Robotic intelligent control system for energy saving and efficiency improvement at the cold end of indirect cooling auxiliary units

Abstract

This invention discloses a robotic intelligent control system for improving energy efficiency at the cold end of an indirect-cooled auxiliary unit, belonging to the field of intelligent control technology. It includes an initial evaluation module, a dynamic planning module, and a feedback monitoring module. The initial evaluation module performs a global preliminary scan to obtain an initial planning image set. Combined with the fin geometric model, it obtains a scale thickness distribution map and scale coverage rate, generating a fin flow channel matrix model. The dynamic planning module divides virtual work blocks, constructs a scale planning path based on a cleaning index, performs rapid coarse cleaning, and provides cavitation early warning. A hierarchical control architecture is used to handle cavitation early warnings. The feedback monitoring module constructs a quality deviation set, filters out substandard virtual work blocks, calculates correction priorities, sorts and ranks them, optimizes correction parameters using a genetic algorithm, and performs feedback regulation. This improves cold-end heat exchange efficiency to achieve energy-saving goals while protecting the fin structure and extending equipment lifespan.

Description

Technical Field

[0001] This invention relates to a robotic intelligent control system for energy saving and efficiency improvement at the cold end of an indirect cooling auxiliary unit, belonging to the field of intelligent control technology. Background Technology

[0002] With the rapid development of industrial automation and intelligent manufacturing, mobile robots have been widely used in various industrial scenarios due to their advantages of automation and intelligent operation. These include the cleaning of cold-end fins in intermediate cooling auxiliary units of thermal power plants. The cold end of the intermediate cooling auxiliary unit is the core link that determines the energy utilization efficiency of the unit. As a key heat exchange component, the fin array of its cold-end equipment is exposed to the complex outdoor environment for a long time. It is prone to heat exchange efficiency decline due to problems such as scale accumulation and corrosion, which directly affects the energy-saving and efficiency-enhancing capabilities of the cold-end system.

[0003] However, existing technologies for cleaning the cold-end fins of indirect-cooled auxiliary units do not fully consider the issues of sensitive control precision and poor cleaning effect stability during the cleaning process. Specifically, if cavitation bubble cleaning technology is used, the generation of cavitation bubbles depends on sudden changes in water pressure and must be below the saturated vapor pressure of the liquid. However, the flow channel structure of the fin matrix is ​​very complex, which can easily lead to local water pressure fluctuations. For example, if there are differences in the spacing between different fin groups, without real-time compensation for water pressure fluctuations, some fins may be damaged due to over-cleaning or have scale residue due to insufficient cleaning, directly reducing the heat exchange efficiency of the fins. In addition, cavitation bubble cleaning requires simultaneous coordination of multiple parameters such as nozzle movement speed, water pressure, and water temperature. The coupling relationship between these parameters is quite complex. If any parameter deviates from the appropriate range, the stability of the cavitation bubbles will decrease significantly, leading to a significant reduction in the overall cleaning effect. Ultimately, this will adversely affect the safe and efficient operation of the unit. It is difficult to meet the energy-saving and efficiency-enhancing requirements of the cold-end system, and it is also difficult to maximize the energy utilization efficiency of the cold-end system through intelligent control while ensuring the quality of fin cleaning. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a robotic intelligent control system for energy saving and efficiency improvement at the cold end of indirect cooling auxiliary units. This system constructs a precise finned flow channel matrix model through multimodal calibration and scanning to resolve data baseline deviations. It divides virtual work blocks according to scaling and geometric attributes, plans priority paths to avoid resource misallocation, monitors cavitation in real time, and implements layered decoupled control, overcoming parameter coupling limitations. It also quantifies the quality of collected data, optimizes parameters through simulation to achieve secondary fine cleaning, replacing experience-based adjustments, and comprehensively solving problems related to control accuracy and performance stability.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The robotic intelligent control system for energy saving and efficiency improvement of the cold end of the indirect cooling auxiliary unit includes: an initial assessment module, a dynamic planning module, and a feedback monitoring module;

[0007] The initial evaluation module is used to drive the mobile robot to perform a global preliminary scan of the cold end fin matrix, obtain an initial planning image set, identify and segment the scaling area, and combine the constructed fin geometric model to obtain the scaling thickness distribution map and the overall scaling coverage, and generate a fin flow channel matrix model.

[0008] The dynamic programming module is used to divide virtual work blocks, calculate the cleaning index of each virtual work block, construct a scaling planning path, drive the mobile robot to perform rapid coarse cleaning, provide cavitation warning through dynamic cavitation threshold and cavitation parameters, and process cavitation warning through a hierarchical control architecture.

[0009] The feedback monitoring module is used to acquire a set of comparison images, construct a set of quality deviations, screen out virtual work blocks that do not meet the standards, calculate correction priorities and sort them by level, optimize correction parameters using a genetic algorithm, and perform feedback regulation.

[0010] Specifically, the dynamic programming module includes a cleaning planning unit and a decoupling control unit;

[0011] The cleaning planning unit is used to divide virtual work blocks, construct an attribute label set for each virtual work block and assign weights to each attribute, calculate the cleaning index by combining the corresponding weight coefficients, divide priority levels, introduce a priority factor to set a heuristic function, and plan the path using the A* algorithm.

[0012] The decoupled control unit is used to construct initial reference parameters, perform coarse scanning on the virtual work block, calculate geometric deviations in conjunction with the fin geometric model, generate local point clouds, construct local sub-models for the virtual work block, set dynamic cavitation thresholds for real-time cavitation monitoring, construct a hierarchical control architecture, and periodically evaluate the control effect.

[0013] Specifically, the steps of a preliminary global scan include:

[0014] The mobile robot is calibrated to standard specifications, and safe scanning boundaries are set, scanable grids are divided, and safe paths are generated.

[0015] Executing the security path triggers the multimodal sensor to collect data, acquiring the three-dimensional point cloud, texture image, and infrared thermal image of the fin matrix surface to generate an initial planning image set;

[0016] The initial planning image set is preprocessed to generate a global planning dataset;

[0017] Construct a triangular mesh model of the fin matrix, extract geometric parameters, simplify the mesh, and generate the fin geometric model.

[0018] Specifically, the steps of the global preliminary scan also include:

[0019] A scale recognition model is constructed, and a scale binary segmentation map is generated by combining the texture image;

[0020] The surface of the fin geometry model is divided into multiple grid blocks. The scale thickness of each grid block is calculated. The scale level is classified by a secondary thickness threshold, and a scale thickness distribution map is generated.

[0021] The total surface area of ​​the fin flow channel is obtained from the fin geometry model, and the total fouling area of ​​all fouling grid blocks is counted to calculate the overall fouling coverage rate.

[0022] Based on the scaling level, stratified sampling was performed to obtain multiple verification points, and the simulated values ​​of each verification point were recorded.

[0023] The mobile robot is used to obtain actual values ​​and calculate verification indicators. If the verification indicator of a certain verification point exceeds the verification threshold, the process returns to the corresponding step for optimization until all indicators meet the standards.

[0024] By integrating the fin geometry model, scale thickness distribution map, scale coverage, and global temperature distribution map, a fin flow channel matrix model is generated.

[0025] Specifically, the steps for planning a route include:

[0026] Based on the aforementioned finned flow channel matrix model, virtual work blocks are divided according to the principles of natural flow channel separation and scaling continuity to generate a virtual work set;

[0027] For each virtual job block The average thickness was calculated using the scale thickness distribution map. The scaling binary segmentation map is called to calculate the scaling area ratio. ;

[0028] Based on the aforementioned finned flow channel matrix model, the average flow channel spacing is calculated. Simultaneously, the total length of the flow channel and the number of flow channel corners are obtained, and combined with the cornerless DC channel, the flow channel complexity coefficient is calculated. .

[0029] Specifically, the steps for planning a route also include:

[0030] Construct virtual job blocks The attribute tag set, combined with the secondary threshold, is used to divide the scale thickness weight. Scale area weight Flow channel complexity weight ;

[0031] Set weighting coefficients, calculate cleaning index, sort the virtual job set, and divide it into priority levels, including high priority, medium priority, and low priority.

[0032] Introduce a priority weighting factor, set a heuristic function, and use the A* algorithm to plan the path and generate a scaling planning path.

[0033] The scaling planning path is imported into the finned flow channel matrix model to verify the feasibility of the path.

[0034] Specifically, the steps for real-time cavitation monitoring include:

[0035] Obtain the local coordinates of the current virtual job block to be processed, move to the guide positioning point, and obtain the initial reference parameters;

[0036] Set the scanning range, acquire a rapid coarse scan point cloud, and calculate the geometric deviation from the fin geometric model;

[0037] Once the geometric deviation exceeds the geometric deviation threshold, the scanning parameters are adjusted based on the coarse scan point cloud to perform a complete fine scan.

[0038] Acquire local point clouds, extract real-time geometric parameters, and generate local geometric features;

[0039] Calculate the thickness deviation, and if the thickness deviation exceeds a preset thickness deviation threshold, perform local fine-tuning to generate local scaling characteristics;

[0040] By combining the local geometric features and the local scaling features, a local sub-model of the current virtual work block is generated.

[0041] Specifically, the steps for real-time cavitation monitoring also include:

[0042] Cavitation bubble images under different working conditions are collected, an associated dataset is constructed, and based on the constructed cavitation bubble recognition model, real-time images during cleaning are obtained to generate cavitation bubble mask images.

[0043] Based on the scaling level of the virtual work block, the corresponding optimal cavitation interval is called as the dynamic cavitation threshold.

[0044] Calculate the total amount of cavitation bubbles, and combine this with the total area of ​​the jet's effective region to calculate the cavitation intensity index;

[0045] The cavitation uniformity index is calculated based on the ratio of the standard deviation of the cavitation bubble area to the average area of ​​the cavitation bubble.

[0046] By combining dynamic cavitation thresholds for real-time monitoring, a cavitation warning signal is triggered once the cavitation intensity index or cavitation uniformity index exceeds the dynamic cavitation threshold.

[0047] Specifically, the steps for real-time cavitation monitoring also include:

[0048] Set a look-ahead time window and extract look-ahead geometric parameters by combining local sub-models;

[0049] The local resistance of the virtual working block is calculated using the local resistance loss formula. Combined with the length of the look-ahead channel and the average channel spacing, the friction loss is calculated.

[0050] Calculate the forward drag loss, simultaneously obtain the current flow channel resistance, calculate the drag change, and thus generate the frequency adjustment amount;

[0051] Set a data acquisition cycle, collect cavitation parameter data and physical parameter data in real time, and analyze the correlation between physical parameters and cavitation parameters in real time;

[0052] Predict changes in cavitation state; if the predicted cavitation value deviates from the target value, initiate the optimization process.

[0053] The system automatically evaluates the effectiveness of each virtual work block after half of its cleaning area has been completed.

[0054] Specifically, the steps of feedback regulation include:

[0055] After the virtual job block is cleaned, multimodal data is collected in real time, and a comparison image set is constructed.

[0056] Obtain the scaling map, calculate the scaling removal rate of the virtual work block, and combine it with the scaling removal rate of a single grid to obtain the removal uniformity index;

[0057] Effective damage is screened based on damage depth, the proportion of damaged area is calculated, the heat exchange efficiency recovery rate is calculated based on temperature distribution, a quality deviation set is constructed, and virtual work blocks that do not meet the standards are screened out.

[0058] Calculate the correction priority index and combine it with the secondary correction threshold to generate a correction priority table;

[0059] Set the adjustable range and optimization target of the correction parameters, and use a genetic algorithm to iteratively optimize the operation variables to generate the combination of correction parameters for the operation variable sequence;

[0060] A secondary washing path is generated, and multimodal data collection and quality quantification are repeated to verify the optimization objectives.

[0061] The beneficial effects of this invention are:

[0062] By employing multimodal sensor calibration and data fusion, a finned flow channel matrix model integrating geometric, scaling, and temperature attributes is constructed. This addresses the baseline bias caused by traditional initial assessments focusing on only a single dimension, providing unified and reliable data support for all subsequent stages. Virtual work blocks are divided according to scaling and geometric attributes, and priority paths are planned to avoid resource occupation by low-urgency work blocks and reduce heat exchange losses caused by delayed cleaning in high-urgency areas. Through real-time imaging and adaptive decoupling control, the limitations of relying solely on water pressure to judge cavitation effects are overcome, achieving multi-parameter dynamic decoupling and dynamically adapting to flow channel changes and scaling requirements. This solves the cleaning instability issues caused by parameter coupling and environmental fluctuations, ensuring cleaning effectiveness while avoiding over-cleaning damage to the fins. Relying on twin simulation and algorithm optimization to generate correction parameters, replacing traditional experience-based adjustments, the system achieves precise cleaning and intelligent control. This not only improves cold-end heat exchange efficiency to achieve energy-saving goals but also protects the finned structure and extends equipment life, fully meeting the core requirements of cold-end energy saving and efficiency improvement for indirect-cooled auxiliary units. Attached Figure Description

[0063] Figure 1 Structure diagram of a robot-based intelligent control system for improving energy efficiency and reducing cold-end energy consumption in indirect cooling auxiliary units;

[0064] Figure 2 This is a flowchart of the global preliminary scan in this invention;

[0065] Figure 3 This is a flowchart of the path planning process in this invention;

[0066] Figure 4 This is a flowchart of the real-time monitoring of cavitation in this invention. Detailed Implementation

[0067] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0068] Example 1

[0069] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a robotic intelligent control system for energy saving and efficiency improvement at the cold end of an indirect cooling auxiliary unit, including: an initial evaluation module, a dynamic programming module, and a feedback monitoring module.

[0070] The initial evaluation module drives a mobile robot equipped with multimodal sensors to perform a global preliminary scan of the cold-end fin matrix along a preset initial scan path at a global scan speed. It acquires 3D point cloud data of the fin matrix using a 3D structured light camera and 2D texture images using a visible light camera to construct a preliminary framework model, including external dimensions, main flow channel orientation, and approximate fin spacing, thus fully presenting the overall geometric layout of the fins. Simultaneously, an infrared thermal imager is activated to acquire the temperature distribution field on the fin matrix surface and generate an infrared thermal image to reveal the uneven heat transfer caused by varying degrees of fouling on the fins, thereby generating an initial planning image set. Preprocessing is performed, and a fin geometric model with precise geometric dimensions is generated through a reconstruction algorithm. Feature extraction and analysis are performed on the acquired two-dimensional images. Based on a pre-trained deep learning model, the fouling area is identified and segmented. At the same time, combined with the fin geometric model, the initial fouling thickness distribution map and the overall fouling coverage are quantitatively calculated. This provides initial input and a benchmark for subsequent differentiated cleaning strategies. It solves the limitation of traditional initial assessments that only focus on fin geometry and ignore thermal performance. Through cross-dimensional perception, it provides a rich information base for subsequent accurate decision-making and avoids cleaning strategy deviations caused by incomplete understanding of the initial state.

[0071] Specifically, the steps of a preliminary global scan include:

[0072] The mobile robot is fixed at a standard calibration station, where a calibration plate for the fins and a scale test block are placed. Multimodal sensors, including a 3D structured light camera, an industrial visible light camera, and an infrared thermal imager, are integrated into the robot's end effector. The structured light data, 2D texture images, and infrared temperature data must be correlated in the same time and space. Otherwise, it is easy to cause misalignment between the scale texture and the 3D position, and the temperature anomaly points cannot correspond to specific fins. The multimodal sensors are connected to the same time synchronizer via industrial Ethernet for equipment calibration and calibration result verification. A set of synchronized data including the calibration plate and the scale test block is captured to verify that the equipment measurement deviation and time deviation are less than the preset deviation threshold. If the standard is not met, the calibration is repeated.

[0073] Obtain the CAD drawing of the cold end fin matrix of the indirect cooling auxiliary unit, set the safe scanning boundary of the mobile robot to avoid collisions, divide the fin matrix into several grid units of the same size, mark all scannable grids, exclude equipment occlusion areas, take the initial docking point of the mobile robot as the starting point and the diagonal vertex of the fin matrix as the ending point, use the A* algorithm to generate a safe path covering the scannable grids to ensure that the point cloud stitching is without misalignment;

[0074] The mobile robot is controlled to move along a safe path, pausing at each path node for a fixed scanning interval, and simultaneously triggering multimodal sensors to collect data. A 3D structured light camera collects the three-dimensional point cloud of the fin matrix surface, a visible light camera collects the two-dimensional texture image of the fin matrix surface, and an infrared thermal imager collects the temperature distribution of the fin matrix surface to generate an infrared thermal image. If the data at a certain node is abnormal, such as missing point cloud or blurred image, the mobile robot automatically returns to the node to re-collect data to ensure data integrity. The collected data from the fin matrix surface are integrated to generate an initial planning image set.

[0075] The initial planning image set is preprocessed, including: applying a statistical filtering algorithm to each group of 3D point clouds to remove noise points that deviate from the average distance of the point cloud cluster by more than 3 standard deviations, such as environmental dust and light interference points; using voxel mesh downsampling to compress the point cloud data volume while preserving fin channel details, such as spacing and tilt angle, to reduce subsequent computational pressure; using the extrinsic parameter matrix of the 3D structured light camera, employing the ICP iterative nearest point algorithm to stitch all path node downsampled point clouds into a global point cloud to eliminate node misalignment; applying Gaussian filtering to remove Gaussian noise from each texture image, and then using median filtering to eliminate salt-and-pepper noise, improving the clarity of the scale texture; using histogram equalization to enhance image contrast and highlight the grayscale difference between scaled areas and clean fins, such as low grayscale values ​​in scaled areas and high grayscale values ​​in clean areas; taking the average of 3 frames of infrared thermal imaging images for each path node to eliminate environmental infrared interference, such as local high temperatures caused by direct sunlight; calling the temperature calibration curve to convert the image grayscale values ​​into actual temperature values, generating a global temperature distribution map, and generating a global planning dataset;

[0076] Based on a global planning dataset, a Poisson surface reconstruction algorithm is used to generate a triangular mesh model of the fin matrix to preserve flow channel details during reconstruction, such as fin root fillets and flow channel corners. Through mesh segmentation and feature detection, key geometric parameters of the fins are extracted from the triangular mesh model. A mesh simplification algorithm is used to reduce the number of triangular meshes to generate a fin geometric model, ensuring that the model can be rotated and scaled in real time. Multiple fin feature points, such as fin tips and flow channel midpoints, are randomly selected, and their actual dimensions are measured using a laser rangefinder. The dimensions of the fin geometric model are compared with the actual dimensions. If the dimensional error does not exceed the preset value, the model is considered qualified.

[0077] A dataset of fin fouling images is obtained, containing multiple fin texture images with different degrees of fouling. The convolutional neural network model is pre-trained, such as the U-Net segmentation model. Simultaneously, the texture images in the global planning dataset are fine-tuned to adapt to the fouling texture features of the current unit's fins, such as fouling color and shape, thereby improving the accuracy of fouling area segmentation and generating a fouling recognition model.

[0078] The texture images in the global planning dataset are input into the scaling recognition model to perform semantic segmentation of the scaling area and generate a scaling binary segmentation map. White represents the scaling area and black represents the clean fin area. At the same time, the pixel coordinates of the scaling area in each image are marked.

[0079] The surface of the fin geometry model is divided into multiple uniformly sized grid blocks, each corresponding to a specific pixel region in the scale binary segmentation map. Using scale test blocks, the mapping relationship between texture grayscale and scale thickness is calibrated, and the scale thickness of each grid block is calculated. The thickness values ​​of all grid blocks are mapped to the corresponding positions in the fin geometry model to generate a scale thickness distribution map, which visually displays the scale thickness in different areas. At the same time, a two-level thickness threshold is used to classify the scale level, including light, moderate, and heavy, and the scale is marked on the scale thickness distribution map to achieve accurate identification and quantitative analysis of scale, providing data support for subsequent differentiated cleaning. Heavy scale areas are cleaned more intensively, while light scale areas are cleaned more simply to achieve energy saving and efficiency improvement.

[0080] The total surface area of ​​the fin flow channel is obtained from the fin geometry model, and the total fouling area of ​​all fouling grid blocks is counted. The overall fouling coverage rate is calculated by the ratio of the total fouling area to the total surface area. At the same time, the area ratio of each level of fouling under the total fouling area is counted according to the fouling level.

[0081] The fin geometry model, scale thickness distribution map, scale coverage, and global temperature distribution map are integrated into a fin flow channel matrix model to achieve visualization of the relationship between geometry, scale, and temperature. For example, clicking on a grid cell can display its geometric coordinates, scale thickness, and temperature value.

[0082] Based on the scaling level, stratified sampling is performed, randomly selecting multiple verification points from light, moderate, and heavy scaling areas. The geometric coordinates, scaling thickness, and temperature values ​​of the verification points in the digital twin are recorded. The mobile robot is controlled to move to the actual position of the verification point. A laser rangefinder is used to measure the geometric dimensions, an ultrasonic thickness gauge to measure the scaling thickness, and a high-precision infrared thermometer to measure the temperature. Verification indicators are calculated, including geometric deviation, scaling thickness deviation, and temperature deviation. If the verification indicator of a certain verification point exceeds the verification threshold, the process returns to the corresponding optimization step. For example, if the scaling thickness deviation exceeds the tolerance, the mapping relationship between grayscale and thickness is recalibrated. After optimization, the verification is repeated until all indicators meet the standards, thereby generating the final finned flow channel matrix model. Among them, the geometric deviation is the Euclidean distance between the model coordinates and the actual coordinates of the verification point, the scaling thickness deviation is the absolute difference between the model thickness and the actual thickness, and the temperature deviation is the absolute difference between the model temperature and the actual temperature.

[0083] The dynamic programming module is used to divide virtual work blocks, calculate the scaling and geometric properties of each virtual work block, calculate the cleaning index and divide the cleaning priority by combining the weight coefficients, use the physical boundary as a constraint and the cleaning priority to plan the path using the A* algorithm, generate the scaling planning path, control the robot to move along the scaling planning path to the guide positioning point of the work block, perform rapid coarse cleaning, calculate the geometric deviation of the fin-bonded model, generate local point cloud, and construct local sub-models of the virtual work blocks, set dynamic cavitation threshold, and perform cavitation early warning through the calculated cavitation parameters, and process the cavitation early warning through a hierarchical control architecture.

[0084] Specifically, the dynamic programming module includes a cleaning planning unit and a decoupling control unit;

[0085] The cleaning planning unit is used to divide virtual work blocks based on the finned flow channel matrix model, ensuring that each block has a single fouling level and a uniform flow channel structure, generating a virtual work set, and assigning a two-dimensional attribute label to each virtual work block, including fouling attribute and size attribute, generating an attribute label set for each virtual work block, dividing the weight of each attribute based on a secondary threshold, and calculating the cleaning index by combining the corresponding weight coefficient, quantifying the urgency of each virtual work block, and sorting them in descending order to divide the priority level, using physical boundaries as constraints, introducing a priority factor to set a heuristic function, using the A* algorithm to plan the path, and simulating and verifying the path in the initial model to ensure a safe distance between the robot and the flow channel, full coverage of high-priority blocks, and controllable path growth. If the standards are not met, adjustments are made to effectively improve the cleaning targeting and avoid the continuous impact of fouling on heat exchange;

[0086] The decoupled control unit is used to acquire the local coordinates of the virtual work block, control the robot to move to the guide positioning point to build initial reference parameters, start the 3D structured light camera to perform a coarse scan according to the initial reference parameters, register the coarse scan point cloud with the fin geometric model to calculate geometric deviation, adjust parameters to supplement fine scan or optimize the coarse scan point cloud to generate a local point cloud to extract real-time geometric features, simultaneously compare texture images to correct scale data, generate a local sub-model of the virtual work block, and set a dynamic cavitation threshold for real-time cavitation monitoring. The hierarchical control architecture handles cavitation warnings and periodically evaluates the control effect, solving the parameter coupling problem and eliminating the control lag caused by environmental changes, ensuring stable cleaning effect.

[0087] Specifically, the steps for planning a route include:

[0088] Based on the fin channel matrix model, virtual working blocks are divided according to the principles of natural channel separation and continuous scaling. The fin matrix is ​​divided into multiple fixed-size basic working blocks along the fin channel direction. If the scaling level within a basic working block is not unique (e.g., light and heavy scaling), it is further subdivided into smaller working blocks to ensure that each virtual working block has a uniform scaling level and channel structure, ultimately generating a virtual working set. ;in, For the first A virtual job block, , This represents the number of virtual work blocks in the finned flow channel matrix model.

[0089] For each virtual job block By using the scale thickness distribution map, the scale thickness values ​​of all microgrids covered by the current virtual work block are retrieved to calculate the average thickness. Simultaneously, it retrieves the scaling status labels of all microgrids within the current virtual work block from the scaling binary segmentation map, counts the number of scaling grids, and calculates the scaling area percentage by combining the area of ​​individual grids and the total scaling area. This allows us to obtain the scaling attributes of the virtual work block, reflecting the severity of scaling. Combined with the finned flow channel matrix model, the spacing values ​​of all independent flow channels within the current virtual work block are retrieved to calculate the average flow channel spacing. Simultaneously, the total length of the flow channel and the number of flow channel corners are obtained. Based on the logic that the more corners there are, the more complex the flow channel, a value of 1 is used as the base value for a no-corner DC channel. This value is added to the ratio of the number of flow channel corners to the total length of the flow channel to calculate the flow channel complexity coefficient. This allows us to obtain the geometric properties of the virtual work block, reflecting the overall structural characteristics of the finned flow channels within the virtual work block, and ultimately generating the virtual work block. attribute tag set The expression is as follows:

[0090]

[0091] In the formula, For virtual job blocks The number of flow channel corners inside, For virtual job blocks The total length of the internal flow channels is the virtual work block. The sum of the lengths of all internal flow channels;

[0092] For average thickness, a secondary thickness threshold is set. , ,and Weighting of scale thickness ,like ,make ,like ,make ,like ,make Similarly, by setting secondary area thresholds and secondary complexity thresholds, the scaling area weights are divided. Flow channel complexity weight Based on energy-saving targets, corresponding weighting coefficients are set to calculate virtual work blocks. The cleaning index is used to quantify the urgency of each virtual task block, and the virtual task sets are sorted from largest to smallest according to the cleaning index. According to a preset ratio, the virtual task blocks are divided into high priority, medium priority, and low priority.

[0093] Using the physical boundary of the fin matrix as a constraint, a priority weight factor is introduced. The Euclidean distance between the nodes and the endpoint is considered, and a heuristic function is set to prioritize the traversal of high-priority virtual work blocks. Virtual work blocks of the same priority are connected in series according to the shortest path principle to ensure that the cleaning interval of high-priority areas does not exceed the upper limit of cleaning time, thus avoiding the continuous impact of scaling on heat exchange. The order of low-priority areas can be flexibly adjusted. The A* algorithm is used to plan the path and generate the scaling planning path.

[0094] Import fouling planning paths into the finned flow channel matrix model, simulate the movement process of a mobile robot, and verify the feasibility of the paths, including: the minimum distance between the robot and the finned flow channel exceeds the minimum safe distance, the paths of high-priority virtual work blocks are fully covered, and the overall path length does not exceed the growth limit compared to the traditional shortest path. If the standards are not met, the path is readjusted.

[0095] Specifically, the steps for real-time cavitation monitoring include:

[0096] The fin flow channel matrix model is called to extract the global three-dimensional coordinate range of the current virtual work block to be processed. The coordinate mapping is achieved through the sensor extrinsic parameter matrix of the initial evaluation, and the coordinates are converted into local coordinates in the mobile robot coordinate system. The mobile robot is controlled to move along the scaling planning path to the guide positioning point of the virtual work block, and the fin spacing and tilt angle are obtained as the initial reference parameters for local scanning.

[0097] Because the initial assessment deviates from the actual environment, and the initial assessment is based on global static data, while local fins exhibit dynamic changes such as deformation and temporary scale buildup, direct adjustment based on the initial data would lead to nozzle-channel mismatch (pressure fluctuations) and cleaning intensity inconsistent with actual scale buildup (over-cleaning / under-cleaning). The 3D structured light camera is activated, and the scanning range is set according to the initial reference parameters to cover a certain distance outside the virtual work block boundary, avoiding missed scans. A rapid coarse scan point cloud is acquired, and the coarse scan point cloud is precisely registered with the fin geometric model using the ICP iterative nearest-point algorithm. The geometric deviation between the two is calculated, such as fin spacing deviation and tilt angle deviation. If the geometric deviation exceeds a preset geometric deviation threshold, it is considered significant, and the scanning parameters are adjusted based on the coarse scan point cloud, such as reducing the scanning step size and increasing the point cloud density. A fine scan is then performed to complete the deviation area. If the geometric deviation does not exceed the preset geometric deviation threshold, optimization is directly based on the coarse scan point cloud without fine scanning, thus generating the final local point cloud. Real-time geometric parameters, including actual fin spacing, tilt angle, and surface flatness, are extracted to generate local geometric features.

[0098] Simultaneously acquire local texture images of the virtual working block, compare them with the scale thickness distribution map, calculate the thickness deviation, and once the thickness deviation exceeds the preset thickness deviation threshold, perform local CNN fine-tuning based on multiple new texture images to correct the scale thickness value of the virtual working block and generate local scale features. Among them, the surface flatness is obtained through point cloud curvature analysis to identify the fin deformation area.

[0099] The real-time geometric parameters and the corrected scaling data are embedded into the fin flow channel matrix model to generate a local sub-model of the current virtual working block, which includes real-time geometry, dynamic scaling, and temperature distribution. Multiple feature points in the virtual working block, such as the fin tip and flow channel inflection point, are randomly selected. The actual geometric dimensions are measured with a laser rangefinder and compared with the local sub-model dimensions. If the deviation is less than the preset value, the local sub-model is deemed qualified; otherwise, the scanning and modeling are repeated.

[0100] A high-speed industrial camera and coaxial light source are installed on the side of the mobile robot nozzle to avoid reflection interference and ensure complete capture of the cavitation bubble generation area. Based on the local sub-model, the camera focal length and field of view are adjusted so that the field of view covers the area where the nozzle jet acts on the fins. The mapping relationship between camera pixels and physical size is calibrated through standard cavitation bubble samples.

[0101] Judging cavitation effect indirectly by water pressure alone cannot perceive the actual intensity and uniformity of cavitation bubbles, resulting in insufficient cavitation in heavily scaled areas and excessive cavitation in lightly scaled areas. By collecting cavitation bubble images under different operating conditions, constructing a dataset that associates cavitation bubbles with scale levels, labeling the density, size, distribution and corresponding scale levels of cavitation bubbles, training a lightweight CNN segmentation model, realizing real-time recognition and segmentation of cavitation bubble regions, and finally generating a cavitation bubble recognition model, synchronously inputting real-time images captured by a high-speed camera during cleaning, and outputting a cavitation bubble mask map, labeling the pixel coordinates and area of ​​each cavitation bubble;

[0102] Call the scale thickness distribution map to obtain the scale level of the corresponding virtual work block. Through experiments, determine the optimal cavitation range required for different scale levels, including the cavitation intensity threshold and the cavitation uniformity threshold. Then, use the optimal cavitation range of the corresponding scale level as the dynamic cavitation threshold of the current virtual work block.

[0103] Based on the cavitation bubble mask image, pixels are converted into physical areas to obtain the area of ​​each cavitation bubble. Combined with the number of cavitation bubbles, the total number of cavitation bubbles is calculated. Simultaneously, based on the ratio of the total number of cavitation bubbles to the total area of ​​the jet action region, the cavitation intensity index is calculated. Based on the ratio of the standard deviation of the cavitation bubble area to the average area of ​​the cavitation bubbles, the cavitation uniformity index is calculated by the difference between the base value and the ratio. The total area of ​​the jet action region is fixed as the area of ​​a circle with a preset diameter, reflecting the overall intensity of the cavitation bubbles, with a base value of 1.

[0104] By combining dynamic cavitation thresholds for real-time monitoring, a cavitation warning signal is triggered once the cavitation intensity index or cavitation uniformity index exceeds the dynamic cavitation threshold, and is pushed to subsequent parameter control steps.

[0105] A hierarchical control architecture is constructed, comprising a decision layer, a decoupling layer, and an execution layer, to process cavitation early warnings. The decision layer defines cavitation control targets based on local twins, clarifies control priorities, and outputs control target commands by combining data from local sub-models and dynamic cavitation thresholds. The decoupling layer performs feedforward compensation prediction and feedback rolling optimization to achieve multi-parameter decoupling. A fuzzy adaptive model predictive control algorithm is used for decoupling operations, acquiring real-time monitoring data and forward-looking data to output optimized values ​​of operational variables, thereby generating decoupling commands. The execution layer precisely executes control commands, collects physical parameters in real time, and executes decoupling commands from equipment including nozzle speed controllers, water pressure controllers, and water temperature controllers. This combines forward-looking prediction with real-time correction, solving the parameter coupling problem and eliminating control lag caused by environmental changes. Specifically, forward-looking data of the flow path the nozzle is about to enter is read from the local sub-model, such as the geometry of the flow path it will pass through in the future, including changes in flow path spacing and inclination angle.

[0106] Based on the current moving speed of the mobile robot and the flow channel accuracy of the local sub-model, a look-ahead time window is set to predict the flow channel area that the nozzle will pass through in the future time period. This ensures that the prediction range is both advanced enough to allow for adjustment time and avoids errors caused by predictions that are too far in advance. From the local sub-model, according to the mobile robot's movement trajectory, that is, along the scaling planning path in the virtual work block, the flow channel segment data from the current position to the position reached in the look-ahead time window is extracted, and the look-ahead geometric parameters are extracted, including changes in flow channel spacing, changes in flow channel inclination angle, and local resistance sources in the flow channel.

[0107] The flow channel resistance is mainly composed of friction resistance and local resistance. Friction resistance is related to the flow channel length and roughness, while local resistance is related to corners and diameter changes. The local resistance of the virtual working block is calculated using the local resistance loss formula. The friction resistance coefficient is calculated based on the fin surface roughness of the local sub-model. The friction resistance is calculated by analogy between the product of the friction resistance coefficient and the ratio of the forward flow channel length to the average flow channel spacing. The forward resistance loss is calculated by summing the friction resistance and the local resistance. The current flow channel resistance is obtained simultaneously, and the resistance change is calculated.

[0108] Based on fluid mechanics principles, the pressure of the cleaning water is positively correlated with the flow channel resistance. The greater the resistance, the higher the pressure required to drive the water flow. Based on the change in resistance, the predicted water pressure fluctuation is calculated. To offset the water pressure fluctuation caused by the predicted water pressure fluctuation, the output pressure of the high-pressure water pump needs to be adjusted in advance. The water pressure adjustment is equal in magnitude and opposite in direction to the predicted water pressure fluctuation. Based on the water pump characteristic curve, the water pressure adjustment is converted into the frequency adjustment of the water pump inverter. The decoupling layer sends the frequency adjustment to the lower actuator in the form of high-speed pulse commands to ensure that the pre-adjustment is completed before the nozzle enters the look-ahead section. At the same time, the pre-adjustment command is synchronously fed back to the local sub-model to update the predicted water pressure value and provide a benchmark for subsequent feedback control.

[0109] Set a data acquisition cycle and collect feedback data in real time, including cavitation parameter data and physical parameter data. The cavitation parameter data includes the cavitation intensity index and the cavitation uniformity index. The physical acquisition data includes the water pressure, water temperature and nozzle velocity collected by the sensor. Through the coupling matrix in the decoupling layer, the correlation between physical parameters and cavitation parameters is analyzed in real time to clarify the influence direction and sensitivity of each parameter on the cavitation state.

[0110] Based on the current feedback data and coupling matrix, the cavitation state change in the next 3 cycles is predicted. If the cavitation prediction value deviates from the target value, the optimization process is initiated to construct the objective function to generate the optimal sequence of operational variables. Only the first cycle instruction of the optimal sequence is executed to avoid future prediction errors. The optimal sequence of operational variables is then converted into a signal that the actuator can recognize.

[0111] After half of the cleaning area of ​​each virtual work block is completed, the control effect is automatically evaluated. The deviations of the cavitation intensity index and cavitation uniformity index of the virtual work block from the target values ​​are calculated. If the deviation is less than the preset deviation ratio, the control is deemed qualified. If the deviation is not less than the preset deviation ratio, the coupling matrix parameters are analyzed back and the weight coefficients of the fuzzy logic module are fine-tuned to ensure that the subsequent cleaning effect meets the standard.

[0112] The feedback monitoring module is used to collect multimodal data from virtual work blocks that have completed the cleaning operation. It combines the data collected before cleaning to build an association index, generate a set of comparison images, construct a set of quality deviations, and filter out virtual work blocks that do not meet the standards. It calculates the correction priority and sorts them by level to avoid low-urgency work blocks occupying resources and high-urgency work blocks increasing heat exchange loss. It uses a genetic algorithm to optimize the correction parameters, generate execution signals and secondary fine cleaning paths, control the robot to perform feedback regulation, and repeat the collection and verification after fine cleaning. If the standard is met, it is marked as completed; if the standard is not met, it is re-optimized to ensure accurate correction and help save energy at the cold end.

[0113] Specifically, the steps of feedback regulation include:

[0114] Once a single virtual work block is cleaned, with the cleaning progress reaching 100% as a benchmark, the mobile robot automatically receives the work acquisition instruction issued by the dynamic planning module, moves along the scale planning path to the preset acquisition point of the virtual work block, and overlaps with the path node during the global scan of the initial evaluation module to ensure consistent coordinate reference. This triggers synchronous acquisition by multimodal sensors, real-time acquisition of multimodal data, including 3D point cloud, texture image, infrared temperature map, ultrasonic thickness measurement data, and water / energy consumption data. The data is then preprocessed and combined with the data acquired before cleaning to build an association index. The post-cleaning point cloud, temperature map, and texture image of each virtual work block are bound to the corresponding virtual work block data of the initial evaluation module through a coordinate matching algorithm to form a comparison image set.

[0115] Based on the 3D point cloud after cleaning, a scale map is obtained. Combined with the scale distribution thickness map of the initial evaluation module, the residual thickness of a single grid is calculated using the grid-by-grid comparison method. Combined with all pixels in the virtual work block, the scale removal rate of the virtual work block is calculated by comparing the sum of the residual thickness of a single grid with the sum of the initial scale thickness of a single grid.

[0116] Based on each grid within the virtual work block, the scaling removal rate of a single grid is calculated, along with the average removal rate and the standard deviation of removal. Based on the ratio of the standard deviation of removal to the average removal rate, the difference between the baseline value and the ratio is calculated to obtain the removal uniformity index, which reflects the consistency of the removal rate in different areas within the virtual work block and avoids the problem of uneven local cleaning being ignored.

[0117] By comparing the initial fin geometry model with the 3D point cloud after cleaning, the damage depth is calculated by the absolute difference in the fin surface height before and after cleaning to identify fin damage, screen out effective damage with a damage depth greater than the damage threshold, and calculate the total damage area of ​​all effective damage. Based on the ratio of the total damage area to the total area of ​​the virtual work block, the damage area ratio is calculated to ensure that the integrity of the fin structure is not damaged by excessive cleaning.

[0118] Extract the initial average temperature of the virtual job block from the global temperature distribution map of the initial evaluation module. Simultaneously, the average cleaning temperature of the virtual work block is extracted from the corrected temperature distribution map after cleaning. The average fin temperature under scale-free conditions is obtained from the simulation of the fin geometry model in the initial evaluation module. The difference between the initial average temperature and the cleaning average temperature, and the difference between the initial average temperature and the fin average temperature are calculated. The heat exchange efficiency recovery rate is calculated by the ratio of these two differences. The scale removal rate, removal uniformity index, damage area ratio and heat exchange efficiency recovery rate are used as quality indicators to construct a quality deviation set. At the same time, by using the quality indicator threshold, non-compliance classification rules are set for each quality indicator to filter out and mark the non-compliant virtual work blocks. The control parameter logs of the dynamic programming module, such as cavitation intensity index and water pressure / velocity curve, are associated to preliminarily determine the cause of the deviation. For example, a low scale removal rate is due to insufficient cavitation intensity.

[0119] To avoid low-urgency virtual work blocks consuming resources and high-urgency work blocks continuously generating heat exchange losses, a correction priority index is calculated by weighted summation, combining the cleaning index and the reciprocal of the scaling removal rate. The correction priority indices of all virtual work blocks are collected, sorted in descending order, and divided into high-priority correction, medium-priority correction, and low-priority correction by setting a secondary correction threshold, generating a correction priority table to ensure that resources are prioritized for the work blocks with the greatest impact.

[0120] The control parameter logs and quality deviation sets of the dynamic programming module are imported into the fin flow channel matrix model. The correlation process between the control parameters and the quality deviation is reproduced by parameter replay. Based on computational fluid dynamics, the fin geometric model is imported, the cavitation bubble generation boundary conditions are set, and the micro-jet stripping process of the cavitation bubble collapse on the scale is simulated to construct a cavitation simulation model.

[0121] Based on the sequence of manipulated variables, the adjustable range of the correction parameters is set. With the optimization objectives of scale removal rate being greater than the scale removal threshold and damage area ratio being less than the damage threshold, a genetic algorithm is used to iteratively optimize the manipulated variables, generating multiple sets of manipulated variable combinations. These combinations are input into the cavitation simulation model, and the corresponding scale removal rate and damage area ratio are output. The parameter combinations that meet the objectives are retained until convergence to the optimal solution, thereby generating the correction parameter combination of the manipulated variable sequence, avoiding the over-washing or under-washing problems caused by traditional experience-based adjustments.

[0122] The optimized combination of correction parameters is converted into a signal that the robot actuator can recognize. The scaling planning path of the dynamic planning module is called to generate a secondary fine cleaning path, which is sent to the robot control system. The robot is controlled to perform secondary fine cleaning on the high-priority corrected virtual work block along the original path. At the same time, cavitation parameters are collected in real time to ensure that the scaling removal rate meets the standard.

[0123] After the second fine washing is completed, repeat the multimodal data acquisition and quality quantification to verify the optimization target. If the optimization target is met, mark it as completed; otherwise, return to this step to re-optimize the parameters.

[0124] Working principle and effects:

[0125] The mobile robot is fixed and calibrated, and integrated multimodal sensors are used to complete synchronous calibration. A safe scanning path is then planned using CAD drawings. Data is collected and preprocessed to construct a finned flow channel matrix model containing geometric, fouling, and temperature attributes. This solves the problems of misaligned and inaccurate benchmarks in traditional assessments, providing a reliable basis for subsequent operations. Virtual work blocks are divided based on the finned flow channel matrix model, and their fouling, geometric attributes, and cleaning indices are calculated. Priority paths are planned and verified to avoid resource occupation by low-urgency work blocks and reduce heat exchange losses in high-urgency areas. The robot then moves along the path, first performing a rough scan to verify several... The system identifies deviations and constructs local sub-models. It then monitors cavitation bubbles using a high-speed camera, sets dynamic thresholds based on scaling levels, and employs a hierarchical control architecture to handle cavitation warnings. This achieves multi-parameter decoupled control, overcoming the limitation of relying solely on water pressure to judge cavitation effectiveness and improving cleaning stability. After cleaning, multi-modal data is collected, quality indicators are calculated to construct a deviation set, and priorities are corrected by ranking them in conjunction with the cleaning index. The deviations are then imported into the model to reproduce them, and parameters are optimized through simulation. The system controls a robot for secondary fine cleaning, replacing traditional experience-based adjustments, avoiding over-cleaning and under-cleaning, protecting the fin structure, and simultaneously achieving energy saving and efficiency improvement, ensuring the economical and safe operation of the cold-end system.

[0126] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A robotic intelligent control system for energy saving and efficiency improvement at the cold end of an indirect cooling auxiliary unit, characterized in that: include: Initial evaluation module, dynamic programming module, feedback monitoring module; The initial evaluation module is used to drive the mobile robot to perform a global preliminary scan of the cold end fin matrix, obtain an initial planning image set, identify and segment the scaling area, and combine the constructed fin geometric model to obtain the scaling thickness distribution map and the overall scaling coverage, and generate a fin flow channel matrix model. The dynamic planning module is used to divide virtual work blocks based on the fin flow channel matrix model according to the principle of natural flow channel separation and scale continuity, generate virtual work sets, calculate the cleaning index of each virtual work block, construct scale planning path, drive the mobile robot to perform rapid coarse cleaning, perform cavitation early warning through dynamic cavitation threshold and cavitation parameters, and process cavitation early warning through hierarchical control architecture. The feedback monitoring module is used to acquire a set of comparison images, construct a set of quality deviations, screen out substandard virtual work blocks, calculate correction priorities and sort them by level, optimize correction parameters using a genetic algorithm, and perform feedback regulation. The feedback control steps include: After the virtual job block is cleaned, multimodal data is collected in real time, and a comparison image set is constructed. Obtain the scaling map, calculate the scaling removal rate of the virtual work block, and combine it with the scaling removal rate of a single grid to obtain the removal uniformity index; Effective damage is screened based on damage depth, the proportion of damaged area is calculated, the heat exchange efficiency recovery rate is calculated based on temperature distribution, a quality deviation set is constructed, and virtual work blocks that do not meet the standards are screened out. Calculate the correction priority index and combine it with the secondary correction threshold to generate a correction priority table; Set the adjustable range and optimization target of the correction parameters, and use a genetic algorithm to iteratively optimize the operation variables to generate a combination of correction parameters for the operation variable sequence; A secondary washing path is generated, and multimodal data collection and quality quantification are repeated to verify the optimization objectives.

2. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 1, characterized in that: The dynamic programming module includes a cleaning planning unit and a decoupling control unit; The cleaning planning unit is used to divide virtual work blocks, construct an attribute label set for each virtual work block and assign weights to each attribute, calculate the cleaning index by combining the corresponding weight coefficients, divide priority levels, introduce a priority factor to set a heuristic function, and plan the path using the A* algorithm. The decoupled control unit is used to construct initial reference parameters, perform coarse scanning on the virtual work block, calculate geometric deviations in conjunction with the fin geometric model, generate local point clouds, construct local sub-models for the virtual work block, set dynamic cavitation thresholds for real-time cavitation monitoring, construct a hierarchical control architecture, and periodically evaluate the control effect.

3. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 2, characterized in that, The steps of a global preliminary scan include: The mobile robot is calibrated to a standard standard, and a safe scanning boundary is set, a scannable grid is divided, and a safe path is generated. Executing the security path triggers the multimodal sensor to collect data, acquiring the three-dimensional point cloud, texture image, and infrared thermal image of the fin matrix surface to generate an initial planning image set; The initial planning image set is preprocessed to generate a global planning dataset; Construct a triangular mesh model of the fin matrix, extract geometric parameters, simplify the mesh, and generate the fin geometric model.

4. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 3, characterized in that, The steps of a global preliminary scan also include: A scale recognition model is constructed, and a scale binary segmentation map is generated by combining the texture image; The surface of the fin geometry model is divided into multiple grid blocks. The scale thickness of each grid block is calculated. The scale level is classified by a secondary thickness threshold, and a scale thickness distribution map is generated. The total surface area of ​​the fin flow channel is obtained from the fin geometry model, and the total fouling area of ​​all fouling grid blocks is counted to calculate the overall fouling coverage rate. Based on the scaling level, stratified sampling was performed to obtain multiple verification points, and the simulated values ​​of each verification point were recorded. The mobile robot is used to obtain actual values ​​and calculate verification indicators. If the verification indicator of a certain verification point exceeds the verification threshold, the process returns to the corresponding step for optimization until all indicators meet the standards. By integrating the fin geometry model, scale thickness distribution map, scale coverage, and global temperature distribution map, a fin flow channel matrix model is generated.

5. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 4, characterized in that, The steps for planning a route include: For each virtual job block The average thickness was calculated using the scale thickness distribution map. The scaling binary segmentation map is called to calculate the scaling area ratio. ; Based on the aforementioned finned flow channel matrix model, the average flow channel spacing is calculated. Simultaneously, the total length of the flow channel and the number of flow channel corners are obtained, and combined with the cornerless DC channel, the flow channel complexity coefficient is calculated. .

6. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 5, characterized in that, The steps for planning a route also include: Construct virtual job blocks The attribute tag set, combined with the secondary threshold, is used to divide the scale thickness weight. Scale area weight Flow channel complexity weight ; Set weighting coefficients, calculate cleaning index, sort the virtual job set, and divide it into priority levels, including high priority, medium priority, and low priority. Introduce a priority weighting factor, set a heuristic function, and use the A* algorithm to plan the path and generate a scaling planning path. The scaling planning path is imported into the finned flow channel matrix model to verify the feasibility of the path.

7. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 6, characterized in that, The steps for real-time cavitation monitoring include: Obtain the local coordinates of the current virtual job block to be processed, move to the guide positioning point, and obtain the initial reference parameters; Set the scanning range, acquire a rapid coarse scan point cloud, and calculate the geometric deviation from the fin geometric model; Once the geometric deviation exceeds the geometric deviation threshold, the scanning parameters are adjusted based on the coarse scan point cloud to perform a complete fine scan. Acquire local point clouds, extract real-time geometric parameters, and generate local geometric features; Calculate the thickness deviation, and if the thickness deviation exceeds a preset thickness deviation threshold, perform local fine-tuning to generate local scaling characteristics; By combining the local geometric features and the local scaling features, a local sub-model of the current virtual work block is generated.

8. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 7, characterized in that, The steps for real-time cavitation monitoring also include: Cavitation bubble images under different working conditions are collected, an associated dataset is constructed, and based on the constructed cavitation bubble recognition model, real-time images during cleaning are obtained to generate cavitation bubble mask images. Based on the scaling level of the virtual work block, the corresponding optimal cavitation interval is called as the dynamic cavitation threshold. Calculate the total amount of cavitation bubbles, and combine this with the total area of ​​the jet's effective region to calculate the cavitation intensity index; The cavitation uniformity index is calculated based on the ratio of the standard deviation of the cavitation bubble area to the average area of ​​the cavitation bubble. By combining dynamic cavitation thresholds for real-time monitoring, a cavitation warning signal is triggered once the cavitation intensity index or cavitation uniformity index exceeds the dynamic cavitation threshold.

9. The robotic intelligent control system for energy saving and efficiency improvement at the cold end of the intercooled auxiliary unit according to claim 8, characterized in that, The steps for real-time cavitation monitoring also include: Set a look-ahead time window and extract look-ahead geometric parameters by combining local sub-models; The local resistance of the virtual working block is calculated using the local resistance loss formula. Combined with the length of the look-ahead channel and the average channel spacing, the friction loss is calculated. Calculate the forward drag loss, simultaneously obtain the current flow channel resistance, calculate the drag change, and thus generate the frequency adjustment amount; Set a data acquisition cycle, collect cavitation parameter data and physical parameter data in real time, and analyze the correlation between physical parameters and cavitation parameters in real time; Predict changes in cavitation state; if the predicted cavitation value deviates from the target value, initiate the optimization process. The system automatically evaluates the effectiveness of each virtual work block after half of its cleaning area has been completed.

Images