A data-driven portable car washing dynamic cleaning strategy optimization method

Abstract

The application discloses a kind of based on data-driven portable car washing dynamic cleaning strategy optimization method, it is related to artificial intelligence technical field, including the following steps: generating vehicle dirty detection data, establish cleaning action library and cleaning rule table;Vehicle dirty detection data is executed regional risk weighting, generates necessary cleaning action set;Binary cleaning action decision vector is constructed, and initial legal particle population is obtained, and through fitness evaluation, time window urgency adjustment and sequence constraint repair, the executable cleaning action instruction list is output;Vehicle cleaning strategy online execution and feedback update.The application is by fusing the double-layer weighting model of regional security weight and dirt hazard index, and it is supplemented with local high-value aggregation algorithm, constructs the dirt urgency evaluation mechanism that can accurately reflect safety priority, realizes the automatic balance of effect and efficiency by controlling particle swarm inertia weight, time penalty term and the greedy clipping based on marginal loss rate.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a data-driven, convenient method for optimizing dynamic car wash cleaning strategies. Background Technology

[0002] With the continuous growth of car ownership and the increasing demands of users for service quality, the traditional fixed-package car wash model is becoming increasingly difficult to adapt to complex actual working conditions. Under the current technological system, most automatic car wash machines and store service processes still adopt a rigid "one car, one policy" logic, that is, regardless of the actual dirtiness of the vehicle, a preset standardized procedure is executed. This model leads to significant resource misallocation and safety risks.

[0003] Existing technologies have the following drawbacks: dirt assessment is often based on only a single dimension of area or concentration, failing to quantify and integrate the driving safety level of vehicle body areas with the hazard level of dirt types, resulting in cleaning priority determination being out of touch with actual safety needs; when calculating the combined cleaning effect, most methods use simple linear summation, lacking a realistic portrayal of the diminishing marginal effect produced by repeated superposition of multiple actions, leading to mathematically distorted resource allocation schemes; when faced with tight delivery times, a one-size-fits-all, coarse-grained truncation strategy is often adopted, failing to finely balance time and effect in complex action combinations, and making it difficult to find the trimming scheme with the least effect loss in unnecessary steps; optimization algorithms are out of touch with various real-world engineering rules, and the output results often violate standard operating procedures or vehicle limitations, requiring inefficient manual correction before being converted into executable instructions. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a data-driven, convenient dynamic cleaning strategy optimization method for car washing.

[0005] The technical solution adopted to solve the above technical problems is: a data-driven, convenient dynamic car wash cleaning strategy optimization method, including the following steps:

[0006] S1 generates vehicle dirt detection data and establishes a cleaning action library. With cleaning rules table;

[0007] S2 performs regional risk weighting based on vehicle dirt detection data to generate a set of necessary cleaning actions. ;

[0008] S201, regarding the original dirtiness matrix By applying a two-level weighting of regional safety function weights and the contamination hazard index, a weighted contamination urgency matrix is ​​obtained. ,

[0009] S202, from the weighted dirt urgency matrix The urgency features of each type of dirt are extracted in local high-urgency areas and across the entire vehicle, and the two are weighted and fused to generate a global dirt severity vector that comprehensively reflects the global urgency of dirt treatment for that type of dirt. ;

[0010] S203, generates the necessary cleaning action set by jointly determining the action activation threshold, the high-level forced trigger judgment of dirt type, and the forced trigger judgment of local key areas. ;

[0011] S3, Constructing the binary cleaning action decision vector The initial legal particle population is obtained. Through fitness evaluation, time window urgency adjustment and sequence constraint repair, a list of executable cleaning action instructions is output.

[0012] S4, Online Execution and Feedback Update of Vehicle Cleaning Strategy: Automatically generates an original dirt level matrix upon entering the detection area. The necessary set of cleaning actions is calculated, the list of cleaning action instructions is optimized, and the parameters of the cleaning action library are updated based on the execution feedback.

[0013] Furthermore, the method for generating vehicle dirt detection data in S1 includes the following steps:

[0014] S10101, Acquire multi-view images of the vehicle. , This refers to the set of images of the vehicle to be cleaned, captured by a fixed shooting station or an onboard vision device. The image set includes images of the front of the vehicle, the rear of the vehicle, the left side of the vehicle, the right side of the vehicle, the roof of the vehicle, partial images of the wheel hubs, and partial images of the windshield.

[0015] S10102, vehicle multi-view images Input the vehicle image recognition model, which is used to identify the vehicle's outer contour, body functional areas, and dirty areas. The vehicle image recognition model outputs the body functional area category and dirt type category for each pixel. The body functional area categories include the windshield, rear windshield, headlights, taillights, rearview mirrors, door handles, body sides, roof, wheel rims, lower skirts, and bumpers. The dirt type categories include dust, mud spots, bird droppings, oil film, tree sap, water stains, wheel rim oil stains, and crevices dirt.

[0016] S10103 divides the vehicle surface into multiple body area grids and establishes the correspondence between the body area grids and the body functional areas;

[0017] S10104, Based on the output of the vehicle image recognition model, generate the original dirt level matrix. , This represents the set of dirt levels across all vehicle body grid areas and all dirt types, with a size of [size missing]. , This indicates the total number of categories of dirt / soil. Indicates a dirty type index. , Indicates the first The grid of the vehicle body area in the first... The original level of dirtiness for each type of dirtiness;

[0018] S10105, Obtain vehicle characteristic data table, order constraint data table, store resource data table and weather environment data table;

[0019] The vehicle characteristic data table includes vehicle size, paint type, whether it is tinted, vehicle age, whether it is allowed to be heavily scrubbed, and whether it is allowed to be washed at close range with high pressure.

[0020] The order constraint data table includes the remaining available time of the order, the service level selected by the user, whether dry delivery of the vehicle body is required, and whether delayed delivery is allowed.

[0021] The store resource data table includes the current availability status of workstations, remaining cleaning agent, water and electricity resource status, and available manpower.

[0022] The aforementioned weather and environmental data table includes temperature, humidity, wind speed, and rainfall status.

[0023] Furthermore, a cleaning action library is established in S1. The method for using the cleaning rule table includes the following steps:

[0024] S10201, Establish a cleaning action library , This represents the complete set of cleaning actions available. Indicates the first Each cleaning action, Indicates the cleaning action index. ,

[0025] Indicates the total number of cleaning actions;

[0026] S10202, for cleaning the action library Each cleaning action is configured with standard process parameters, including standard time, standard water and electricity costs, chemical type, standard chemical consumption, applicable vehicle body functional areas, main types of dirt to be treated, and precautions.

[0027] S10203, configures the dirt removal rate for each cleaning action. , Indicates the first The first cleaning action is for the first The projected removal capacity for each type of dirt;

[0028] S10204, Configure the action area coverage coefficient for each cleaning action. , Indicates the first The first cleaning action is for the first The coverage capability of the grid in each vehicle body area;

[0029] S10205, Establish an action pre-dependency rule table and an action mutual exclusion restriction rule table. The action pre-dependency rule table is used to record the basic cleaning actions that must be executed before a certain cleaning action is executed. The action mutual exclusion restriction rule table is used to record cleaning actions that should not be executed simultaneously or are prohibited from being executed under specific vehicle conditions.

[0030] Furthermore, in S201, the original dirtiness matrix... By applying a two-level weighting of regional safety function weights and the contamination hazard index, a weighted contamination urgency matrix is ​​obtained. The method includes the following steps:

[0031] S20101, Read the original dirt level matrix And confirm the original dirt level matrix. The size is ;

[0032] S20102, Configure regional safety function weights for each vehicle body area grid. , Indicates the first The importance of each body area grid to driver visibility, driving safety, vehicle operation, and user perception;

[0033] S20103, configures a dirt hazard index for each type of dirt. , Indicates the first The intensity of the impact of different types of dirt on driving visibility, paint protection, cleaning difficulty, corrosion risk, and user complaint risk;

[0034] S20104, regarding the original dirtiness matrix Perform a two-level weighting to obtain a weighted dirt urgency matrix. , This represents the weighted set of processing urgency for the current vehicle across all body area grids and all dirt types, with a size of [size missing]. , Indicates the first The grid of the vehicle body area in the first... Weighted dirt urgency values ​​for different dirt types.

[0035] Furthermore, in S2, a global severity vector of contamination is generated that comprehensively reflects the urgency of global treatment for this type of contamination. The method includes the following steps:

[0036] S20201, for the first Types of dirt, from the weighted dirt urgency matrix Extract the first Column, to get the first The urgency sequence of different types of pollution, the first The urgency sequence of different types of dirt includes Each weighted dirt urgency value corresponds to one vehicle body area grid.

[0037] S20202, regarding the first Sort the areas with varying degrees of urgency based on their type of contamination from highest to lowest, and select the top... A weighted dirt urgency value was calculated, and the previous... The average of the weighted dirt urgency values ​​is used to obtain the local high severity value. , Indicates the first The severity of each type of dirt on a few high-pressure vehicle body area grids is used to capture dirt in localized critical areas;

[0038] S20203, calculate the first The overall severity distribution is obtained by averaging the dirt severity of each dirt type across all vehicle body grid areas. ,Right now , Indicates the first The average urgency of a type of dirt across the entire vehicle is used to reflect whether that type of dirt is present over a large area.

[0039] S20204, the severity of local high values and overall distribution severity Weighted fusion yields the first Global severity index for each type of dirt , Indicates the first The urgency of addressing different types of dirt in the current vehicle cleaning task;

[0040] S20205, repeat S20201~S20204 to traverse all dirt types and obtain the global dirt severity vector. , Indicates by A vector consisting of global severity indices of dirtiness.

[0041] Furthermore, in S203, a set of necessary cleaning actions is generated. The method includes the following steps:

[0042] S20301, for cleaning the action library Each cleaning action in the configuration is associated with a set of dirt types. , Indicates the first Each cleaning action primarily deals with a set of types of dirt;

[0043] S20302, Configure a low alert threshold for each type of dirt. and high forced threshold , Indicates the first The severity level of the type of dirt that needs to be addressed during cleaning. Indicates the first The severity level at which the type of dirt must trigger an associated cleaning action;

[0044] S20303, based on the global severity vector of contamination Calculate the type activation value for each type of dirt. , Indicates the first Types of dirt relative to low alert thresholds and high forced threshold The degree of activation;

[0045] S20304, according to the... A set of associated dirt types for each cleaning action Calculate action activation score , Indicates the first The degree to which a cleaning action is triggered by the current state of dirt on the vehicle;

[0046] S20305, Configure an action activation threshold for each cleaning action. , Indicates the first Each cleaning action is added to the minimum action activation score of the required cleaning action set;

[0047] S20306, execute the high-forced trigger judgment of dirt type;

[0048] S20307, ​​Perform forced trigger judgment for local critical areas;

[0049] S20308, Set of Necessary Cleaning Actions Perform prerequisite dependency supplementation.

[0050] Furthermore, in S3, a binary cleaning action decision vector is constructed. The initial legal particle population is obtained. The method includes the following steps:

[0051] S30101, Constructing the binary cleaning action decision vector , This represents a candidate combination of cleaning actions. Indicates the first The execution status of each cleaning action. Indicates the cleaning action library The total number of cleaning actions in the process;

[0052] S30102, set of necessary cleaning actions The corresponding action dimension is forcibly set to 1;

[0053] S30103 is a set of cleaning actions that are not required. Calculate the initial selection probability for the cleaning action. , Indicates the first The probability of a non-essential cleaning action being selected from the initial candidate cleaning action combination;

[0054] S30104, generate a particle population, the total number of particles is denoted as... , This represents the number of candidate cleaning action combinations participating in binary particle swarm optimization.

[0055] S30105, perform order constraint repair on each candidate binary cleaning action decision vector to obtain a legal binary cleaning action decision vector.

[0056] S30106, After completing the sequence constraint repair, the necessary cleanup action set will be re-selected. The corresponding action dimension is forcibly set to 1 to obtain the initial legal particle population. , Let represent the set of decision vectors for the 0th generation of legal binary cleaning actions.

[0057] Furthermore, the fitness evaluation method in S3 includes the following steps:

[0058] S30201, Read the cleaning action library Dirt removal rate in the process and action area coverage coefficient ;

[0059] S30202, based on the binary cleaning action decision vector Calculate the cleaning effect score , This indicates the estimated removal value of the current cleaning action combination for the vehicle's dirt and grime.

[0060] S30203, based on the binary cleaning action decision vector Calculate the estimated total time , This indicates the estimated execution time of the current cleaning action combination under the current vehicle, current weather, and current store workstation conditions;

[0061] S30204, based on the binary cleaning action decision vector Computational resource consumption cost , This indicates the combined cost of water, electricity, chemicals, and labor resources expected to be consumed in the current cleaning operation combination;

[0062] S30205, Calculate the time window penalty value based on the order constraint data table. , This indicates the penalty incurred when the current cleaning action combination exceeds the order time window;

[0063] S30206, Calculate the time penalty weight based on the remaining time window. , This indicates the strength of the impact of the time window penalty value on fitness evaluation;

[0064] S30207, based on cleaning effectiveness score Resource consumption costs and time window penalty value Obtain fitness value , Represents the binary cleaning action decision vector Considering all advantages and disadvantages, the optimization objective is to maximize the fitness value. ;

[0065] The method for adjusting the urgency of the time window and repairing the sequence constraints in S3 includes the following steps:

[0066] S30301, Set the maximum number of iterations , This represents the maximum update round number in binary particle swarm optimization, at the [number]th [stage]. The time window urgency factor is calculated at the start of the next iteration. , Indicates the first The urgency of the remaining time for the order in the next iteration;

[0067] S30302, based on the time window urgency coefficient Adjusting inertia weight , Indicates the first The inertia weight of the nth iteration, characterizing the... The degree to which historical search directions are retained in the particle velocity update during each iteration;

[0068] S30303, regarding the first The first particle Update speed of each action dimension , Indicates the first During the nth iteration The particle in the first The velocity value in each action dimension is used to determine the tendency of that action dimension to take a value of 1 in the next sampling.

[0069] S30304, speed value Converted into action execution probability, the action execution probability represents the first action. The particle in the first The probability of taking a value of 1 in each action dimension;

[0070] S30305, Randomly sample non-essential cleaning actions based on the action execution probability to obtain a temporary binary cleaning action decision vector. Perform order constraint repair on the temporary binary cleaning action decision vector, repairing action pre-dependencies, action mutual exclusion restrictions, vehicle feature restrictions, and resource unavailability restrictions, and once again enforce the maintenance of the essential cleaning action set. The corresponding action dimension is 1;

[0071] S30306, Calculate the fitness value corresponding to the repaired binary cleaning action decision vector. ;

[0072] S30307 involves repeatedly performing speed updates, action execution probability transformations, random sampling, sequence constraint repair, and fitness evaluation until the maximum number of iterations is reached. .

[0073] Furthermore, the method for outputting the list of executable cleaning action instructions in S3 includes the following steps:

[0074] S30401, after the iteration, the globally optimal cleaning action decision vector is obtained. , This represents the binary cleaning action decision vector with the highest overall evaluation obtained during the binary particle swarm optimization process.

[0075] S30402, if the globally optimal cleaning action decision vector Corresponding estimated total time Less than or equal to the maximum allowable time for the order time window Then directly follow the cleaning action library. The standard construction sequence converts cleaning actions with a value of 1 into a list of cleaning action instructions;

[0076] S30403, Calculate the marginal effect loss rate for each removable cleaning action. , Indicates the removal of the first The loss of cleaning effect caused by saving time per cleaning action;

[0077] S30404, according to the marginal utility loss rate Remove unnecessary cleaning actions in ascending order of their complexity. After removing each unnecessary cleaning action, recalculate the estimated total time. and cleaning effect score until the estimated total time Less than or equal to the maximum allowable time for the order time window ;

[0078] S30405, Even after all non-essential cleaning actions are removed, the estimated total time is still greater than the maximum allowable time within the order time window. If the minimum required cleaning plan is not found, then the insufficient time window will be marked in the cleaning action instruction list.

[0079] S30406, Output a list of cleaning action instructions: The list includes the name of the cleaning action, execution order, estimated time, main type of dirt to be treated, corresponding functional area of ​​the vehicle body, whether it is a necessary cleaning action, the reason for recommendation, and precautions. The list of cleaning action instructions is based on the cleaning action library. The standard construction sequence is arranged and directly sent to the car wash workers' mobile phones, workstation displays, or automatic car wash control systems.

[0080] Furthermore, the method for online execution and feedback update of the vehicle cleaning strategy in S4 includes the following steps:

[0081] S40201 After cleaning is completed, collect execution result data, which includes the actual time consumed for each cleaning action, the actual amount of chemical agent consumed, whether rework occurred, user satisfaction score, and post-cleaning inspection images.

[0082] S40202, calculate the degree of residual dirt based on the re-inspection image after cleaning;

[0083] S40203, update the standard time based on the actual time consumed;

[0084] S40204, Update the dirt and hazard index based on rework records. and action activation threshold .

[0085] The beneficial effects of the present invention are as follows: (1) The present invention integrates the regional safety weight and the dirt hazard index into a two-layer weighted model, and is supplemented by a local high value aggregation algorithm to construct a dirt urgency assessment mechanism that can accurately reflect the safety priority.

[0086] (2) When calculating the cleaning effect of multi-action combination, the present invention introduces a marginally decreasing model based on probability multiplication and accumulation, which effectively suppresses the tendency of the optimization algorithm to pile up repetitive and inefficient actions for falsely high scores.

[0087] (3) The dynamic adaptive optimization algorithm driven by the remaining time of the order in this invention achieves an automatic balance between effect and efficiency by controlling the particle swarm inertia weight, time penalty term and greedy pruning based on marginal loss rate.

[0088] (4) The present invention abstracts the complex construction process rules into a matrix of prerequisite dependencies and mutual exclusion constraints, and embeds them into the constraint repair operator of the optimization iteration, ensuring that the output scheme naturally satisfies all real constraints without the need for manual secondary verification. Attached Figure Description

[0089] Figure 1 This is a flowchart of an embodiment of the present invention.

[0090] Figure 2 This is a heat map of the dirt urgency matrix after double-layer weighting according to the present invention.

[0091] Figure 3 This is a grouped bar chart showing the severity of local high values ​​and the overall distribution of each type of dirt in this invention.

[0092] Figure 4 This is a histogram of the global severity vector of dirt in this invention.

[0093] Figure 5 It is a graph showing the type activation value, low alert threshold, and high mandatory threshold for each type of dirt.

[0094] Figure 6 This is a horizontal histogram of the candidate binary cleaning action decision vectors for particle 1.

[0095] Figure 7 This is a horizontal histogram of the decision vectors for the candidate binary cleaning actions of particle 2.

[0096] Figure 8 This is a horizontal histogram of the decision vectors for the candidate binary cleaning actions of particle 3.

[0097] Figure 9 This is a horizontal histogram of the decision vectors for the candidate binary cleaning actions of particle 4. Detailed Implementation

[0098] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0099] like Figure 1 As shown in the figure, a data-driven, convenient dynamic car wash cleaning strategy optimization method according to this embodiment includes the following steps:

[0100] S1 generates vehicle dirt detection data and establishes a cleaning action library. With the cleaning rules table.

[0101] Vehicle dirt is not evenly distributed throughout the vehicle, and different areas of the vehicle and different types of dirt have different impacts on safety, paint protection, and user experience. Therefore, it is necessary to establish a unified data input format and cleaning action rules as the basis for subsequent dirt urgency calculation and cleaning strategy optimization.

[0102] The method for generating vehicle dirt detection data includes the following steps:

[0103] S10101, acquire multi-view images of the vehicle, the multi-view images of the vehicle are denoted as... , This refers to the set of images of the vehicle to be cleaned, captured by a fixed shooting station or on-board vision equipment. The image set includes images of the front of the vehicle, the rear of the vehicle, the left side of the vehicle, the right side of the vehicle, the roof of the vehicle, partial images of the wheel hubs, and partial images of the windshield.

[0104] In practice, each image undergoes uniform scaling, brightness normalization, and distortion correction before being input into the vehicle image recognition model, ensuring that images from different shooting distances and under different lighting conditions have a consistent input scale.

[0105] S10102, vehicle multi-view images Input the vehicle image recognition model, which is used to identify the vehicle's outer contour, body functional areas, and dirty areas. The vehicle image recognition model outputs the body functional area category and dirt type category for each pixel. The body functional area categories include the windshield, rear windshield, headlights, taillights, rearview mirrors, door handles, body sides, roof, wheel rims, lower skirts, and bumpers. The dirt type categories include dust, mud spots, bird droppings, oil film, tree sap, water stains, wheel rim oil stains, and crevices dirt.

[0106] Vehicle image recognition models can be implemented using semantic segmentation models or a combination of object detection and region segmentation models. In this embodiment, the vehicle image recognition model is implemented using the DeepLabV3+ semantic segmentation model. This model structure uses ResNet-50 as the backbone network to extract multi-scale features of vehicle images. It is followed by a spatial pyramid pooling module with dilated convolutions to capture contextual information at multiple sampling rates. Finally, a decoder module fuses the feature map output from the pyramid pooling with the low-level feature map of the backbone network and upsamples it to the original image resolution to achieve dense classification of each pixel. During model training, a vehicle image dataset containing dirt type annotations is collected. Each image is pixel-by-pixel labeled with its corresponding vehicle body functional area category and dirt type category. Training is supervised using the cross-entropy loss function, and the optimizer can be stochastic gradient descent with momentum. After training, the model input is a preprocessed single local or overall vehicle image, and the output is a two-channel segmentation map of the same size as the input image: one channel stores the vehicle body functional area category index for each pixel, and the other channel stores the dirt type category index for each pixel. In practice, the vehicle image recognition model can be deployed on a cloud server and inference can be performed by calling the GPU through store workstation cameras or mobile terminals.

[0107] S10103 divides the vehicle surface into multiple body area grids and establishes the correspondence between the body area grids and the functional areas of the vehicle body. The total number of body area grids is denoted as... , This indicates the number of grid regions that the current vehicle surface is divided into, the i-th The index of each vehicle body area grid is denoted as , .

[0108] In practical implementation, safety-critical areas such as the windshield, rearview mirrors, headlights, and door handles can be divided into smaller body area meshes. For example, each headlight can be divided into 4 body area meshes, the windshield into 12 body area meshes, and areas such as the roof, sides, and lower skirts can be divided into relatively larger body area meshes to reduce computational load.

[0109] S10104, Based on the output of the vehicle image recognition model, generate the original dirt level matrix. , This represents the set of dirt levels across all vehicle body grid areas and all dirt types, with a size of [size missing]. , This indicates the total number of categories of dirt / soil. Indicates a dirty type index. , Indicates the first The grid of the vehicle body area in the first... The original dirt level value for each type of dirt is within a certain range. arrive The larger the value, the higher the coverage area, color depth, or recognition confidence of the dirt type within the grid area of ​​the vehicle body.

[0110] In specific implementation, for the first The body area grid and the first First, count the types of dirt and grime identified in the grid area of ​​the vehicle body. The percentage of pixel area for each type of dirt is used to obtain the percentage of dirt area. , Indicates the first Within the grid of the vehicle body area, the first The average dirt confidence score is obtained by calculating the ratio of the number of pixels of each type of dirt to the total number of pixels in the mesh of that vehicle body area, and then calculating the average recognition confidence score of these dirt pixels. , This indicates that the vehicle image recognition model is for the first... Within the grid of the vehicle body area, the first The average reliability of the results for determining the type of dirtiness is expressed as follows: The calculation method for the original dirtiness level value is as follows: ,in, This represents the area proportion adjustment coefficient, used to balance the impact of coverage area and identification confidence level on the original dirtiness value, for example... 0.7 is acceptable.

[0111] In practical implementation, the average recognition confidence level Based on the output of the vehicle image recognition model, the semantic segmentation model outputs a probability vector when classifying each pixel. The category corresponding to the maximum value is the predicted category of that pixel, and this maximum value is the prediction confidence of that pixel. For the For each of the vehicle body area grids, count all those predicted as the [number]th [unit]. For each type of dirt, calculate the arithmetic mean of the confidence scores for those pixels; this yields the average dirt confidence score. .

[0112] In this embodiment, as an example, if the area ratio of oil film pixels in a certain windshield body area grid is... Average soiling confidence level ,Pick The original dirt level value of the mesh area on the vehicle body in terms of oil film dirt type is then determined. This value reflects both the oil film coverage area and the reliability of the identification, avoiding misjudgment based solely on a small number of high-confidence pixels or a large area of ​​low-confidence regions.

[0113] S10105: Obtain vehicle characteristic data, order constraint data, store resource data, and weather environment data. The vehicle characteristic data table includes vehicle size, paint type, whether it has been tinted, vehicle age, whether strong brushing and high-pressure close-range washing are permitted. The order constraint data table includes remaining order time, the user's selected service level, whether dry delivery is required, and whether delayed delivery is permitted. The store resource data table includes the current workstation availability, cleaning agent remaining, water and electricity availability, and manpower availability. The weather environment data table includes temperature, humidity, wind speed, and rainfall status, which will be used in subsequent cleaning action selection, time estimation, and feasibility repair processes.

[0114] The method for establishing a cleaning action library and a cleaning rule table includes the following steps:

[0115] S10201, Establish a cleaning action library , This represents the complete set of cleaning actions available. Indicates the first Each cleaning action, Indicates the cleaning action index. ,

[0116] This indicates the total number of cleaning actions.

[0117] In practical implementation, the cleaning action library Arranged according to standard construction sequence, such as cleaning action library The process includes, in sequence, pre-rinsing, foam soaking, soft water washing, high-pressure washing, wheel cleaning, degreasing of glass, fine cleaning of crevices, water wax coating, strong air drying, and manual repair.

[0118] S10202, for cleaning the action library Each cleaning action is configured with standard process parameters, including standard time, standard water and electricity costs, chemical type, standard chemical consumption, applicable vehicle body functional areas, main types of dirt to be treated, and precautions.

[0119] For example, the glass oil film remover is suitable for the windshield and rear windshield, primarily treating oil film. The wheel cleaning service is suitable for the wheels, primarily treating oil stains and mud. The foam soaking service is suitable for most exterior surfaces, primarily treating bird droppings, tree sap, mud, and dust.

[0120] It should be noted that the standard process parameters provide a baseline quantitative description for each cleaning action, supporting multiple calculations in subsequent steps. Specifically, the standard time consumption is used to calculate the estimated total time consumption. This serves as the basis for vehicle size correction and weather correction. Standard water and electricity costs, chemical types, and standard chemical consumption are used to calculate resource consumption costs. The applicable vehicle body functional areas and the main types of dirt to be treated are used to configure the associated dirt type set. This also helps determine the matching relationship between cleaning actions and dirt. The "Notes" section is used to output a list of cleaning action instructions, providing operators with key information.

[0121] S10203, configures the dirt removal rate for each cleaning action. , Indicates the first The first cleaning action is for the first The projected removal capacity for each type of dirt. The value ranges from 0 to 1.

[0122] In one implementation, the action dirt removal rate can be preset based on the store's historical cleaning results, technician experience, or test calibration results. For example, the action dirt removal rate for mud spots by high-pressure washing can be 0.85, the action dirt removal rate for floating dust can be 0.80, the action dirt removal rate for oil film can be 0.20, the action dirt removal rate for oil film by glass degreasing can be 0.90, and the action dirt removal rate for mud spots can be 0.10.

[0123] S10204, Configure the action area coverage coefficient for each cleaning action. , Indicates the first The first cleaning action is for the first The coverage capability of the mesh for each vehicle body area, with a value ranging from 0 to 1. When the... The first cleaning action can completely cover the second When the body area grid is used, Take 1. When the first... The first cleaning action cannot cover the second. When the body area grid is used, Take 0. When the first The first cleaning action can only partially cover the second. When the body area grid is used, A value of 0.3 to 0.8 is acceptable.

[0124] In practical implementation, The process is pre-configured and stored in a rule table based on equipment specifications and construction experience using a manual preset method. For example, the action area coverage coefficient for wheel hub cleaning is set to 1 for the wheel hub body area grid and 0 for the windshield body area grid. Similarly, the action area coverage coefficient for glass degreasing is set to 1 for the windshield body area grid and 0 for the roof body area grid.

[0125] S10205, Establish an action pre-dependency rule table and an action mutual exclusion restriction rule table. The action pre-dependency rule table is used to record the basic cleaning actions that must be executed before a certain cleaning action is executed. For example, water wax plating requires rinsing and drying actions to be executed first, strong wind drying requires at least one rinsing action to exist beforehand, and manual fine touch-up needs to be executed after the main cleaning action.

[0126] The action mutual exclusion restriction rule table is used to record cleaning actions that should not be performed simultaneously or are prohibited under specific vehicle conditions. For example, high-temperature glass treatment is prohibited for vehicles with tinted windows, strong brushing is prohibited for vehicles with special paint, and close-range high-pressure washing is restricted for vehicles with high age.

[0127] It should be noted that the action pre-dependency rule table and the action mutual exclusion restriction rule table are used to ensure that the cleaning action combination obtained from subsequent optimization can be directly implemented. Both rule tables are preset manually, transforming the car wash standard operating procedures and vehicle safety restrictions into structured rule entries and storing them in the database.

[0128] S2 performs regional risk weighting based on vehicle dirt detection data to generate a set of necessary cleaning actions. .

[0129] Vehicle cleaning decisions should not be based solely on the average level of dirt across the entire vehicle. The windshield, rearview mirrors, headlights, and door handles directly affect driving visibility, road safety, and user experience; bird droppings, tree sap, and oil films are more likely to cause paint corrosion, obstruct vision, or require rework than ordinary dust.

[0130] S201, regarding the original dirtiness matrix By applying a two-level weighting of regional safety function weights and the contamination hazard index, a weighted contamination urgency matrix is ​​obtained. .

[0131] By introducing safety function weights for vehicle body areas and hazard indices for different types of dirt, the original dirt level matrix was analyzed. Double weighting is performed to generate a weighted dirt urgency matrix. This quantifies the urgency of treating contamination in different areas and of different types, and analyzes the original contamination level matrix. The method for implementing a two-tiered weighting of regional safety function weights and dirt hazard indices includes the following steps:

[0132] S20101, Read the original dirt level matrix And confirm the original dirt level matrix. The size is If a vehicle body area mesh fails to output a valid result due to occlusion or reflection, that mesh is marked as an area requiring manual review. The average original dirt level value of adjacent vehicle body area meshes within the same functional area is then used to temporarily fill the gap, preventing null values ​​from appearing in subsequent matrix calculations.

[0133] S20102, Configure regional safety function weights for each vehicle body area grid. , Indicates the first The importance of individual vehicle body grid patterns to driver visibility, driving safety, vehicle operation, and user perception. The value ranges from 0 to 1.

[0134] In one implementation, regional security function weights This can be obtained by looking up the weight table for the functional areas of the vehicle body. For example, the grid value for the windshield area can be 1.0, the grid value for the rearview mirror area can be 0.95, the grid value for the headlight area can be 0.90, the grid value for the door handle area can be 0.75, the grid value for the wheel hub area can be 0.60, the grid value for the side areas can be 0.50, the grid value for the roof area can be 0.30, and the grid value for the lower skirt area can be 0.40.

[0135] It should be noted that the vehicle body functional area weight table is a pre-set lookup table used to determine the area safety function weights corresponding to various vehicle body functional areas. This table is preset based on experience, taking into account factors such as driving safety, vehicle operation, and user experience.

[0136] S20103, configures a dirt hazard index for each type of dirt. , Indicates the first The impact of different types of dirt on driving visibility, paint protection, cleaning difficulty, corrosion risk, and user complaint risk. The value ranges from 0 to 1.

[0137] In one implementation, the dirt hazard index The hazard index can be obtained from a dirt type hazard table. For example, bird droppings can be assigned a hazard of 1.0 due to their corrosive nature; oil film can be assigned a hazard of 0.95 due to its obstruction of the windshield; tree sap can be assigned a hazard of 0.90 due to its strong adhesion; wheel rim oil stains can be assigned a hazard of 0.75 due to their impact on the appearance of the wheel rims; mud spots can be assigned a hazard of 0.60; water stains can be assigned a hazard of 0.50; and dust can be assigned a hazard of 0.30. If the store has a history of rework, the dirt hazard index can be adjusted at fixed intervals based on the rework rate. Updates are performed, but the dirt hazard index is adjusted during the same cleaning strategy generation process. It remains unchanged.

[0138] It should be noted that the dirt type hazard table is a pre-defined lookup table used to determine the dirt hazard index corresponding to each dirt type. This table is based on factors such as the corrosive effect of dirt on car paint, obstruction of vision, difficulty of cleaning, and risk of user complaints, and is preset by human experience.

[0139] S20104, regarding the original dirtiness matrix Perform a two-level weighting to obtain a weighted dirt urgency matrix. , This represents the weighted set of processing urgency for the current vehicle across all body area grids and all dirt types, with a size of [size missing]. , Indicates the first The grid of the vehicle body area in the first... Weighted dirt urgency values ​​for different dirt types The value ranges from 0 to 1.

[0140] In practical implementation, the regional security function weights can be calculated first. and dirt and filth hazard index The corresponding comprehensive weighting factor is then used to adjust the original dirtiness value. To avoid infinite amplification of the value, a normalized weighting method can be used. The calculation method is expressed as follows:

[0141]

[0142] in, Indicates the first The grid of the vehicle body area in the first... Weighted dirt urgency values ​​for different dirt types Indicates the first The grid of the vehicle body area in the first... The original dirt level value for each type of dirt. Indicates the first The regional safety function weights of each vehicle body area grid Indicates the first The hazard index of each type of dirt. This represents the area weighting amplification factor, used to control the influence of the importance of different vehicle body areas on the weighted dirt urgency value. For example, it can be set to 1.0. This represents the soil hazard amplification factor, used to control the intensity of the influence of soil hazard type on the weighted soil urgency value. For example, it can be set to 1.0. This means that the calculation result will be limited to the range of 0 to 1.

[0143] It should be noted that, This component represents a comprehensive weighting factor, which integrates the regional importance weight and the pollution hazard index to form a coefficient greater than 0, used to adjust the original pollution level value. When... and When the factor reaches its maximum value of 1, it is at its maximum; when it reaches its minimum value of 0, it is at its minimum value. .

[0144] In this embodiment, as an example, the original dirt level of the oil film on the mesh of a certain windshield body area is [value missing]. Safety function weighting of the corresponding area of ​​the windshield Oil film corresponds to dirt hazard index ,Pick and The overall weighting factor is The corresponding weighted dirt urgency value is 0.4875. If the mesh on the roof and body areas also exists... Dust, and the corresponding safety function weight of the roof area. Dust corresponds to the dirt hazard index The overall weighting factor is The corresponding weighted urgency value for soiling is 0.21125. Therefore, given the same initial soiling level, safety-critical areas and high-hazard soiling receive higher treatment priority.

[0145] It should be noted that, Represents the weighted dirt urgency matrix Included The matrix form consisting of all elements, including the matrix itself, is the standard matrix representation, meaning " It is The matrix whose first... Line 1 The elements of the column are Regional security function weights and dirt and filth hazard index It doesn't simply replace the original dirt level value, but rather incorporates both "where the dirt is located" and "what type of dirt is" into the cleaning strategy generation process, ensuring that localized safety-critical dirt is not masked by the overall vehicle average.

[0146] S202, from the weighted dirt urgency matrix The urgency features of each type of dirt are extracted in local high-urgency areas and across the entire vehicle, and the two are weighted and fused to generate a global dirt severity vector that comprehensively reflects the global urgency of dirt treatment for that type of dirt. .

[0147] S20201, for the first Types of dirt, from the weighted dirt urgency matrix Extract the first Column, to get the first The urgency sequence of different types of pollution, the first The urgency sequence of different types of dirt includes Each weighted dirt urgency value corresponds to one vehicle body area grid.

[0148] In one embodiment, for example, suppose the vehicle has 100 grid areas and 8 types of dirt, for the second type of dirt (e.g., "mud spots"). ),from Extracting the second column yields a sequence containing 100 weighted dirt urgency values, i.e. .

[0149] S20202, regarding the first Sort the areas with varying degrees of urgency based on their type of contamination from highest to lowest, and select the top... A weighted dirt urgency value was calculated, and the previous... The average of the weighted dirt urgency values ​​is used to obtain the local high severity value. , Indicates the first The severity of each type of dirt on a few high-pressure vehicle body area grids is used to capture dirt in localized critical areas.

[0150] in, This indicates the number of meshes used to aggregate locally high values ​​in the vehicle body area. Can be taken , Indicates the proportion of high-value areas, for example 0.05 is acceptable. This indicates rounding up to the nearest integer.

[0151] In this embodiment, the previous example is used. ,but Sort the 100 values ​​from the previous step in descending order, take the top 5 maximum values, and calculate the arithmetic mean of these 5 values. This gives the severity of the local high value of the mud spot. .

[0152] S20203, calculate the first The overall severity distribution is obtained by averaging the dirt severity of each dirt type across all vehicle body grid areas. ,Right now , Indicates the first The average urgency of a type of dirt across the entire vehicle is used to reflect whether that type of dirt is widespread.

[0153] S20204, the severity of local high values and overall distribution severity Weighted fusion yields the first Global severity index for each type of dirt , Indicates the first The degree of urgency of each type of dirt in the current vehicle cleaning task, ranging from 0 to 1, is expressed as follows:

[0154]

[0155] in, Indicates the first The urgency of addressing different types of dirt in current vehicle washing tasks. Indicates the first The severity of localized high values ​​for various types of dirt. Indicates the first The overall distribution and severity of various types of dirt and grime This represents the high-value severity fusion weight, used to control the influence of local key regions on the global severity index, for example... 0.6 is acceptable.

[0156] S20205, repeat S20201~S20204 to traverse all dirt types and obtain the global dirt severity vector. , Indicates by A vector composed of global severity indicators of dirtiness, with size [value missing]. Global severity vector of contamination This is used to determine whether each cleaning action needs to be enforced or prioritized.

[0157] In this embodiment, as an example, an oil film is only present in a localized area of ​​the windshield of a vehicle, with less oil film in other areas. If only the average grid value of the entire vehicle body area is used, the overall severity of the oil film distribution is... The value may be too low, causing the degreasing action on the glass to not be triggered. This invention calculates the severity of localized high values. The high weighted dirt urgency value of the mesh in the windshield body area will enter the front A weighted dirt urgency value determines the global severity index corresponding to the oil film. The height increases, making it easier to trigger the degreasing action on the glass.

[0158] It should be noted that the global severity vector of contamination It does not simply indicate the size of the dirty area, but rather the urgency of cleaning after the combined effects of "original degree of dirt, importance of the vehicle body area, severity of dirt, local high value distribution, and average distribution throughout the vehicle".

[0159] S203, generates the necessary cleaning action set by jointly determining the action activation threshold, the high-level forced trigger judgment of dirt type, and the forced trigger judgment of local key areas. .

[0160] This embodiment is based on the global severity vector of dirt. Based on the configured thresholds, the activation level of each type of dirt is quantified, and the action activation score is calculated by combining the correlation between cleaning actions. Finally, through threshold judgment and forced triggering rules, a set of mandatory cleaning actions that cannot be skipped is generated. The specific method includes the following steps:

[0161] S20301, for cleaning the action library Each cleaning action in the configuration is associated with a set of dirt types. , Indicates the first Each cleaning action primarily deals with a set of types of dirt.

[0162] In one implementation, a set of associated dirt types is used. Based on the main functions and standard operating procedures of the cleaning actions, the process is preset manually. For example, the associated dirt types for pre-rinsing may include dust and mud spots, the associated dirt types for foam soaking may include bird droppings, tree sap, and mud spots, the associated dirt types for removing oil film from glass may include oil film, the associated dirt types for wheel cleaning may include wheel oil stains and mud spots, and the associated dirt types for crevices cleaning may include crevices dirt and water stains.

[0163] S20302, Configure a low alert threshold for each type of dirt. and high forced threshold , Indicates the first The severity level of the type of dirt that needs to be addressed during cleaning. A value of 0.25 to 0.40 is acceptable. Indicates the first The severity level at which the type of dirt must trigger an associated cleaning action. For highly hazardous contaminants such as bird droppings, oil films, and tree sap, a lower low alert threshold can be set between 0.75 and 0.90. (e.g., 0.2) so that processing can be triggered earlier.

[0164] S20303, based on the global severity vector of contamination Calculate the type activation value for each type of dirt. , Indicates the first Types of dirt relative to low alert thresholds and high forced threshold The degree of activation, The value ranges from 0 to 1.

[0165] In specific implementation, when season .when season .when lie in and When the distance is between, it is obtained by mapping according to a linear ratio. ,Right now, For example, the low warning threshold of the oil film. High forced threshold If the global severity index of the oil film The activation value of the oil film type .

[0166] S20304, according to the... A set of associated dirt types for each cleaning action Calculate action activation score , Indicates the first The degree to which a cleaning action is triggered by the current state of dirt on the vehicle. The value ranges from 0 to 1.

[0167] In the actual implementation, the set of associated dirt types is first calculated. The average of all activation values ​​of all types is used to calculate the associated dirty type set. The maximum value of all activation values ​​of all types is calculated, and finally the average and the maximum value are weighted and combined. The calculation method is expressed as follows:

[0168]

[0169] in, This is the average value weighting coefficient, used to balance the influence of the mean and the maximum value; a value of 0.5 is acceptable. Represents a set The number of types of dirt, This indicates retrieving the set of associated dirt types. Type activation values ​​corresponding to all types of dirt The maximum value.

[0170] It should be noted that the average value is used to handle situations where multiple types of moderate dirt trigger the same action, while the maximum value is used to handle situations where a single high-risk dirt strongly triggers the same action.

[0171] In this embodiment, as an example, the associated dirt type set for high-pressure washing includes mud spots and dust. If the type activation value of mud spots is 0.9 and the type activation value of dust is 0.2, the action activation score of high-pressure washing will increase due to the higher maximum value, avoiding the dilution of heavy mud spots by the low average value of dust. If the type activation values ​​of both mud spots and dust are 0.5, the action activation score of high-pressure washing will also reach a moderate level, reflecting that various common dirt types require high-pressure washing treatment.

[0172] S20305, Configure an action activation threshold for each cleaning action. , Indicates the first Each cleaning action is added to the minimum action activation score of the required cleaning action set, for example... A score of 0.65 to 0.80 is acceptable. The score is determined when the action is activated. Greater than or equal to the action activation threshold At that time, the first Each cleaning action is added to the necessary cleaning action set. , This represents the set of cleaning actions that cannot be skipped in the current vehicle cleaning task.

[0173] In one embodiment, for example, suppose an action activation threshold is used. For "high-pressure washing", its action activation score Therefore, "high-pressure rinsing" was added to the essential cleaning action set. .

[0174] S20306, execute the high-forced trigger judgment of the dirt type.

[0175] Specifically, for the first Types of dirt, if the global severity index Greater than or equal to the high forced threshold Then all related numbers Add cleaning actions for different types of dirt to the necessary cleaning action set For example, when the global severity index of bird droppings exceeds the high mandatory threshold, foam soaking and manual finishing can be directly added to the necessary cleaning action set. When the global severity index of the oil film exceeds the high-force threshold, the glass degreasing action can be directly added to the necessary cleaning action set. .

[0176] S20307, ​​Perform forced trigger judgment for local critical areas.

[0177] For example, for safety-critical areas such as the windshield, rearview mirrors, headlights, and door handles, if the weighted dirt urgency value of a certain area of ​​the vehicle body is... Exceeding the local forced threshold Then it will be able to handle the first Types of dirt and the coverage coefficient of the action area Cleaning actions with a value greater than 0 are added to the required cleaning action set. This addresses the issue of "not being a serious problem globally but being critical in certain local areas." This represents the threshold for forced triggering in a local critical area, for example, it can be 0.60.

[0178] In this embodiment, as an example, there is a small oil film in the upper left area of ​​the windshield, and the global severity index of the oil film is... The high-pressure threshold was not reached, but the weighted dirt urgency value of the mesh in the windshield body area was [not specified]. Exceeding the local forced threshold Furthermore, the coverage coefficient of the glass degreasing film over the action area of ​​the mesh on the vehicle body is... Therefore, removing the grease film from glass is still included in the necessary cleaning process. .

[0179] S20308, Set of Necessary Cleaning Actions Perform prerequisite dependency supplementation.

[0180] Specifically, if a set of cleaning actions is necessary... If any cleaning action requires basic prerequisite actions, then the minimum number of basic prerequisite actions should be added according to the action prerequisite dependency rule table. For example, water wax coating requires the vehicle body surface to be cleaned and dried first. If water wax coating is added to the set of necessary cleaning actions... However, if pre-rinse and strong air drying are not selected, then pre-rinse and strong air drying will be added. Degreasing the glass requires a basic rinse of the windshield surface first. If degreasing the glass is added to the necessary cleaning action set... If pre-rinse is not selected, then perform a pre-rinse.

[0181] It should be noted that this invention generates the necessary cleaning action set by combining action activation threshold, high-level forced trigger judgment of dirt type, and forced trigger judgment of local key areas. Action activation threshold is suitable for handling situations where multiple moderate dirt types trigger actions together; high dirt type forced trigger judgment is suitable for handling situations where a single severe dirt must be removed; local critical area forced trigger judgment is suitable for handling situations where local dirt is found in safety-critical areas, which can reduce the risk of missed washing and accidental omission of critical actions.

[0182] like Figure 2As shown, the heatmap of the dirt urgency matrix after double-weighting is analyzed. The horizontal axis represents the grid index of the vehicle body area (unitless), and the vertical axis represents the dirt type index (unitless). The color intensity represents the weighted dirt urgency value, ranging from 0 to 1. Double-weighting effectively incorporates the importance of the area and the hazard of the dirt into the data, naturally highlighting safe and high-risk dirt in terms of values, while suppressing low-priority dirt, thus providing the correct direction for subsequent severity aggregation.

[0183] like Figure 3 As shown, a grouped bar chart is used to compare the local high value severity and overall distribution severity of each type of dirt. The horizontal axis represents the dirt type, and the vertical axis represents the severity value. The values ​​are dimensionless and range from 0 to 1. The dark red bars represent the average value of the top 5% of high-value areas, and the blue bars represent the average value of the entire vehicle. Figure 3 The high values ​​of localized oil film and bird droppings are significantly higher than the overall average, indicating that dirt is highly concentrated in a few grid areas, while dust shows little variation. Using only the overall vehicle average would mask localized high risks, highlighting the necessity of considering both localized high values ​​and overall distribution and weighting them together. This ensures that the global severity index can capture the severity of the dirtiest areas as well as reflect the overall situation of widespread dirt.

[0184] like Figure 4 As shown in this embodiment, a bar chart of the global severity vector of dirt is analyzed. The horizontal axis represents the type of dirt, and the vertical axis represents the global severity index, which is dimensionless and ranges from 0 to 1. Figure 4 The high global severity of bird droppings and oil film demonstrates that after weighted fusion of local high values ​​and overall distribution, the system can arrange various types of dirt according to their urgency, providing a direct basis for subsequent judgment on which actions must be performed. High severity types are more likely to trigger associated cleaning actions.

[0185] S3, Constructing the binary cleaning action decision vector The initial legal particle population is obtained. By evaluating fitness, adjusting time window urgency, and repairing sequence constraints, a list of executable cleaning action instructions is output.

[0186] After obtaining the necessary cleaning action set Afterwards, it is still necessary to select appropriate actions from the non-essential cleaning procedures to achieve a balance between cleaning effectiveness, order time window, remaining chemical agent, weather conditions, and vehicle characteristics. Relying entirely on manual experience can easily lead to retaining too many non-critical actions when time is tight, or excessively compressing the cleaning process when the soil is heavily soiled. Using random search directly can easily result in action combinations that do not conform to the construction sequence or vehicle limitations.

[0187] Constructing binary cleaning action decision vectors The initial legal particle population is obtained. The method includes the following steps: transforming the cleaning action selection problem into the form of a binary vector, generating an initial population by forcing the inclusion of necessary actions, calculating the selection probability of non-necessary actions, randomly sampling, and repairing the order constraints on each individual, thereby constructing a legal and diverse initial solution set.

[0188] S301, Construct the binary cleaning action decision vector , This represents a candidate combination of cleaning actions. Indicates the first The execution status of a cleaning action, when When it indicates the execution of the first Each cleaning action, when The time indicates that the first step will not be executed. Each cleaning action, Indicates the cleaning action library Total number of cleaning actions, binary cleaning action decision vector This only indicates whether the cleaning action is executed; it does not change the cleaning action library. The standard construction sequence.

[0189] In this embodiment, as an example, it is assumed that the cleaning action library has 8 actions. It is an 8-dimensional vector, such as, This indicates that actions 1, 2, 4, and 7 (such as pre-rinse, foam soaking, high-pressure rinsing, and strong air drying) will be performed, and other actions will not be performed.

[0190] S302, the necessary cleaning action set The corresponding action dimension is forcibly set to 1. Specifically, for each cleaning action... ,like This is a set of necessary cleaning actions. Then let And ensure that the dimension of this action cannot be set to 0 during subsequent initialization, particle update and pruning processes.

[0191] S303 is a set of cleaning actions that are not required. Calculate the initial selection probability for the cleaning action. , Indicates the first The probability of a non-essential cleaning action being selected from the initial candidate cleaning action combination.

[0192] In the specific implementation, first read the... Action activation score for each cleaning action Then, adjustments are made based on remaining resources and time window status to activate the score. The higher the initial selection probability The higher. If the first If the remaining amount of chemical agent required for a cleaning action is insufficient or the expected time is long, then the initial selection probability... Decrease. Initial selection probability. It can be limited to between 0.05 and 0.95 to avoid all particles being exactly the same.

[0193] In this embodiment, as an example, let's assume the action activation score for the action "water wax coating" is... Currently, there is sufficient water wax remaining, but the process is expected to be lengthy and the time window is tight, thus the initial selection probability is low. It may be lowered from the base value of 0.6 to 0.3.

[0194] S304, generates a particle population; the total number of particles is denoted as... , This represents the number of candidate cleaning action combinations participating in binary particle swarm optimization, for example... A range of 20 to 50 is acceptable. For the first... One particle, According to the initial selection probability corresponding to each action dimension Perform random sampling to obtain the first Decision vector of candidate binary cleaning actions for each particle .

[0195] Specifically, for the first Generate a random number between 0 and 1 for each action dimension. If the random number is less than or equal to Then let Otherwise ,in, Indicates the first The particle in the first The execution state of each action dimension.

[0196] It should be noted that, It is the first Particles In the The specific values ​​in each dimension, that is, It consists of all dimensions. The vector formed, for example, for the 3rd particle ( ), in the 5th action dimension ( Sampling is performed, and a random number of 0.5 is generated. If this dimension ,because Then let .

[0197] S305, perform order constraint repair on each candidate binary cleaning action decision vector to obtain a legal binary cleaning action decision vector.

[0198] Specifically, the sequential constraint repair follows the cleaning action library. The standard construction sequence is checked from front to back. If a selected cleaning action lacks necessary prerequisite actions, the missing basic prerequisite actions are supplemented according to the action prerequisite dependency rule table. If the supplemented actions conflict with vehicle characteristic constraints, the set of necessary cleaning actions is retained first. Remove unnecessary and conflicting actions. If two mutually exclusive cleaning actions are selected simultaneously, the necessary cleaning action is retained. If both mutually exclusive cleaning actions are unnecessary, the cleaning action with higher expected cleaning effect or shorter expected time is retained.

[0199] In this embodiment, as an example, suppose a candidate vector selects both "strong air drying" and "high pressure flushing", but these two actions are set to be mutually exclusive in the rule table (e.g., they cannot be performed simultaneously in some cases). If "high pressure flushing" is a necessary action to deal with the current severe mud spots, while "strong air drying" is not a necessary action, then the repair operation will remove "strong air drying" and retain "high pressure flushing".

[0200] S306, After completing the sequence constraint repair, the necessary cleanup action set will be re-selected. The corresponding action dimension is forcibly set to 1 to obtain the initial legal particle population. , This represents the set of decision vectors for the 0th generation of legal binary cleaning actions. The current position of each particle is taken as its individual optimal position, starting from the initial legal particle population. The particle with the highest fitness value is selected as the global optimal position.

[0201] In this embodiment, as an example, the cleaning action library... The process includes, in sequence, pre-rinsing, foam soaking, high-pressure rinsing, degreasing of the glass, wheel cleaning, water wax coating, high-pressure drying, and manual touch-up. If the randomly sampled candidate binary cleaning action decision vector selects high-pressure drying but not any rinsing action, the sequence constraint repair will check the pre-dependencies of high-pressure drying and supplement with either pre-rinsing or high-pressure rinsing. If the current vehicle has special paint and close-range high-pressure rinsing is prohibited, pre-rinsing will be prioritized over high-pressure rinsing. Based on this, unexecutable solutions such as "drying before cleaning" or "violating paint restrictions" can be avoided.

[0202] For each binary cleaning action decision vector x, the cleaning effect score is arithmetically calculated by introducing the diminishing marginal effect, and its expected time consumption, resource consumption cost, and time window penalty are comprehensively calculated to form a comprehensive score, namely the fitness value F(x), which is used to evaluate the merits of the scheme. The fitness evaluation method includes the following steps:

[0203] S30201, Read the cleaning action library Dirt removal rate in the process and action area coverage coefficient .

[0204] Among them, the dirt removal rate of the action Indicates when the first When a cleaning action completely covers a certain area, the first cleaning action within that area... The ideal removal ratio of different types of dirt characterizes the cleaning action's ability to remove different types of dirt, and the coverage coefficient of the action area. Indicates the first The first cleaning action is for the first The physical coverage of a vehicle body area grid represents whether a cleaning action can cover the corresponding vehicle body area grid. Together, they determine the actual cleaning contribution of a certain cleaning action to the removal of a certain type of dirt within a certain vehicle body area grid.

[0205] S30202, based on the binary cleaning action decision vector Calculate the cleaning effect score , This indicates the expected cleaning value of the current cleaning action combination for removing dirt from the vehicle; the higher the value, the better the cleaning effect.

[0206] In practical implementation, for each vehicle body area grid and each type of dirt, the overall removal rate of the selected cleaning action is calculated. If multiple cleaning actions can handle the same type of dirt, they are not directly added together, but calculated using a marginal decrease method to avoid overestimating the cleaning effect due to repeated actions. The cleaning effect score is calculated as follows:

[0207] ;

[0208] in, This indicates the estimated cleaning value of the current combination of cleaning actions for removing dirt from the vehicle. This indicates the number of grid regions that the current vehicle surface is divided into. Indicates the first Index of each vehicle body area grid, This indicates the total number of categories of dirt / soil. Indicates a dirty type index. Indicates the first The grid of the vehicle body area in the first... Weighted dirt urgency values ​​for different dirt types Indicates all Perform the corresponding operation for the cleaning action. Indicates all Multiply the items corresponding to the cleaning actions together to determine the dirt removal rate of each action. Indicates when the first When the first cleaning action completely covers a certain area, the first cleaning action within that area... The ideal removal ratio of various types of dirt and grime, and the coverage coefficient of the action area. Indicates the first The first cleaning action is for the first The physical coverage of the grid in each vehicle body area.

[0209] In this embodiment, as an example, if there are mud spots in a certain area of ​​the vehicle body grid, the cleaning rate of high-pressure washing is 0.85, and the cleaning rate of foam soaking is 0.50. Since both cleaning actions cover the vehicle body grid area, the combined cleaning rate is... Instead This calculation method characterizes the pattern of cleaning effect gradually approaching saturation and can suppress meaningless repeated cleaning actions.

[0210] S30203, based on the binary cleaning action decision vector Calculate the estimated total time , This indicates the estimated execution time of the current cleaning action combination under the current vehicle, current weather, and current store workstation conditions.

[0211] In practice, for each selected cleaning action, the standard time is first read, then multiplied by a vehicle size correction factor and a weather correction factor. The vehicle size correction factor reflects the surface area differences between small cars, mid-size SUVs, and large MPVs. The weather correction factor reflects the impact of humidity, rainfall, and wind speed on the drying action. For example, under high humidity or rainfall conditions, the weather correction factor for drying in strong winds can be taken as 1.2 to 1.5. Under low humidity and natural wind conditions, the weather correction factor for drying in strong winds can be taken as 0.8 to 0.9.

[0212] In this embodiment, as an example, it is assumed that a "high-pressure wash" is performed on a mid-size SUV, with a standard time of 4 minutes. The size correction factor for a mid-size SUV is 1.2, and due to the current high humidity weather, the weather correction factor is 1.3. Therefore, the estimated time for this action is... minute.

[0213] S30204, based on the binary cleaning action decision vector Computational resource consumption cost , This represents the combined cost of water, electricity, chemicals, and human resources expected to be consumed in the current cleaning operation combination.

[0214] In practice, the standard water and electricity costs, chemical costs, and labor costs are accumulated for each selected cleaning action. If the current remaining proportion of a certain type of chemical is lower than the low remaining chemical threshold, the chemical cost corresponding to the cleaning action using that chemical is increased, thus increasing the resource consumption cost. The calculation method is expressed as follows:

[0215]

[0216] in, This represents the combined cost of water, electricity, chemicals, and labor resources expected to be consumed by the current cleaning operation combination. Indicates all Perform the corresponding operation for the cleaning action. It is the dynamic cost coefficient of the chemical agent. When the current residual proportion of the chemical agent is lower than the preset low residual threshold of the chemical agent, (For example, take 2), otherwise take 1. Indicates the first The standard water cost per cleaning action, expressed in resource consumption cents. Indicates the first The standard electricity cost of a single cleaning operation Indicates the first The standard consumption cost of a particular chemical agent for each cleaning action. Indicates the first The standard labor cost for each cleaning action.

[0217] In this embodiment, as an example, the action of "foam soaking" consumes a certain cleaning agent, and there is a sufficient amount of the cleaning agent remaining. The standard chemical cost is 2 yuan. The water-wax coating process consumes less than 20% of the required amount of water-wax. The standard chemical reagent cost is 3 yuan, so the actual cost included is... Yuan.

[0218] S30205, Calculate the time window penalty value based on the order constraint data table. , This indicates the penalty incurred when the current cleaning action combination exceeds the order time window.

[0219] In practical implementation, the maximum allowed time for the order time window is denoted as: , This indicates the maximum allowed time from the start of cleaning to delivery for the current order. If Then let .like First, calculate the timeout duration. Then, the time window penalty value increases rapidly with the timeout duration; for example, it can be set to... .in, This indicates the duration of the current cleaning action combination relative to the order time window.

[0220] In this embodiment, as an example, the maximum allowed time for the order time window is... For 25 minutes, if the decision vector Estimated total time If it is 30 minutes, then the overtime duration is... Minutes, Time Window Penalty Value .

[0221] S30206, Calculate the time penalty weight based on the remaining time window. , This indicates the strength of the time window penalty value in fitness evaluation; the shorter the remaining time window, the greater the weight of the time penalty. The larger.

[0222] In practical implementation, the remaining time window is denoted as... , This represents the available time between the current moment and the order's required delivery time, after deducting queuing time and workstation switching. When there is sufficient time, the time penalty weighting Close to the base value. When Time penalty weight when the cleaning time is close to or less than the normal cleaning time The increased size makes the optimization process more inclined to choose solutions with shorter processes but retaining key cleaning actions.

[0223] In one implementation, time penalty weights Based on the remaining time window Compared to regular cleaning time The ratio is calculated in segments, where, This represents the average time required for a store to complete a routine car wash under standard operating conditions, and is a preset constant. When Greater than or equal to season , This represents the base value of the time penalty weight, for example... A value of 0.3 to 0.5 is acceptable. When Less than Time penalty weight It increases linearly as the remaining time window decreases, and the calculation method is expressed as follows:

[0224]

[0225] in, This indicates the strength of the impact of the time window penalty value on fitness evaluation. This represents the upper limit of the time penalty weight, used to significantly increase the penalty intensity when time is extremely tight, for example... Version 2.0 is acceptable. This indicates the available time between the current moment and the order's required delivery time, after deducting queuing time and workstation switching.

[0226] In this embodiment, as an example, let the normal cleaning time be... For 30 minutes, Take 0.4, Take 2.0; if the current order has a remaining time window For 40 minutes (or more) ),but ;like 18 minutes (less than) ),but Based on this, the closer an order is to its timeout, the greater the weight of the time penalty, and the more likely it is to generate a cleaning solution that takes less time.

[0227] S30207, based on cleaning effectiveness score Resource consumption costs and time window penalty value Obtain fitness value , Represents the binary cleaning action decision vector Considering all advantages and disadvantages, the optimization objective is to maximize the fitness value. The calculation method is expressed as follows:

[0228]

[0229] in, This indicates the estimated cleaning value of the current combination of cleaning actions for removing dirt from the vehicle. Indicates the weight of resource consumption costs, for example 0.1 is acceptable. Indicates the time penalty weight, for example 0.5 is acceptable. This represents the combined cost of water, electricity, chemicals, and labor resources expected to be consumed by the current cleaning operation combination. This indicates the penalty incurred when the current cleaning action combination exceeds the order time window.

[0230] It's important to note that fitness evaluation isn't simply about maximizing the number of cleaning actions, but rather about striking a balance between cleaning effectiveness, resource consumption, and order time windows. A diminishing marginal score for cleaning effectiveness can reduce repetitive cleaning actions, dynamic cost of remaining chemical agents can prevent overuse of scarce cleaning agents when resources are limited, and time window penalties can prevent solutions from failing to be delivered on time.

[0231] The binary particle swarm optimization algorithm is used to iteratively optimize candidate cleaning action combinations. During the iteration process, parameters such as inertia weight are dynamically adjusted according to the remaining time window to balance the breadth of the search and the convergence speed, ultimately approximating the global optimum. The methods for adjusting the time window urgency and repairing the order constraints include the following steps:

[0232] S30301, Set the maximum number of iterations , This represents the maximum number of update rounds in binary particle swarm optimization, for example... A value of 30 to 80 can be taken, in the... The time window urgency factor is calculated at the start of the next iteration. , Indicates the first The urgency level of the remaining time for the order at the next iteration, ranging from 0 to 1. Time window urgency coefficient. The closer the value is to 1, the tighter the order timeline; the more urgent the time window. The closer it is to 0, the more ample the order time.

[0233] In practical implementation, if the remaining time window for an order is short, the maximum number of iterations can be reduced. This allows the system to quickly output executable results. If the remaining time window for the order is ample, the maximum number of iterations can be increased. This allows the system to have more opportunities for searching.

[0234] Based on the remaining time window The ratio of the remaining time window to the store's regular cleaning time indicates a higher urgency level. The larger the value, the more it is calculated as follows:

[0235]

[0236] in, Indicates the first The urgency of the remaining time for the order in the next iteration. This means limiting the calculation result to the range of 0 to 1. This indicates the average time required for a store to complete a routine car wash under standard operating conditions. This indicates the available time between the current moment and the order's required delivery time, after deducting queuing time and workstation switching.

[0237] S30302, based on the time window urgency coefficient Adjusting inertia weight , Indicates the first The inertia weight of the nth iteration, characterizing the... The degree to which historical search directions are retained in the particle velocity update during each iteration.

[0238] In one implementation, the time window urgency coefficient At lower levels, inertia weight A larger value can be chosen to preserve exploration capabilities. Time window urgency coefficient. At higher levels, inertia weight A smaller value can be chosen to speed up convergence. Inertia weight. Based on the time window urgency factor Linear adjustment is performed, and the calculation method is expressed as follows:

[0239]

[0240] in, This represents the maximum inertia weight, used to maintain strong global exploration capabilities when time is ample, for example... 0.85 is acceptable. This represents the minimum inertia weight, used to accelerate convergence when time is limited, for example... 0.45 is a good value.

[0241] It should be noted that when the time window urgency coefficient At lower (with ample time) inertia weight near To preserve the ability to explore. When the time window urgency factor When the inertia weight is high (time is tight), near This is to accelerate the convergence speed.

[0242] S30303, regarding the first The first particle Update speed of each action dimension , Indicates the first During the nth iteration The particle in the first The velocity value in each action dimension is used to determine the tendency of that action dimension to take a value of 1 in the next sampling.

[0243] In the actual implementation, the speed update simultaneously considers the current speed, the individual's optimal position, and the global optimal position. The individual's optimal position represents the position of the first... The global optimal position represents the binary cleanup action decision vector with the highest fitness value in the history of all particles. After velocity update, Limit it to a preset range, such as between -6 and 6, to avoid the probability of subsequent action selections being too close to 0 or 1 and thus losing the search capability.

[0244] In one implementation, the first The particle in the first Speed ​​in each action dimension The update calculation method is expressed as follows:

[0245]

[0246] in, Indicates the first During the nth iteration The particle in the first Speed ​​values ​​in each action dimension No. During the nth iteration The particle in the first Speed ​​values ​​in each action dimension Indicates the first The inertia weight of the next iteration, The individual learning factor represents the weight by which a particle learns from its own historical best position; for example, it can be 1.5. This represents a random number between 0 and 1, used to increase the randomness of the search. Indicates the first The best historical position of each particle is at the _ The value (0 or 1) is taken in each action dimension. Indicates the first During the nth iteration The particle in the first The position value (0 or 1) is taken in each action dimension. The social learning factor represents the weights a particle learns to achieve the global optimal position, for example... 1.5 is acceptable. Represents a random number between 0 and 1. Indicates the globally optimal position at the th position. The value (0 or 1) is taken in the action dimension.

[0247] In one embodiment, as an example, suppose , , In the third action dimension, the velocity of a certain particle in the previous round Previous position its own best historical position Global optimal position Random numbers , Then the speed of this round. If the speed limit is set to 6, then Within the permissible range.

[0248] S30304, speed value Converted into action execution probability, the action execution probability represents the first action. The particle in the first The probability of taking a value of 1 in each action dimension.

[0249] In one implementation, an S-shaped function can be used for the transformation; the higher the speed value, the higher the probability of the action being executed; the lower the speed value, the lower the probability of the action being executed. This applies to sets of necessary cleaning actions. For the action dimension, no random sampling is performed; the action execution state is directly kept at 1. Specifically, the velocity value... Convert to action execution probability Using an S-shaped function, the calculation method is expressed as follows:

[0250]

[0251] in, Indicates the first During the nth iteration The particle in the first The probability of a value of 1 in each action dimension, for a set of necessary cleaning actions. Action Dimensions Without random sampling, directly let Furthermore, this dimension does not participate in speed updates. Indicates the first During the nth iteration The particle in the first The speed value in each motion dimension.

[0252] In this embodiment, as an example, if the velocity of a particle in the 5th action dimension... Then the probability of action execution If the action dimension is a non-essential cleaning action, then generate a random number between 0 and 1. If the random number is less than or equal to 0.887, then let... Otherwise .

[0253] S30305, Randomly sample non-essential cleaning actions based on the action execution probability to obtain a temporary binary cleaning action decision vector. Perform order constraint repair on the temporary binary cleaning action decision vector, repairing action pre-dependencies, action mutual exclusion restrictions, vehicle feature restrictions, and resource unavailability restrictions, and once again enforce the maintenance of the essential cleaning action set. The corresponding action dimension is 1.

[0254] In practical implementation, the temporary binary cleaning action decision vector After obtaining the data through dimensional sampling, sequential constraint repair is performed. The repair process consists of three steps: First, according to the cleanup action library... The standard construction sequence is traversed from front to back. If a selected action lacks a necessary prerequisite action, that prerequisite action is added. The second step checks if there are mutually exclusive action pairs selected simultaneously; if so, one unnecessary action is deleted according to the rules. The third step checks if the repaired vector violates vehicle feature restrictions or resource unavailability restrictions; if so, the corresponding unnecessary action is set to 0. After repair, all actions belonging to... The action dimension is forcibly set to 1 to obtain the legal binary cleaning action decision vector. .

[0255] In one embodiment, as an example, suppose Includes "high-pressure flushing", after the current particle is sampled. The "water wax coating" dimension is 1, but the "pre-rinse" dimension is 0. Sequence constraint repair detects that "water wax coating" depends on "pre-rinse," so the "pre-rinse" dimension is set to 1. If the current vehicle paint type is a special paint and high-pressure washing is prohibited, but "high-pressure washing" is... If "high-pressure washing" is also 1 and conflicts with special paint restrictions, and "strong washing" does not belong to... If so, set "Strong Scrub" to 0. After the repair is complete, ensure that the "High Pressure Wash" dimension is 1 again.

[0256] S30306, Calculate the fitness value corresponding to the repaired binary cleaning action decision vector. .

[0257] In practice, if the fitness value of the current particle is higher than the individual best fitness value of the particle, then the individual best position of the particle is updated; if the fitness value of the current particle is higher than the global best fitness value, then the global best position is updated.

[0258] Specifically, in the In the nth iteration, for the th Each particle calculates its decision vector for the legal binary cleaning action after repair. Corresponding fitness value .like Then Updated to ,in Indicates the first The optimal position of each individual particle. If Then Updated to ,in This indicates the globally optimal position.

[0259] In this embodiment, as an example, let the current fitness value of the third particle be... The individual's best fitness value in history Then let If the current global optimal fitness value ,because Then let .

[0260] S30307 involves repeatedly performing speed updates, action execution probability transformations, random sampling, sequence constraint repair, and fitness evaluation until the maximum number of iterations is reached. Or, the global optimal fitness value improves below a preset improvement threshold after several consecutive iterations. The preset improvement threshold can be set to... The number of consecutive stops can be set to 5 to 10, thereby reducing invalid calculations in time-sensitive orders.

[0261] It should be noted that the time window urgency factor The system directly participates in the binary particle swarm optimization process, enabling it to automatically adjust its search strategy based on the remaining time for each order. When time is ample, the system retains more exploration capabilities, tending to seek combinations of cleaning actions with higher efficiency. When time is tight, the system accelerates convergence, prioritizing the output of executable solutions that meet the necessary cleaning actions and delivery time windows.

[0262] The globally optimal solution obtained through optimization undergoes a final compliance check. If the solution takes longer than the order time window, non-essential actions are greedily pruned until the time requirement is met or only essential actions remain. Finally, a structured list of cleaning action instructions is output. The method for outputting the list of executable cleaning action instructions includes the following steps:

[0263] S30401, after the iteration, the globally optimal cleaning action decision vector is obtained. , This represents the binary cleaning action decision vector with the highest overall evaluation obtained during the binary particle swarm optimization process.

[0264] In practical implementation, the globally optimal cleaning action decision vector When the binary particle swarm optimization process ends, The corresponding binary cleaning action decision vector. After the optimization process is complete, directly take... That is, the decision vector of the legal binary cleaning action with the highest fitness value in all iteration rounds.

[0265] In this embodiment, as an example, assuming the cleaning action library contains 8 actions, the globally optimal position is determined at the end of the 50th iteration. ,but This indicates that actions 1, 2, 4, and 7 (such as pre-rinse, foam soaking, high-pressure rinsing, and strong air drying) will be performed, and other actions will not be performed.

[0266] Global optimal cleaning action decision vector Perform a final feasibility check, including the necessary cleaning action set. Whether all actions are executed, whether pre-action dependencies are satisfied, whether action mutual exclusion restrictions are satisfied, whether vehicle characteristic restrictions are satisfied, whether store resources are sufficient, and the estimated total time. Does it not exceed the maximum allowed time within the order time window? .

[0267] S30402, if the globally optimal cleaning action decision vector Corresponding estimated total time Less than or equal to the maximum allowable time for the order time window Then directly follow the cleaning action library. The standard construction sequence converts cleaning actions with a value of 1 into a list of cleaning action instructions.

[0268] It should be noted that the list of cleaning action instructions is based on the cleaning action library. The standard construction sequence is to extract in sequence. The cleaning action information corresponding to the action dimension with a value of 1 is used to generate structured instruction entries; each instruction includes action name, execution sequence number, estimated time, main type of dirt to be treated, corresponding vehicle body functional area, whether it is a necessary cleaning action, recommendation reason and precautions.

[0269] In one embodiment, as an example, suppose If the actions with a value of 1 are pre-rinse (index 1) and high-pressure rinse (index 3), then the cleaning action instruction list output is:

[0270] Step 1: Pre-rinse, estimated 2 minutes, to remove dust and mud, covering the entire exterior surface of the vehicle. This is a necessary cleaning action (preliminary step). It is recommended as a basic preliminary step for subsequent cleaning actions.

[0271] Step 2: High-pressure rinsing, estimated 4 minutes, to remove mud and oil stains from the wheel hubs, covering the sides of the vehicle and the wheel hubs. This is a necessary cleaning action (action activation score exceeds the threshold). The reason for the recommendation is that the global severity index of mud spots exceeds the high mandatory threshold, and high-pressure rinsing is required to remove them.

[0272] If the globally optimal cleaning action decision vector Corresponding estimated total time Greater than the maximum allowed time for the order time window If not, then a greedy timeout pruning is performed on non-essential cleaning actions. Greedy timeout pruning only allows the removal of actions that are not part of the essential cleaning action set. The cleaning process cannot remove the necessary cleaning action set. The cleaning action in the process.

[0273] In this embodiment, as an example, let's assume... Corresponding estimated total time Minutes, maximum allowed time for the order time window Minutes, timeout required for cropping. Currently... The non-essential actions include "water wax coating" (index 6) and "artificial finishing" (index 8). Calculate the marginal effect loss rate for both actions, assuming that "artificial finishing" has a marginal effect loss rate of 100%. "Water wax coating" .according to Sort by size from smallest to largest, first remove "water wax coating", then recalculate. minutes, still greater than Then remove "artificial supplementation". Reduce the time to 25 minutes to meet the time window requirement, then the cropping is complete.

[0274] S30403, Calculate the marginal effect loss rate for each removable cleaning action. , Indicates the removal of the first The loss of cleaning effect caused by time saving per unit after each cleaning action; the smaller the value, the better. The more suitable the cleaning action is, the more likely it is to be removed.

[0275] In the specific implementation, first calculate the removal of the first... Difference in cleaning effectiveness score before and after each cleaning action , Indicates the removal of the first The decrease in cleaning effect score caused by each cleaning action, then read the... The current estimated time for each cleaning action. , This indicates the number after considering weather and vehicle model adjustments. Each cleaning action takes time; then let... ,in, This represents a very small constant, used to avoid a denominator of 0; for example, it can be taken as... .

[0276] S30404, according to the marginal utility loss rate Remove unnecessary cleaning actions in ascending order of their complexity. After removing each unnecessary cleaning action, recalculate the estimated total time. and cleaning effect score until the estimated total time Less than or equal to the maximum allowable time for the order time window Or there are no unnecessary cleaning actions that can be removed.

[0277] In one embodiment, as an example, if the globally optimal cleaning action decision vector The cleaning process includes pre-rinse, foam soaking, high-pressure rinsing, wheel cleaning, water wax coating, strong air drying, and manual touch-up. However, if the order time window is insufficient, the system first calculates the marginal effect loss rate of non-essential cleaning actions such as water wax coating and manual touch-up. If the current vehicle's main dirt consists of mud spots and wheel oil stains, and water wax coating contributes little to the current dirt removal and takes a long time, then the marginal effect loss rate of water wax coating is low, and the system prioritizes removing water wax coating. If removing water wax coating still results in a timeout, then removing manual touch-up, which contributes less to the current dirt removal, is considered. If high-pressure rinsing is a necessary cleaning action... If so, high-pressure flushing will not remove it.

[0278] S30405, Even after all non-essential cleaning actions are removed, the estimated total time is still greater than the maximum allowable time within the order time window. If the minimum required cleaning plan is not met, the time window will be output and a warning message will be marked in the cleaning action instruction list. The time window shortage warning message may include "The current order time window is insufficient. The required cleaning actions for safety-critical areas and high-hazard dirt have been reserved. Please manually confirm whether to extend the delivery time or adjust the workstation arrangement."

[0279] S30406, Output a list of cleaning action instructions: The list includes the name of the cleaning action, execution order, estimated time, main type of dirt to be treated, corresponding functional area of ​​the vehicle body, whether it is a necessary cleaning action, the reason for recommendation, and precautions. The list of cleaning action instructions is based on the cleaning action library. The standard construction sequence is arranged and directly sent to the car wash workers' mobile phones, workstation displays, or automatic car wash control systems.

[0280] In one embodiment, as an example, the list of cleaning action instructions can be output as follows: Pre-rinse, estimated 2 minutes, to remove dust and mud, a prerequisite basic action; High-pressure rinse, estimated 4 minutes, to remove mud, a necessary cleaning action because the global severity index of mud exceeds the high-force threshold; Wheel hub cleaning, estimated 3 minutes, to remove wheel hub oil stains, a necessary cleaning action because the local weighted dirt urgency value of the wheel hub area exceeds the local forced threshold; Strong air drying, estimated 3 minutes, a delivery drying requirement action; Abandon water wax coating because the current order time window is tight and water wax coating contributes little to the current main dirt removal.

[0281] like Figure 5As shown, the analysis compares the type activation value, low alert threshold, and high enforcement threshold for each type of dirt in the same coordinate system. The horizontal axis represents the dirt type, and the vertical axis represents the activation value or threshold, which is dimensionless. The green bars represent the activation value, the blue line represents the low alert threshold, and the red line represents the high enforcement threshold. When the activation value exceeds the high enforcement threshold, the corresponding dirt will be forcibly cleaned; when it is between the two thresholds, a linear mapping is used.

[0282] like Figures 6 to 9 As shown, the diagram displays the candidate binary cleaning action decision vectors for four particles in a given iteration step. The horizontal axis of each subgraph represents the execution state, with values ​​of 0 or 1 (0 for no execution, 1 for execution), and is dimensionless. The vertical axis represents the name of the cleaning action. A gold semi-transparent horizontal bar marks the action dimension belonging to the required action set, which is forcibly kept at 1. Different particles choose different non-required actions, reflecting population diversity.

[0283] S4 executes vehicle cleaning strategies and updates feedback online.

[0284] After generating the cleaning action instruction list, this invention can be deployed in car wash shop workstation systems, mobile car wash operation systems, or automatic car wash control systems. The system automatically generates an initial dirt level matrix after each vehicle enters the detection area, calculates the necessary cleaning action set, optimizes the cleaning action instruction list, and updates the cleaning action library parameters based on execution feedback.

[0285] S401, the specific steps for generating an online cleaning strategy are as follows:

[0286] S40101, After the vehicle enters the detection area, the image acquisition device is triggered to acquire multi-view images of the vehicle. Simultaneously, it reads the vehicle characteristic data table, order constraint data table, store resource data table, and weather environment data table. If the user order specifies a service level, such as a regular quick wash, a fine wash, or a maintenance wash, the service level is written into the order constraint data table to influence the range of optional washing actions and time window constraints.

[0287] S40102, Generate the original dirt level matrix based on S10104. A weighted dirt urgency matrix is ​​generated based on S201. Generate a global severity vector of contamination based on S202. Generate the necessary cleaning action set according to S203. If the vehicle image recognition model identifies an area with insufficient image quality, such as severe glare from the windshield causing unreliable oil film recognition, then that area of ​​the vehicle body will be marked as a manually reviewed area, and a manual inspection prompt will be added to the cleaning action instruction list.

[0288] S40103: Based on S3, a cleaning action instruction list is generated and distributed to the car wash worker's mobile phone, workstation display screen, or automatic car wash control system. The cleaning action instruction list not only provides the action name but also a recommended reason, enabling the car wash worker to understand why a certain cleaning action should be performed or skipped. For example, when removing oil film from the windshield is listed as a mandatory cleaning action, the recommended reason could be displayed as "The oil film on the windshield is located in a safety-critical area, and the local weighted dirt urgency value exceeds the local mandatory threshold."

[0289] S40104: During the cleaning process, if the car wash worker finds a significant discrepancy between the actual dirt condition and the vehicle image recognition model result, they can manually correct the issue on their mobile device. Manual corrections include adding dirt types, reducing dirt levels, forcibly adding cleaning actions, or forcibly disabling cleaning actions. After manual correction, the system re-executes S2 to S3 to generate an updated list of cleaning action instructions, preventing incorrect recognition results from directly affecting the final cleaning process.

[0290] S402, Execution result feedback and parameter update:

[0291] After a single cleaning task is completed, the system collects actual execution data and post-cleaning results to perform closed-loop feedback correction on standard process parameters and dirt hazard indices in the cleaning action library, so that the model parameters gradually approximate real working conditions. The specific steps are as follows:

[0292] S40201, After cleaning is completed, collect execution result data. This data includes the actual time consumed for each cleaning action, the actual chemical consumption, whether rework occurred, user satisfaction rating, and post-cleaning inspection images. The post-cleaning inspection images are used to determine whether the main types of dirt have been effectively removed. The actual time and actual resource consumption are used to correct the cleaning action library. The standard process parameters in the process.

[0293] S40202, calculate the degree of residual dirt based on the re-inspection image after cleaning.

[0294] Specifically, if a certain type of dirt frequently remains in a specific functional area of ​​the vehicle body, it indicates that the corresponding cleaning action's actual removal capability for that type of dirt or that functional area is lower than the preset value. The system can reduce the dirt removal rate of the corresponding action. or action area coverage coefficient This makes subsequent cleaning strategies more inclined to increase supplementary actions or manual finishing.

[0295] S40203, update the standard time based on the actual time consumed.

[0296] Specifically, for the first If the actual time taken for the first cleaning action exceeds the standard time in multiple consecutive orders, then the time for the second cleaning action will be increased. The standard time for each cleaning action. If the actual time for multiple consecutive orders is less than the standard time, then the time for the first cleaning action will be reduced. The standard time for each cleaning operation. A moving average method can be used during updates to avoid significant parameter fluctuations caused by a single abnormal operation.

[0297] S40204, Update the dirt and hazard index based on rework records. and action activation threshold .

[0298] Specifically, if a certain type of dirt frequently leads to rework or user complaints, the dirt hazard index for that type of dirt will be increased. This assigns a higher weighted dirt urgency value to subsequent cleaning actions of the same type. If a cleaning action is frequently manually performed, the action activation threshold for that cleaning action is lowered. This makes it easier for the system to automatically add the cleaning action to the set of necessary cleaning actions. Or in the optimization results.

[0299] It should be noted that the feedback update does not change the list of cleaning action instructions already generated for a single cleaning task. Instead, it gradually corrects the cleaning action library parameters, dirt hazard index, and action activation threshold in subsequent orders. Based on this, the system can continuously improve the accuracy and feasibility of the dynamic cleaning strategy by combining the actual construction results in the store.

[0300] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A data-driven, convenient dynamic car wash cleaning strategy optimization method, characterized in that: Includes the following steps: S1 generates vehicle dirt detection data and establishes a cleaning action library. With cleaning rules table; S2 performs regional risk weighting based on vehicle dirt detection data to generate a set of necessary cleaning actions. ; S201, regarding the original dirtiness matrix By applying a two-level weighting of regional safety function weights and the contamination hazard index, a weighted contamination urgency matrix is ​​obtained. , S202, from the weighted dirt urgency matrix The urgency features of each type of dirt are extracted in local high-urgency areas and across the entire vehicle, and the two are weighted and fused to generate a global dirt severity vector that comprehensively reflects the global urgency of dirt treatment for that type of dirt. ; S203, generates the necessary cleaning action set by jointly determining the action activation threshold, the high-level forced trigger judgment of dirt type, and the forced trigger judgment of local key areas. ; S3, Construct the binary cleaning action decision vector The initial legal particle population is obtained. Through fitness evaluation, time window urgency adjustment and sequence constraint repair, a list of executable cleaning action instructions is output. S4, Online Execution and Feedback Update of Vehicle Cleaning Strategy: Automatically generates an original dirt level matrix upon entering the detection area. The necessary set of cleaning actions is calculated, the list of cleaning action instructions is optimized, and the parameters of the cleaning action library are updated based on the execution feedback.

2. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, The method for generating vehicle dirt detection data in S1 includes the following steps: S10101, Acquire multi-view images of the vehicle. The image set includes images of the front of the vehicle, the rear of the vehicle, the left side of the vehicle, the right side of the vehicle, the roof of the vehicle, partial images of the wheel hubs, and partial images of the windshield. S10102, vehicle multi-view images Input the vehicle image recognition model, which is used to identify the vehicle's outer contour, body functional areas, and dirty areas. The vehicle image recognition model outputs the body functional area category and dirt type category for each pixel. The body functional area categories include the windshield, rear windshield, headlights, taillights, rearview mirrors, door handles, body sides, roof, wheel rims, lower skirts, and bumpers. The dirt type categories include dust, mud spots, bird droppings, oil film, tree sap, water stains, wheel rim oil stains, and crevices dirt. S10103 divides the vehicle surface into multiple body area grids and establishes the correspondence between the body area grids and the body functional areas; S10104, Based on the output of the vehicle image recognition model, generate the original dirt level matrix. ; S10105, Obtain vehicle characteristic data table, order constraint data table, store resource data table and weather environment data table; The vehicle characteristic data table includes vehicle size, paint type, whether it is tinted, vehicle age, whether it is allowed to be heavily scrubbed, and whether it is allowed to be washed at close range with high pressure. The order constraint data table includes the remaining available time of the order, the service level selected by the user, whether dry delivery of the vehicle body is required, and whether delayed delivery is allowed. The store resource data table includes the current availability status of workstations, remaining cleaning agent, water and electricity resource status, and available manpower. The aforementioned weather and environmental data table includes temperature, humidity, wind speed, and rainfall status.

3. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, The cleaning action library is established in S1. The method for using the cleaning rule table includes the following steps: S10201, Establish a cleaning action library , This represents the complete set of cleaning actions available. Indicates the first Each cleaning action, Indicates the cleaning action index. , Indicates the total number of cleaning actions; S10202, for cleaning the action library Each cleaning action is configured with standard process parameters, including standard time, standard water and electricity costs, chemical type, standard chemical consumption, applicable vehicle body functional areas, main types of dirt to be treated, and precautions. S10203, configures the dirt removal rate for each cleaning action. ; S10204, Configure the action area coverage coefficient for each cleaning action. ; S10205, Establish an action pre-dependency rule table and an action mutual exclusion restriction rule table. The action pre-dependency rule table is used to record the basic cleaning actions that must be executed before a certain cleaning action is executed. The action mutual exclusion restriction rule table is used to record cleaning actions that should not be executed simultaneously or are prohibited from being executed under specific vehicle conditions.

4. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, The original dirt level matrix in S201 By applying a two-layer weighting of regional safety function weights and the contamination hazard index, a weighted contamination urgency matrix is ​​obtained. The method includes the following steps: S20101, Read the original dirt level matrix And confirm the original dirt level matrix. The size is ; S20102, Configure regional safety function weights for each vehicle body area grid. , Indicates the first The importance of individual body area grids to driver visibility, driving safety, vehicle operation, and user perception; S20103, configures a dirt hazard index for each type of dirt. , Indicates the first The intensity of the impact of different types of dirt on driving visibility, paint protection, cleaning difficulty, corrosion risk, and user complaint risk; S20104, regarding the original dirtiness matrix Perform a two-level weighting to obtain a weighted dirt urgency matrix. , This represents the weighted set of processing urgency for the current vehicle across all body area grids and all dirt types. Indicates the first The grid of the vehicle body area in the first... Weighted dirt urgency values ​​for different dirt types.

5. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, In S2, a global severity vector of contamination is generated that comprehensively reflects the urgency of global treatment for this type of contamination. The method includes the following steps: S20201, for the first Types of dirt, from the weighted dirt urgency matrix Extract the first Column, to get the first The urgency sequence of different types of pollution, the first The urgency sequence of different types of dirt includes Each weighted dirt urgency value corresponds to one vehicle body area grid. S20202, regarding the first Sort the areas with varying degrees of urgency based on their type of contamination from highest to lowest, and select the top... A weighted dirt urgency value was calculated, and the previous... The average of the weighted dirt urgency values ​​is used to obtain the local high severity value. ; S20203, calculate the first The overall distribution severity is obtained by averaging the dirt severity of each dirt type across all vehicle body grid areas. ,Right now , Indicates the first The average urgency of each type of dirt across the entire vehicle. Indicates the first The grid of the vehicle body area in the first... Weighted dirt urgency values ​​for different types of dirt; S20204, the severity of local high values and overall distribution severity Weighted fusion yields the first Global severity index for each type of dirt ; S20205, repeat S20201~S20204 to traverse all dirt types and obtain the global dirt severity vector. .

6. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, In step S203, the necessary cleaning action set is generated. The method includes the following steps: S20301, for cleaning the action library Each cleaning action in the configuration is associated with a set of dirt types. ; S20302, Configure a low alert threshold for each type of dirt. and high forced threshold ; S20303, based on the global severity vector of contamination Calculate the type activation value for each type of dirt. ; S20304, according to the... A set of associated dirt types for each cleaning action Calculate action activation score ; S20305, Configure an action activation threshold for each cleaning action. ; S20306, execute the high-forced trigger judgment of dirt type; S20307, ​​Perform forced trigger judgment for local critical areas; S20308, Set of Necessary Cleaning Actions Perform prerequisite dependency supplementation.

7. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, The binary cleaning action decision vector is constructed in S3. The initial legal particle population is obtained. The method includes the following steps: S30101, Constructing the binary cleaning action decision vector , This represents a candidate combination of cleaning actions. Indicates the first The execution status of each cleaning action. Indicates the cleaning action library The total number of cleaning actions in the process; S30102, set of necessary cleaning actions The corresponding action dimension is forcibly set to 1; S30103 is a set of cleaning actions that are not required. Calculate the initial selection probability for the cleaning action. ; S30104, generate a particle population, the total number of particles is denoted as... ; S30105, Perform order constraint repair on each candidate binary cleaning action decision vector to obtain a legal binary cleaning action decision vector; S30106, After completing the sequence constraint repair, the necessary cleanup action set will be re-selected. The corresponding action dimension is forcibly set to 1 to obtain the initial legal particle population. .

8. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, The fitness evaluation method in S3 includes the following steps: S30201, Read the cleaning action library Dirt removal rate in the process and action area coverage coefficient ; S30202, based on the binary cleaning action decision vector Calculate the cleaning effect score ; S30203, based on the binary cleaning action decision vector Calculate the estimated total time ; S30204, based on the binary cleaning action decision vector Computational resource consumption cost ; S30205, Calculate the time window penalty value based on the order constraint data table. ; S30206, Calculate the time penalty weight based on the remaining time window. ; S30207, based on cleaning effectiveness score Resource consumption costs and time window penalty value Obtain fitness value .

9. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, The method for adjusting the urgency of the time window and repairing the sequence constraints in S3 includes the following steps: S30301, Set the maximum number of iterations In the The time window urgency factor is calculated at the start of the next iteration. ; S30302, based on the time window urgency coefficient Adjusting inertia weight Spend; S30303, for the first The first particle Update speed of each action dimension ; S30304, speed value Converted into action execution probability, the action execution probability represents the first action. The particle in the first The probability of taking a value of 1 in each action dimension; S30305, Randomly sample non-essential cleaning actions based on the action execution probability to obtain a temporary binary cleaning action decision vector, and repair the execution order constraint of the temporary binary cleaning action decision vector; S30306, Calculate the fitness value corresponding to the repaired binary cleaning action decision vector. ; S30307 involves repeatedly performing speed updates, action execution probability transformations, random sampling, sequence constraint repair, and fitness evaluation until the maximum number of iterations is reached. .

10. The data-driven, convenient dynamic car wash cleaning strategy optimization method according to claim 1, characterized in that, The method for online execution and feedback update of the vehicle cleaning strategy in S4 includes the following steps: S40201 After cleaning is completed, collect execution result data, which includes the actual time consumed for each cleaning action, the actual amount of chemical agent consumed, whether rework occurred, user satisfaction score, and post-cleaning inspection images. S40202, calculate the degree of residual dirt based on the re-inspection image after cleaning; S40203, update the standard time based on the actual time consumed; S40204, Update the dirt and hazard index based on rework records. and action activation threshold .

Images