 m / s; for abnormal checks or high turbidity conditions, the velocity is reduced to 0.2 m / s. Sampling parameters: image acquisition frame rate 5~10 fps; depth data sampling frequency 20Hz; dissolved oxygen sensor sampling frequency 1Hz; The path arrangement adopts a "bow" shaped coverage strategy, starting from one end of the aeration tank and scanning line by line along the length direction, with each line spacing equal to the grid width; the inter-line overlap rate is ≥10%; after the inspection is completed, automatic resampling around the edge of the tank wall is performed.

4. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 2, characterized in that, A fusion localization framework is used to fuse inertial navigation and visual SLAM for localization. The specific process is as follows: (1) Inertial navigation provides three-axis acceleration and angular velocity, and integrates to obtain position, velocity and attitude; (2) Visual SLAM: ORB-SLAM3 improved version is used to extract ORB feature points of underwater images, match features of previous and next frames, and estimate the relative pose of the underwater robot by bundle adjustment. (3) Fusion of state variables, defining the EKF state vector as follows: ;in, For three-dimensional position coordinates, The velocity of motion in the three coordinate axes. Euler angles, Zero bias for accelerometers and gyroscopes; (4) Prediction step: IMU measurements drive state prediction and update the covariance matrix; (5) Update step: When the keyframe pose estimation of visual SLAM is obtained, it is used as the observation value, the Kalman gain is calculated, the IMU integral drift is corrected and the zero bias is re-estimated; (6) Underwater adaptation: a depth sensor is introduced as an absolute constraint on the z-axis to suppress vertical drift.

5. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 1, characterized in that, Step S2 specifically includes: The acquired underwater images were sequentially subjected to dark channel defogging, adaptive histogram equalization, and contrast enhancement to reduce the interference of water turbidity. The enhanced underwater image is input into the improved Mask R-CNN segmentation model to complete pixel-level segmentation of three types of targets: bubbles, aeration heads, and sludge flocs, and outputs the target mask region and category label.

6. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 5, characterized in that, The process of the improved Mask R-CNN segmentation model is as follows: (1) Input an enhanced image, with the size normalized to 1024×1024 pixels; (2) Feature extraction network: ResNet-101-FPN is used as the backbone network to output multi-scale feature maps; (3) Region Proposal Network generates anchor boxes on each feature map pixel and outputs candidate target regions; (4) ROI Align: precisely maps each candidate region to the feature map and outputs a fixed-size feature block; (5) Parallel branches, including: The classification branches predict the categories, namely, bubbles, aeration heads, and sludge flocs. Bounding box regression, refining candidate box positions; The mask branch generates a 28×28 pixel binary mask through a fully convolutional network; (6) Mask upsampling: The 28×28 mask is restored to the original ROI size through bilinear interpolation to achieve the category assignment of each pixel.

7. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 1, characterized in that, Step S3 specifically includes: The characteristic parameters of bubble quantity, average size, and area ratio in each sampling area are statistically analyzed, and the bubble density calculation results are corrected by combining real-time depth data. Based on a preset aeration status feature library, it matches differentiated aeration intensity thresholds according to different process sections and water depths. At the same time, it is compatible with various types of aeration heads and has built-in exclusive identification templates to quantitatively determine the aeration status of normal aeration, insufficient aeration, excessive aeration, and aeration head blockage.

8. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 7, characterized in that, The calculation of bubble density and depth correction includes: (1) Initial bubble density: Within the image mask area, the total area of ​​the bubble pixels is counted. Divide by the corresponding pool bottom area to obtain the apparent bubble coverage rate. ; (2) Depth correction factor: Let the water depth at the sampling point be... As water depth increases, the volume of bubbles shrinks due to pressure, resulting in a lower actual escape frequency, but a visually increased density in the image; therefore, reverse correction is required. ； in, , which is standard atmospheric pressure; , where is the density of water; , is the acceleration due to gravity; For reference water depth; (3) Actual implementation: The robot records each sampling point ; Original bubble coverage Multiply by the pressure ratio to obtain the true coverage after depth correction. ; Then, the bubble number density is calculated based on the average area of ​​a single bubble. (4) Threshold correction: when When the depth is greater than 5m, the pressure ratio is less than 0.

67.

9. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 1, characterized in that, Step S4 specifically includes: Bubble tracking and matching are performed by selecting multiple consecutive frames of images to analyze the rising speed and diffusion range of bubbles, identify hidden faults in the aeration area, quantify the degree of blockage of the aeration head, and complete high-precision positioning.

10. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 1, characterized in that, Step S5 specifically includes: The image recognition results, dissolved oxygen measurement data, and SCADA system blower operation data are compared and correlated three times. If the deviation exceeds the preset threshold, a second verification is automatically triggered. For extreme conditions with high turbidity in water, a multi-frame overlay enhancement and noise reduction algorithm is automatically activated to ensure the accuracy of target recognition and the reliability of detection in complex water environments.

11. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 1, characterized in that, Step S6 specifically includes: Automatically generate thermal maps of aeration tanks, anomaly point markings, and structured inspection reports, and push them to the operation and maintenance platform in a standardized format; An aeration energy consumption optimization model is constructed by linking long-term aeration status, treated water volume, and effluent water quality data. The aeration volume of each area is dynamically adjusted according to the influent water quality, and the aeration control strategy is adaptively adjusted for low-temperature conditions in winter. This ensures that the effluent water quality meets the standards while achieving energy-saving and intelligent operation of the aeration system.

12. The wastewater treatment plant aeration identification method based on an underwater robot according to claim 11, characterized in that, The input to the aeration energy consumption optimization model is historical data and real-time data; The aeration energy consumption optimization model defines the zone aeration efficiency as follows: ； in, For the first The oxygen mass transfer coefficient of the partition was calibrated by the DO rise rate experiment; The aeration power allocated to this zone; The objective function of the aeration energy consumption optimization model is: ； The constraints are: , And the effluent water quality meets the standards; in, This represents the total number of zones in the aeration tank.

Images