A robot path planning method for incinerator decoking based on a dual-mode adaptive optimization algorithm

Abstract

This invention discloses a path planning method for a coke removal robot in an incinerator based on a dual-mode adaptive optimization algorithm, belonging to the field of boiler maintenance and industrial robot control technology. The method includes: acquiring the characteristic information of each coke block on the inner wall of the furnace and initializing the coke removal task; dividing the working area into multiple sub-regions based on location information, and determining the boundaries with the goal of balancing the workload of each sub-region; constructing a multi-objective evaluation function including movement cost components, coke removal time cost components, and priority cost components; applying the dual-mode adaptive optimization algorithm, using the multi-objective evaluation function as the fitness function to solve for the optimal coke removal path; and controlling the robot to execute the optimal coke removal path. This invention combines the dual-mode adaptive optimization algorithm with the multi-objective path planning cost function, which can balance movement distance, coke removal time, and coke block priority, reducing the robot's ineffective movement and total operation time, ensuring that key coke blocks are cleaned first, thereby achieving precise control of coke removal quality.

Description

Technical Field

[0001] This invention relates to the field of boiler maintenance and industrial robot control technology, and more specifically to a path planning method for a coke removal robot in an incinerator based on a dual-mode adaptive optimization algorithm. Background Technology

[0002] With the widespread adoption of waste-to-energy incineration technology, coking and ash accumulation on boiler heating surfaces have become key challenges restricting the operational efficiency and safety of waste-to-energy plants. During waste incineration, ash particles in the flue gas melt under high temperatures and deposit and sinter on boiler heating surfaces such as water-cooled walls and superheaters, forming hard coke lumps. The presence of coke lumps not only reduces the heat transfer efficiency of the heating surfaces, leading to a decrease in boiler thermal efficiency and an increase in flue gas temperature, but may also cause serious safety accidents such as localized overheating of the heating surfaces, pipe deformation, and even pipe rupture. Furthermore, it increases flue gas flow resistance, affecting the stable operation of the incinerator.

[0003] Traditional manual coke removal methods have inherent drawbacks such as high safety risks, high labor intensity, and long operation cycles. Manual coke removal requires workers to enter the furnace after the boiler has been shut down and cooled, resulting in a harsh working environment and difficulty in effectively cleaning high areas of the furnace wall, severely impacting the continuous and stable operation and economic efficiency of the incineration plant. Existing technologies include mechanized coke removal equipment to replace manual labor, such as telescopic arm coke removal robots and dry ice coke removal robots. While these devices improve the safety and automation of coke removal operations to some extent, their path planning intelligence is low, often relying on predefined fixed trajectories or manual remote control by operators.

[0004] This "one-size-fits-all" operating mode fails to perceive the actual distribution, thickness, and hardness of coke blocks inside the furnace, resulting in a large number of ineffective idle strokes during the cleaning process, leading to severe resource waste and low efficiency. Furthermore, fixed-track operations cannot prioritize the cleaning of thick coke blocks in critical areas such as weld joints and areas prone to overheating, making it difficult to completely eliminate safety hazards. In addition, existing coke removal equipment lacks adaptive adjustment capabilities during operation, failing to dynamically optimize the coke removal path based on real-time operating conditions, resulting in poor adaptability to the variable internal environment of the incinerator.

[0005] Therefore, existing technologies lack an intelligent planning method capable of adaptively generating optimal paths based on real-time operating conditions. Developing an intelligent path planning method that can accurately sense coking conditions and achieve efficient, energy-saving, and comprehensive cleaning is an urgent industrial need for improving the automation level and overall efficiency of incinerator coking removal operations. Summary of the Invention

[0006] To address the problem that existing incinerator decoking robots mostly operate on fixed trajectories, making it difficult to plan targeted paths based on the actual distribution and severity of coke blocks, resulting in numerous ineffective idle strokes, insufficient decoking in key areas, and low overall operating efficiency, this invention provides a path planning method for incinerator decoking robots based on a dual-mode adaptive optimization algorithm. This method features a reasonable algorithm design, high decoking efficiency, and the ability to intelligently plan the optimal decoking path based on the spatial distribution, type characteristics, and severity of coke blocks.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a path planning method for a coke removal robot in an incinerator based on a dual-mode adaptive optimization algorithm, comprising: Step S1: Obtain the characteristic information of each coke block on the inner wall of the furnace. The characteristic information includes at least the coke block location information and the evaluation index reflecting the priority of coke block cleaning. The coke removal task is initialized based on the characteristic information. Step S2: Based on the coke block location information, the working area to be decoked in the furnace is divided into multiple sub-areas, and the boundary of each sub-area is determined with the balance of workload between each sub-area as the dividing objective. Step S3: Construct a multi-objective evaluation function for evaluating the quality of the desiccation path. The multi-objective evaluation function includes at least a movement cost component that represents the cost of robot movement, a time cost component that represents the time cost of desiccation operation, and a priority cost component that represents the priority cost of the desiccation block. Step S4: Apply the dual-mode adaptive optimization algorithm to solve for the optimal defocusing path using the multi-objective evaluation function as the fitness function. The dual-mode adaptive optimization algorithm performs a random search in the solution space through a random exploration mode and a fine search near the current optimal solution through a local exploration mode. The execution ratio of the random exploration mode and the local exploration mode is controlled by adaptive parameters that change dynamically with the iteration process. Step S5: Control the robot to perform the decoking operation according to the optimal decoking path.

[0009] In one embodiment, it also includes: Step S6: During the decoking operation, adjust the path parameters based on the real-time monitored environmental feedback information, and record the execution data after the decoking operation is completed to optimize the system parameters.

[0010] In one embodiment, in step S1, the evaluation index is obtained in the following way: The acquired images of the furnace inner wall are preprocessed, and the coke block areas in the images are identified. Extract image features from each focal block region, the image features including color features, texture features and / or size features; The positional importance weight is determined based on the location of the coke block on the inner wall of the furnace. The evaluation index is calculated based on the image features and positional importance weights of each focal block region.

[0011] In one embodiment, step S2, determining the boundaries of each sub-region, includes: The individual workload of each coke block is calculated based on its area, type coefficient, and evaluation indicators. The boundary positions of each sub-region are determined by minimizing the standard deviation of the workload of all sub-regions as the objective function. The workload of a single sub-region is the sum of the individual workloads of all the coke blocks contained in that sub-region.

[0012] In one embodiment, step S2 employs a four-quadrant region segmentation strategy to divide the work area into four sub-regions, with the goal of minimizing the standard deviation of workload in the four sub-regions:

[0013] in, The standard deviation of workload; Minimize the function; Calculate the square root; For the first The estimated total workload for each sub-region, where k takes values ​​in the range [1,4]; The average estimated workload for all sub-regions; , coke area, coke Type coefficient, coke Severity index The serial number of the coke block; A one-dimensional search method is used to dynamically determine the boundaries of each sub-region.

[0014] In one embodiment, the multi-objective evaluation function is: J = α·C m +β·C t +γ·C p Where J is the total cost; α, β, and γ are weighting coefficients; C m For the moving cost component; C t C represents the time cost component for decoking; p The priority cost component is a negative correlation function, which means that the higher the evaluation index of the coke block, the lower the corresponding priority cost.

[0015] In one embodiment, the formula for calculating each cost component is as follows: Moving cost components:

[0016] in, is the Euclidean distance between adjacent focal blocks i and j in the path sequence; η is the turning angle between adjacent path segments at focal block i and focal block j in the path sequence; η is the turning penalty coefficient, used to convert the turning angle into an equivalent additional movement distance with the same dimension as the Euclidean distance; De-coking time cost components:

[0017] Among them, A i coke i The area, K i coke i Type coefficient, K m The coefficient of performance of the heated surface material. To improve the efficiency of decoking operations; Priority cost components:

[0018] Among them, S i coke i Evaluation indicators.

[0019] In one embodiment, the dual-mode adaptive optimization algorithm includes an initialization phase and an iterative optimization phase; Initialization phase: A set of initial solutions is randomly generated as the initial population, and algorithm parameters are set, including the total number of iterations T and the maximum decay parameter e. max and minimum attenuation parameter e min ; Iterative optimization phase: In each iteration, the adaptive parameter e(t) is determined based on the current iteration number t, and the search mode to be used in this iteration is determined from the random exploration mode and the local exploration mode based on the value of e(t), generating a new candidate solution; the fitness value of the new candidate solution is compared with the fitness value of the current solution. If the fitness value of the new candidate solution is better than the fitness value of the current solution, the current solution is replaced by the new candidate solution; after the population update is completed, it is determined whether the termination condition is met. If so, the current optimal solution is output as the optimal defocusing path; where the fitness value is the multi-objective evaluation function J, the smaller the value, the better.

[0020] In one embodiment, the formula for generating candidate solutions in the random exploration mode is:

[0021] in, To randomly explore newly generated candidate solutions; x is the current solution; a and b are both exploration intensity coefficients; and All numbers are random numbers in the range [0,1]; L(β) is the random step size following a Lévy distribution, where β is the Lévy distribution exponent, ranging from (1,2]; δ is the step size scaling factor that decreases with the iteration process; x max and x min These are the upper and lower bounds of the solution space, respectively; The formula for generating candidate solutions in the local exploration mode is:

[0022] in, To explore newly generated candidate solutions locally; is the historical best solution in the current population; c and d are development accuracy coefficients; rand is a random number in the range [0,1].

[0023] In one embodiment, the adaptive parameter e(t) is calculated as follows:

[0024] Where λ is the decay constant and T is the total number of iterations; the value of e(t) decays as the number of iterations increases, and the algorithm dynamically adjusts the execution ratio of random exploration mode and local exploration mode according to the value of e(t): when the value of e(t) is greater than the preset threshold, the current iteration executes random exploration mode; when the value of e(t) is not greater than the preset threshold, the current iteration executes local exploration mode.

[0025] In one embodiment, the termination condition is: reaching the total number of iterations T, or the improvement of the optimal solution within N consecutive iterations being less than a preset percentage ε.

[0026] In one embodiment, in step S4, the solution adopts a random key-based encoding method to encode the access order of the focal blocks into a continuous real number vector, wherein the order of the values ​​of each dimension of the continuous real number vector corresponds to the access order of each focal block.

[0027] In one embodiment, the individual workload of the coke block is calculated as follows: the individual workload is calculated based on the area of ​​the coke block, the coke block type coefficient, and the evaluation index, wherein the coke block type coefficient is determined based on the hardness or density of the coke block, and the evaluation index is obtained based on the visual characteristics and positional importance of the coke block.

[0028] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages: 1. This invention obtains the characteristic information of coke blocks and initializes the coke removal task, so that path planning no longer depends on a preset fixed trajectory, but can combine the actual distribution of coke blocks in the furnace to model the task, thereby improving the pertinence and flexibility of path planning.

[0029] 2. This invention determines the boundaries of each sub-region by using a region segmentation strategy based on the workload of the coke block, which helps to improve the problem of uneven workload between different regions, avoids excessive concentration of tasks in local areas, and improves the rationality of overall job scheduling.

[0030] 3. This invention, by constructing a multi-objective evaluation function that includes robot movement cost, desiccation operation cost, and coke block priority cost, can reduce ineffective idle travel and improve the priority processing capability of key coke blocks and critical areas; wherein the priority cost component adopts (1-S i ) / S i The functional form of the function significantly enhances the driving force for prioritizing high-severity coke blocks.

[0031] 4. This invention employs a dual-mode adaptive optimization algorithm that includes both random and local exploration modes, which balances global search capability and local convergence capability, thus improving the quality and efficiency of path solving. By using a random key-based encoding method, the discrete path permutation problem is transformed into a continuous real vector optimization problem, providing a foundation for continuous numerical computation of the algorithm. Through a greedy selection strategy and an elite retention mechanism, the convergence and stability of the algorithm are guaranteed.

[0032] 5. This invention enhances the robot's adaptability to complex working conditions inside the furnace by dynamically adjusting the path based on environmental feedback during task execution. At the same time, the system has closed-loop optimization capabilities, which can continuously learn from historical task data to optimize algorithm parameters and cost weights. In long-term operation, resource utilization efficiency and overall performance are continuously improved, effectively reducing water and electricity consumption and equipment wear. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0034] Figure 1 This is a flowchart of a path planning method for a coke removal robot in an incinerator based on a dual-mode adaptive optimization algorithm, provided in an embodiment of the present invention. Figure 2 A schematic diagram of the four-quadrant region segmentation strategy provided in an embodiment of the present invention; Figure 3 The flowchart is provided for the dual-mode adaptive optimization algorithm in the embodiments of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1: like Figure 1 As shown, this embodiment provides a path planning method for a coke removal robot in an incinerator based on a dual-mode adaptive optimization algorithm, including the following steps: Step S1: Obtain the characteristic information of each coke block on the inner wall of the furnace. The characteristic information includes at least the coke block location information and the evaluation index reflecting the priority of coke block cleaning. The coke removal task is initialized based on the characteristic information.

[0037] Task initialization: Obtain the characteristic information of each coke block on the inner wall of the furnace.

[0038] Taking a waste-to-energy plant as an example, the boiler model is SLC500-4.0 / 400, and the furnace water-cooled wall is a membrane water-cooled wall structure. The coke removal robot adopts a tracked mobile platform equipped with a dry ice spray coke removal device. During operation, it moves along the surface of the water-cooled wall and removes the coke by impacting it with high-pressure sprayed dry ice particles.

[0039] After the robot is started, its visual recognition subsystem begins to operate. It acquires images of the furnace interior wall using a high-definition visual sensor. After image preprocessing such as Gaussian filtering for noise reduction, illumination compensation, and contrast enhancement, it applies a deep learning-based object detection algorithm, such as the YOLO model, to identify coke block regions in the images. In this embodiment, the visual recognition subsystem identified a total of 16 coke blocks.

[0040] For each identified focal spot, the system extracts its structural and visual features, mainly including: (1) Physical position coordinates: The coordinates of the center position of the coke block in the furnace coordinate system, in millimeters. For example, the coordinates of coke block P1 are (1200, 3500), which means that it is located 1200 mm from the bottom and 1800 mm from the left end of the left wall of the furnace. The physical position coordinates determine the robot's movement path and turning angle.

[0041] (2) Geometric dimensions: including the pixel area of ​​the focal block in the image, which is converted into physical area by combining the actual depth information, with the unit being square millimeters. The outline shape and thickness information of the focal block are also recorded. In this embodiment, the focal block area ranges from 520 mm² to 8600 mm².

[0042] (3) Evaluation indicator S i The calculation is based on the combined weights of the color features (grayscale value), texture features (surface roughness), and positional importance of the focal block, as follows: S i =min(1,k1×(A)) i / Aref)+k2×Dcolor+k3×Ttexture+k4×Iposition) Among them, A i The area is denoted as Aref, which is the reference area (taking the maximum coke area of ​​8600mm²) and used to normalize the area factor. Dcolor is the color depth factor (range 0-0.3), with a larger value for darker coke colors (lower grayscale values). Ttexture is the texture roughness factor (range 0-0.2), with a larger value for rougher surfaces. Iposition is the position importance weight (range 0-1.5), with 1.5 for coke located in the boiler pipe weld interface area, 1.2 for coke located in the critical heat exchange surface, 1.0 for coke located in the overheating zone, 0.8 for coke located in the general water-cooled wall area, and 0.5 for coke located in the furnace corner and other secondary areas. k1, k2, k3, and k4 are normalization weight coefficients. In this embodiment, k1 = 0.5, k2 = 0.15, k3 = 0.1, and k4 = 0.25. These coefficients can be adjusted through experimental calibration or learning from historical data.

[0043] Taking a specific coke block as an example: coke block P4 has an area of ​​A = 8600 mm², a color depth factor Dcolor = 0.2, a texture roughness factor Ttexture = 0.1, and is located in the weld interface area (Iposition = 1.5), then: S4=min(1,0.5×(8600 / 8600)+0.15×0.2+0.1×0.1+0.25×1.5) = min(1,0.5+0.03+0.01+0.375) = min(1,0.915) =0.915 This coke block is large, dark in color, and strategically located, with an evaluation index as high as 0.915, and should be prioritized for removal. In this embodiment, the evaluation index ranges from 0.25 to 0.915 for the 16 coke blocks. The structured data of all coke blocks is organized into JSON format and sent to the path planning system to initialize this coke removal task.

[0044] Step S2: Based on the coke block location information, the working area to be decoked in the furnace is divided into multiple sub-regions, and the boundary of each sub-region is determined with the balance of workload among the sub-regions as the dividing objective. The multiple sub-regions include, but are not limited to, four-quadrant, bisection, grid division, and division based on coke block clustering.

[0045] Region segmentation: For example, a four-quadrant region segmentation strategy can be used to divide the work area into four sub-regions.

[0046] The region segmentation processing subsystem executes a four-quadrant segmentation strategy, the principle of which is as follows: Figure 2 As shown, Figure 2 In this diagram, P represents the coke block, and Q1 to Q4 represent four sub-regions. Considering that the inner wall of the incinerator furnace is a curved structure rather than an ideal plane, this embodiment uses a "projection + piecewise fitting" method to determine the segmentation boundary: first, the curved inner wall is projected onto a plane coordinate system along the unfolding direction, and the optimal segmentation line position is solved in the plane coordinate system. Then, the segmentation line is mapped back to the actual curved inner wall, and the curve segments between adjacent segmentation points are piecewise linearly fitted along the shape of the furnace, thereby forming a continuous segmentation boundary at the corresponding position on the actual curved inner wall.

[0047] First, calculate the individual workload W for each coke block. i =A i ×K i ×S i K i This is the coke type coefficient. The coke type coefficient is determined based on the density and hardness of the coke: for hard sintered coke (dark red in color and hard in texture), K is taken as the coefficient. i =1.5; Loose ash (grayish-white, loose texture) take K i =1.0; K is taken for medium-dense coke. i =1.2. The density of the coke blocks can be automatically classified by the visual recognition subsystem based on the surface color, reflective properties, and texture of the coke blocks.

[0048] Taking coke block P4 as an example, its area A = 8600 mm², and its type is hard sintered coke block (K). i =1.5), evaluation index S i =0.915, then its individual workload is W4 = 8600 × 1.5 × 0.915 = 11803.5.

[0049] Subsequently, with the objective of minimizing the standard deviation σ of the workload in the four sub-regions Q1, Q2, Q3, and Q4, an iterative optimization method is used to dynamically determine the segmentation boundaries. Specifically, the system solves the following optimization problem:

[0050] Where σ is the standard deviation of workload; Wk Let μ be the estimated total workload for the k-th sub-region, which is the sum of the individual workloads of all focal blocks within that sub-region; μ is the average workload of the four sub-regions; k = 1, 2, 3, 4. The position of the dividing line is used as an optimization variable, and can be achieved using one-dimensional search, the golden section, particle swarm optimization, etc. For example, a one-dimensional search method can be used to find the coordinates of the dividing line that minimizes σ on the unfolded plane. The constraints of the dividing boundary include: (1) Sub-areas must not cross areas that are inaccessible inside the furnace, such as the location of explosion-proof doors; (2) The shape of the sub-region must meet the reachability constraints of the robot arm, such as the working radius not exceeding 3m; (3) The dividing line should avoid falling in areas with dense coke blocks as much as possible in order to reduce the generation of cross-regional paths.

[0051] In this embodiment, the workload data for the four sub-regions obtained after optimization are, for example: W1 = 24680, W2 = 25890, W3 = 25120, W4 = 24350. The standard deviation of the workload is calculated as follows: μ = (24680 + 25890 + 25120 + 24350) / 4 = 25010 σ = sqrt[((24680-25010)² + (25890-25010)² + (25120-25010)² + (24350-25010)²) / 4] = sqrt[(108900 + 774400 + 12100 + 435600) / 4] = sqrt(83187.5) ≈288.5 The standard deviation of the workload is only about 1.15% of the average of 25010, indicating that the workload is highly balanced across sub-regions. By employing a hierarchical processing strategy, the problem in the entire working area is decomposed into four independent sub-regions, each with its own path planning function. This effectively reduces the problem size and improves the system's real-time performance and solution quality.

[0052] Step S3: Construct a multi-objective evaluation function for evaluating the quality of the desiccation path. The multi-objective evaluation function includes at least a movement cost component representing the cost of robot movement, a time cost component representing the time cost of desiccation operation, and a priority cost component representing the priority cost of the desiccation block.

[0053] Constructing a multi-objective evaluation function: The path planning and optimization subsystem models the defocusing task as a multi-objective optimization problem, and the constructed evaluation function is as follows: J = α·C m + β·C t + γ·C p Where J is the total cost, and the algorithm optimizes to minimize J; α, β, and γ are weighting coefficients (in this embodiment, α = 0.4, β = 0.4, and γ = 0.2); C m For the moving cost component, C t As a component of decoking time cost, C p These are the priority cost components. The specific calculations for each cost component are as follows: Moving cost component C m Calculate the sum of Euclidean distances between all consecutive focal points in the path sequence, and add a penalty term for turns exceeding 45°.

[0054]

[0055] in, is the Euclidean distance between adjacent focal blocks i and j in the path sequence; η is the turning angle between adjacent path segments at focal block i and focal block j in the path sequence; η is the turning penalty coefficient, used to convert the turning angle into an equivalent additional movement distance with the same dimension as the Euclidean distance; (in this example, η = 0.05). and Since they have the same dimensions, they can be directly added together. Combining them is to allow the algorithm to find an optimal or near-optimal path that is "short in distance and has few / smooth turns," avoiding the generation of invalid paths with frequent twists and turns.

[0056] Taking two adjacent coke blocks P1(1200,3500) and P6(1500,3000) as an example, d(1,6)=sqrt((1500-1200)²+(3000-3500)²)≈583.1mm.

[0057] Decoking time cost component C t The decoking time for each coke block is calculated and summed based on the coke block area, type coefficient, heating surface material coefficient, and decoking efficiency.

[0058]

[0059] Among them, A i K represents the area of ​​the coke block. i K is the coke block type coefficient. m K is the material coefficient of the heating surface, reflecting the difference in decoking difficulty among different heating surface materials. For carbon steel water-cooled walls, K is taken as... m =1.0, K is taken for stainless steel superheater tube bundle m =1.3, K is used for chrome plating. m =0.8, material information can be read from the boiler design drawing database or pre-entered by the operator; To improve the efficiency of the decoking process (based on the performance parameters of the decoking device, in this embodiment...) =500mm² / s). Taking coke block P4 as an example, the decoking time is (8600×1.5×1.0) / 500 = 25.8 seconds.

[0060] Priority cost component C p :

[0061] This component is related to the coke evaluation index S. i Inversely proportional. For S i The cost of a coke block with a density of 0.9 is approximately 0.11; for S i A focal block with a severity of 0.3 costs approximately 2.33, a difference of about 21 times, which makes the algorithm try its best to avoid placing low-severity focal blocks at the beginning of the path.

[0062] The initial values ​​of the weighting coefficients α, β, and γ are set empirically and continuously optimized by the parameter adaptive adjustment subsystem based on historical task data after the system is running.

[0063] Step S4: Apply the dual-mode adaptive optimization algorithm to solve for the optimal defocusing path using the multi-objective evaluation function as the fitness function. The dual-mode adaptive optimization algorithm performs a random search in the solution space through a random exploration mode and a fine search near the current optimal solution through a local exploration mode. The execution ratio of the random exploration mode and the local exploration mode is controlled by adaptive parameters that change dynamically with the iteration process. Path finding: A dual-mode adaptive optimization algorithm is applied, the process is as follows: Figure 3 As shown.

[0064] 1. Encoding and initialization.

[0065] Considering that focal block de-spotting path planning is essentially a combinatorial optimization problem (i.e., a permutation problem of the Traveling Salesman Problem (TSP)), the solution is represented as a permutation of the focal block access order. To adapt to the continuous numerical computation characteristics of the random search formula and the local exploration formula in the dual-mode adaptive optimization algorithm, this embodiment adopts a random key-based encoding method to encode the focal block access order into a continuous real number vector. The order of the values ​​of each dimension of the vector corresponds to the access order of each focal block.

[0066] For example, an 8-dimensional random key encoding vector x = [3.2, 0.5, 1.8, 2.7, 0.9, 4.1, 2.3, 1.5] corresponds to 8 focal blocks P1 to P8. After arranging the values ​​of each dimension in ascending order, the second dimension value of 0.5 is the smallest (focal block P2 is visited first), the fifth dimension value of 0.9 is the next smallest (focal block P5 is visited second), and so on. This uniquely maps to a valid defocal path sequence. Through this encoding method, any continuous real number vector can be uniquely mapped to a valid path sequence, and the order in which focal blocks are visited depends only on the relative magnitude of each dimension value, and is independent of the absolute magnitude. This makes it feasible for the algorithm to perform scalar addition and subtraction operations in the real number field.

[0067] A random initial defocusing path sequence is generated as the initial population; in this embodiment, the population size N = 50. The total number of iterations T = 1000, the decay constant λ = 3, and the maximum value of the adaptive parameter e are set. max =0.7, minimum value e min =0.3.

[0068] 2. Iterative optimization.

[0069] In each iteration, the adaptive parameter e(t) is first calculated for the current iteration number t:

[0070] Where e max =0.7, e min =0.3, λ=3, T=1000. Examples of e(t) values ​​for each generation are as follows: e(0)=0.7; e(100)≈0.596; e(500)≈0.389; e(900)≈0.327; e(1000)≈0.320 It can be seen that in the early stage of the iteration, e(t) is close to 0.7, and the algorithm performs random exploration with a high probability; in the later stage of the iteration, e(t) drops to about 0.32, and the algorithm switches to performing local exploration for refinement with a high probability. The continuous decay of e(t) realizes the adaptive transition from exploration to development.

[0071] Random exploration mode, also known as the random exploration phase (Levi flight): For each individual x in the population, the random exploration mode is executed with probability e(t), introducing random perturbations based on the current solution to generate a new candidate solution x1:

[0072] in, The newly generated candidate solution; x is the current solution; a and b are the exploration intensity coefficients (in this example, a = 0.9, b = 2); and is a random number in the range [0,1]; L(β) is the random step size following a Lévy distribution; β is the Lévy distribution exponent (range (1,2], in this example β=1.5); δ is the Lévy step size scaling factor that decreases with the iteration process; x max and x min These are the upper and lower bounds of the solution space, respectively.

[0073] L(β) is generated using the Mantegna algorithm:

[0074] Where u and v are random numbers following a normal distribution, i.e., u ~ N(0, σ). u ²), v~N(0,σ v ²), σ u and σ v Calculated using the following formula:

[0075] σ v Let σ be 1, i.e., v ~ N(0,1). v Setting it to 1 is the standard setting when the Mantegna algorithm generates the Lévy random step size, which is used to ensure the scale stability of the random variable in the denominator of the Lévy step size generation formula.

[0076] Where Γ is the gamma function. Substituting β = 1.5, we can calculate σ. u ≈0.696. The Lévy step size scaling factor δ adopts a linear decreasing strategy, with an initial value δ0 = 0.1.

[0077] At t=100, δ=0.09, and at t=500, δ=0.05. This strategy allows the algorithm to exert a strong stochastic exploration effect during the initial Lévy flight iterations, while its influence gradually weakens in the later stages of iterations, which is beneficial for stable convergence. To avoid the denominator (1-b·rand1) approaching zero, the algorithm resamples rand1 during initialization to ensure computational stability.

[0078] Local exploration phase: This phase directs newly generated solutions toward the historical best solution x in the current population. m As the directions converge, a refined search is performed to accelerate convergence. The local exploration formula is:

[0079] in, x represents the newly generated candidate solution; x represents the current solution. Let c be the historical best solution in the current population; c and d are development accuracy coefficients (c = 0.9 and d = 0.1 in this example), satisfying c² + d² < 1, thus statistically ensuring that the algorithm has contractility and can converge toward the optimal solution. This formula enables the new solution to perform interpolation search between the current solution and the historical best solution, quickly converging to the vicinity of the current best solution.

[0080] Selection Operator and Elite Retention: After generating a new candidate solution in each iteration, a greedy selection strategy is adopted to compare the fitness value (i.e., total cost J) of the new candidate solution with the fitness value of the current solution. If the fitness value of the new candidate solution is better (J is smaller), it is used to replace the current solution. At the same time, an elite retention mechanism is introduced to directly retain the global optimal solution in the current population to the next generation, preventing the optimal solution from degenerating and being lost, thus forming a complete closed loop in the algorithm.

[0081] 3. Termination and output.

[0082] The algorithm terminates when the total number of iterations T is reached, or when the improvement of the optimal solution within 50 consecutive iterations is less than 1%, and outputs the current optimal defocusing path sequence. In this embodiment, after being solved by the dual-mode adaptive optimization algorithm, for example, the output optimal defocusing path sequence is: P4→P7→P2→P9→P1→P6→P11→P3→P8→P14→P5→P10→P13→P12→P15→P16; The total cost of the path sequence is J = 0.4 × 186.4 + 0.4 × 125.6 + 0.2 × 45.2 ≈ 131.36, which is better than the solution obtained by using a greedy strategy (J ≈ 220.5) and a single genetic algorithm (J ≈ 165.8).

[0083] Step S5: Control the robot to perform the decoking operation according to the optimal decoking path.

[0084] Step S6: During the decoking operation, adjust the path parameters based on the real-time monitored environmental feedback information, and record the execution data after the decoking operation is completed to optimize the system parameters.

[0085] For example, the robot control and execution subsystem takes over the optimized path sequence. The path tracking module controls the robot to move precisely to each focal point; the motion control module calculates the wheel speed and controls the defogging mechanism (dry ice spraying mechanism) to work. In this embodiment, the robot's movement speed is 0.3 m / s. The status monitoring module evaluates the defogging effect in real time through periodic image analysis, and performs an effect evaluation after defogging operations on every 5 focal points are completed. If an unforeseen focal point is found or the defogging effect is poor at a certain point, local path replanning can be triggered—the new focal point is inserted into the position with the minimum cost increment, and an updated path is quickly generated.

[0086] After the task is completed, all execution data, such as actual time consumption, energy consumption, and omission rate, are recorded and sent to the parameter adaptive adjustment subsystem. In this embodiment, the actual total descorching time is 21.4 minutes, the total moving distance is 68.5m, and the descorching coverage rate is 96.8%. Based on the deviation between the actual execution data and the theoretical planned data, the system uses the gradient descent method to continuously optimize the weight coefficients α, β, γ and the algorithm parameters a, b, c, d, achieving closed-loop optimization and continuous learning of the system.

[0087] Comparison data from simulation experiments: To further verify the technical effectiveness of the method of this invention, three sets of comparative experiments were conducted in this embodiment. Each set of experiments used the same 16 focal block data for path planning: Experiment A used the method of this invention (dual-mode adaptive optimization algorithm, parameters as above); Experiment B used a single genetic algorithm (GA, population size 50, crossover probability 0.8, mutation probability 0.1, iteration count 1000); Experiment C used a fixed trajectory operation method (a serpentine traversal order from top to bottom and from left to right). The comparison results of the three sets of experiments are shown in Table 1. Table 1 Performance Comparison of Three Path Planning Methods

[0088] As shown in Table 1, after adopting the method of the present invention, the total travel distance is reduced by approximately 51.9% compared to the fixed trajectory method and by approximately 21.4% compared to the GA method; the total descorching time is reduced by approximately 44.0% compared to the fixed trajectory method and by approximately 19.2% compared to the GA method; and the proportion of invalid empty travel is reduced from 35.0% to 12.1%. Especially in terms of high-priority coke block coverage, the method of the present invention ensures that all high-priority coke blocks are prioritized for cleaning within the first 30% of the path sequence, a targeted priority scheduling that neither the fixed trajectory method nor the GA method can achieve. The above data fully verify that the method of the present invention has significant technical effects in improving descorching efficiency, reducing resource consumption, and ensuring the quality of descorching in key areas.

[0089] Example 2: Consider the detailed segmentation of the curved furnace inner wall; This embodiment, based on Embodiment 1, further refines the four-quadrant segmentation implementation method of step S2 in the case of a curved furnace inner wall, to better meet the actual engineering application scenario. Step S1 in Embodiment 1 serves as a prerequisite step in this embodiment, and its specific implementation method is the same as in Embodiment 1.

[0090] The refinement of step S2 in this embodiment includes: S2.1 Furnace Inner Wall Unfolding and Mapping. The inner wall of an incinerator furnace is typically a curved surface structure with a rectangular or circular cross-section. To simplify calculations, this embodiment uses an equidistant mapping method to unfold the actual curved inner wall: unfolding along the horizontal circumferential direction of the furnace, mapping the curved surface into a planar rectangular region. Distinguishing between square and circular cross-section furnaces—for a rectangular cross-section furnace, its inner wall consists of four planes with rounded transitions between adjacent planes (radius typically 50–100 mm). The rounded transition area is assigned to the larger of the adjacent planes. For a circular or elliptical cross-section furnace, an unfolding sampling point is taken every 1° along the circumferential direction. The abscissa on the unfolded plane is the arc length coordinate s = R × θ (θ is the central angle corresponding to the sampling point, and R is the radius of the arc). Through arc length coordinate mapping, a one-to-one correspondence is formed between the points on the curved inner wall and the points on the unfolded plane.

[0091] S2.2 Candidate Segmentation Boundary Generation. On the unfolded plane, the positions of two segmentation lines are determined (the first segmentation line along the height direction and the second segmentation line along the width direction). The two segmentation lines intersect at a point, dividing the unfolded plane into four rectangular sub-regions. In the mapping scenario of the curved furnace inner wall, the actual segmentation lines are curve segments on the unfolded plane. In this embodiment, a piecewise linear fitting method is used to approximate the curved segmentation boundary: the segmentation line direction is uniformly discretized into several (e.g., 5 to 10) control points, and the position coordinates of the control points are used as optimization variables; in each generation of optimization, the control points are connected sequentially and smoothed to form candidate segmentation boundaries; after solving, the coordinates of the segmentation boundaries are mapped back to the actual three-dimensional curved inner wall to form segmentation instructions that the robot can execute in actual space.

[0092] S2.3 Solving the objective function. After generating candidate segmentation boundaries, calculate the workload of individual coke blocks contained in each sub-region and evaluate the objective function σ. The same iterative optimization method as in Example 1 is used to optimize the coordinates of the segmentation control points. The constraints include: (1) The segmentation line must not cross obstacle areas (such as explosion-proof doors, fire viewing holes, and maintenance manholes) and their safety buffer areas (the buffer distance is not less than 100mm); (2) The area of ​​the segmented sub-region must not be less than the minimum operating coverage area threshold of the robot (such as 1m²); (3) The segmentation line should avoid dense coke block areas and avoid dividing one coke block into two sub-regions. It is stipulated that there should be no coke block center point within the buffer zone of 20mm on both sides of the segmentation line; (4) The aspect ratio of each sub-region after segmentation should not be greater than 3:1.

[0093] The optimization process converged after 300 iterations. The final determined segmentation boundary parameters are: the coordinates of the first segmentation control point are [(1200,3200),(1350,3250),(1500,3180),(1650,3220),(1800,3200)], and the coordinates of the second segmentation control point are [(1400,2500),(1380,2650),(1450,2800),(1420,2950),(1400,3100)]. After mapping the segmentation boundaries back to the inner wall of the 3D curved surface, the actual surface areas of the four sub-regions Q1 to Q4 are 12.3m², 11.8m², 11.2m², and 11.1m², respectively, with an area difference of less than 10%. The workload data for the four sub-regions are W1 = 24810, W2 = 25760, W3 = 24690, and W4 = 24740, with a workload standard deviation σ ≈ 488.5 and a relative standard deviation of approximately 2.0%, which is better than the 6.8% before optimization. The execution method of subsequent steps S3 to S5 is the same as in Example 1.

[0094] Example 3: Parameter selection and sensitivity analysis of dual-mode adaptive optimization algorithm; This embodiment further illustrates the basis for the selection of key parameters (exploration intensity coefficients a and b, development accuracy coefficients c and d, and decay constant λ) in the dual-mode adaptive optimization algorithm, and the impact of parameter changes on algorithm performance. It should be noted that the specific values ​​of a, b, c, and d must meet certain constraints to ensure the convergence of the algorithm and the feasibility of the solution. Taking the local exploration formula as an example, when c = 0.9 and d = 0.1, c² + d² = 0.82 < 1, and statistically, the expected distance of the new solution x2 is no greater than the distance from the current solution x to the optimal solution x. m The distance between c and d ensures that the algorithm is contractile and can converge toward the optimal solution; if c or d is too large (e.g., c² + d² > 1), the algorithm may diverge.

[0095] The parameter sensitivity analysis experiment was conducted using data from 16 focal blocks in Example 1, and the results are as follows: (1) When a takes 0.5 to 1.0 and b takes 1.5 to 2.5, the algorithm can converge to a better solution. However, the convergence speed is the fastest when a = 0.9 and b = 2, with an average convergence algebra of 142. (2) When c is 0.7 to 1.0 and d is 0.05 to 0.3, the algorithm converges stably. If c is too small (e.g., c = 0.5), the convergence speed will be significantly slower; if d is too large (e.g., d = 0.5), the search step size will be too large in the later stages, and the improvement of the optimal solution will decrease. (3) When λ takes values ​​from 1 to 5, the algorithm can converge within 1000 generations. The larger λ is, the faster e(t) decays, and the earlier the algorithm switches from exploration to local refinement. When λ = 3, the algorithm balances sufficient exploration in the early stage with high convergence accuracy in the later stage; when λ = 5, although the convergence is relatively fast, there is a risk of premature convergence in the local stage; when λ = 1, the exploration time in the early stage is too long, and the convergence speed is slow.

[0096] In practical applications, the above parameters can be adaptively adjusted according to specific working conditions and robot performance indicators. The system has a parameter adaptive adjustment submodule, which can reverse-calibrate each parameter based on actual execution data after each task, thereby achieving continuous parameter optimization.

[0097] Therefore, this invention, by collecting structured data of coke blocks and initializing the coke removal task, enables path planning to no longer rely on preset fixed trajectories but can instead model the task based on the actual distribution of coke blocks in the furnace, thereby improving the targeting of path planning. By determining the boundaries of each sub-region through a four-quadrant region segmentation strategy based on coke block workload, it helps to improve the problem of uneven workload between different regions, avoids excessive concentration of tasks in local areas, and improves the overall rationality of task scheduling. By constructing a multi-objective path planning cost function that includes robot movement cost, coke removal operation cost, and coke block priority cost, it can reduce ineffective idle travel and improve the priority processing capability of key coke blocks and critical areas. By adopting a dual-mode adaptive optimization algorithm that includes random exploration and local refinement, it takes into account both global search capability and local convergence capability, which helps to improve the quality and efficiency of path solution. By updating the path in conjunction with environmental feedback during task execution, it can enhance the robot's adaptability to complex changes in furnace conditions and improve the stability and reliability of coke removal operations.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0099] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A path planning method for a coke removal robot in an incinerator based on a dual-mode adaptive optimization algorithm, characterized in that, include: Step S1: Obtain the characteristic information of each coke block on the inner wall of the furnace. The characteristic information includes at least the coke block location information and the evaluation index reflecting the priority of coke block cleaning. The coke removal task is initialized based on the characteristic information. Step S2: Based on the coke block location information, the working area to be decoked in the furnace is divided into multiple sub-areas, and the boundary of each sub-area is determined with the balance of workload between each sub-area as the dividing objective. Step S3: Construct a multi-objective evaluation function for evaluating the quality of the desiccation path. The multi-objective evaluation function includes at least a movement cost component that represents the cost of robot movement, a time cost component that represents the time cost of desiccation operation, and a priority cost component that represents the priority cost of the desiccation block. Step S4: Apply the dual-mode adaptive optimization algorithm to solve for the optimal defocusing path using the multi-objective evaluation function as the fitness function. The dual-mode adaptive optimization algorithm performs a random search in the solution space through a random exploration mode and a fine search near the current optimal solution through a local exploration mode. The execution ratio of the random exploration mode and the local exploration mode is controlled by adaptive parameters that change dynamically with the iteration process. Step S5: Control the robot to perform the decoking operation according to the optimal decoking path.

2. The method according to claim 1, characterized in that, Also includes: Step S6: During the decoking operation, adjust the path parameters based on the real-time monitored environmental feedback information, and record the execution data after the decoking operation is completed to optimize the system parameters.

3. The method according to claim 1, characterized in that, In step S1, the evaluation indicators are obtained in the following way: The acquired images of the furnace inner wall are preprocessed, and the coke block areas in the images are identified. Extract image features from each focal block region, the image features including color features, texture features and / or size features; The positional importance weight is determined based on the location of the coke block on the inner wall of the furnace. The evaluation index is calculated based on the image features and positional importance weights of each focal block region.

4. The method according to claim 1, characterized in that, In step S2, determining the boundaries of each sub-region includes: The individual workload of each coke block is calculated based on its area, type coefficient, and evaluation indicators. The boundary positions of each sub-region are determined by minimizing the standard deviation of the workload of all sub-regions as the objective function. The workload of a single sub-region is the sum of the individual workloads of all the coke blocks contained in that sub-region.

5. The method according to claim 4, characterized in that, Step S2 employs a four-quadrant region segmentation strategy to divide the work area into four sub-regions, with the goal of minimizing the standard deviation of workload in each of the four sub-regions. in, The standard deviation of workload; Minimize the function; Calculate the square root; For the first The estimated total workload for each sub-region, where k takes values ​​in the range [1,4]; The average estimated workload for all sub-regions; , coke area, coke Type coefficient, coke Severity index The serial number of the coke block; A one-dimensional search method is used to dynamically determine the boundaries of each sub-region.

6. The method according to claim 1, characterized in that, The multi-objective evaluation function is: J = a - C m + β - C t + γ - C p Wherein, J is the total cost; a, b, g are weight coefficients; C m is the moving cost component; C t is the defocus time cost component; C p is the priority cost component; the priority cost component is a negative correlation function, so that the higher the evaluation index of the focus block, the lower the corresponding priority cost.

7. The method according to claim 6, characterized in that, The formulas for calculating each cost component are as follows: Moving cost components: in, Let be the Euclidean distance between adjacent focal blocks i and j in the path sequence; η is the turning angle between adjacent path segments at focal block i and focal block j in the path sequence; η is the turning penalty coefficient, used to convert the turning angle into an equivalent additional movement distance with the same dimension as the Euclidean distance; De-coking time cost components: Among them, A i coke i The area, K i coke i Type coefficient, K m The coefficient of performance of the heated surface material. To improve the efficiency of decoking operations; Priority cost components: Among them, S i coke i Evaluation indicators.

8. The method according to claim 1, characterized in that, The dual-mode adaptive optimization algorithm includes an initialization phase and an iterative optimization phase; Initialization phase: A set of initial solutions is randomly generated as the initial population, and algorithm parameters are set, including the total number of iterations T and the maximum decay parameter e. max and minimum attenuation parameter e min ; Iterative optimization phase: In each iteration, the adaptive parameter e(t) is determined based on the current iteration number t, and the search mode to be used in this iteration is determined from the random exploration mode and the local exploration mode based on the value of e(t), generating new candidate solutions; The fitness value of the new candidate solution is compared with the fitness value of the current solution. If the fitness value of the new candidate solution is better than the fitness value of the current solution, then the current solution is replaced by the new candidate solution. After the population update is complete, determine whether the termination condition is met. If so, output the current optimal solution as the optimal path to remove the focus. The fitness value is the multi-objective evaluation function J, and the smaller the value, the better.

9. The method according to claim 8, characterized in that, The formula for generating candidate solutions in the random exploration mode is: in, To randomly explore newly generated candidate solutions; x is the current solution; a and b are both exploration intensity coefficients; and All numbers are random numbers in the range [0,1]; L(β) is the random step size following a Lévy distribution, where β is the Lévy distribution exponent, ranging from (1,2]; δ is the step size scaling factor that decreases with the iteration process; x max and x min These are the upper and lower bounds of the solution space, respectively; The formula for generating candidate solutions in the local exploration mode is: in, To explore newly generated candidate solutions locally; is the historical best solution in the current population; c and d are development accuracy coefficients; rand is a random number in the range [0,1].

10. The method according to claim 8, characterized in that, The formula for calculating the adaptive parameter e(t) is: Where λ is the decay constant and T is the total number of iterations; the value of e(t) decays as the number of iterations increases, and the algorithm dynamically adjusts the execution ratio of random exploration mode and local exploration mode according to the value of e(t): when the value of e(t) is greater than the preset threshold, the current iteration executes random exploration mode; when the value of e(t) is not greater than the preset threshold, the current iteration executes local exploration mode.

Images