A deep learning driven intelligent judgment system for vehicle cleaning needs of unmanned car washing

Abstract

The application belongs to the technical field of unmanned intelligent control, and particularly relates to a deep learning driven unmanned vehicle cleaning demand intelligent judgment system, which comprises a multi-modal anti-interference data acquisition module, an edge computing processing module, a car paint-pollution dual-branch decoupling deep learning model library, a pollution-car paint coupling feature database, a multi-target Pareto optimal cleaning scheme generation module, a washing-judging closed loop execution and feedback module and a federal incremental learning updating module; by adopting the multi-modal anti-interference data acquisition module integrating a polarization high-definition imaging unit and a vehicle body micro-motion IMU synchronous unit, in combination with a polarization differential imaging algorithm and a multi-sensor space-time synchronous registration technology, the problems of car paint high mirror reflection, environmental light mutation, vehicle dynamic driving in vibration leading to collected data distortion and high pollution leakage and misdiagnosis rate in the prior art are solved.

Description

Technical Field

[0001] This invention belongs to the field of unmanned intelligent control technology, specifically a deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs. Background Technology

[0002] With the continuous increase in the number of civilian vehicles in my country, the demand for automotive aftermarket services is experiencing explosive growth in scale and diversification. Car wash services, as a high-frequency, essential service in the automotive aftermarket, are rapidly transforming from traditional manual car washes to unmanned, automated, and intelligent models. Furthermore, the upgrading demands of end-users for more refined, personalized, and safe car wash services are placing higher requirements on the cleaning efficiency, resource utilization, service adaptability, and safety protection levels of unmanned car wash technology. This invention relates to the interdisciplinary fields of computer vision, deep learning, edge computing, and intelligent car wash equipment control, specifically an intelligent vehicle cleaning demand judgment system for unmanned car wash scenarios.

[0003] The core application areas of this technical solution cover the entire scenario of unmanned car wash service industry, including unmanned car wash stations in underground parking lots of urban commercial complexes and residential communities, 24-hour unmanned car wash outlets in chain gas stations, and community self-service cleaning service stations; at the same time, it can be extended to adapt to automated cleaning scenarios of commercial freight vehicles in logistics parks, municipal sanitation vehicles, and special engineering vehicles, and can also be deeply integrated with smart parking operation platforms, vehicle networking terminals, and new energy vehicle after-sales service systems to provide unmanned cleaning support services for intelligent connected vehicles.

[0004] Current mainstream unmanned car wash technologies still suffer from numerous unresolved industry-wide pain points, severely hindering the industry's high-quality development. Firstly, at the perception level, existing solutions mostly employ a single RGB visual acquisition method, which cannot effectively address data distortion caused by high reflectivity of car paint, sudden changes in ambient light, and vehicle dynamics during entry. This results in high rates of missed and false detections of dirt, hindering accurate identification of cleaning needs. Secondly, at the decision-making level, existing solutions only focus on basic dirt detection, failing to consider crucial information such as paint material, aging level, protective film properties, and original vehicle damage. They cannot distinguish between cleanable dirt and non-cleanable original vehicle damage, easily leading to irreversible damage to the paint and protective film due to over-cleaning, posing serious service safety risks. Furthermore, at the execution level, existing solutions mostly adopt an open-loop control mode of "pre-detection - fixed process execution," which cannot judge the cleaning effect in real time and dynamically adjust the cleaning parameters during the cleaning process. This generally leads to the contradiction of "one-size-fits-all" cleaning resulting in waste of water resources and cleaning agents, or incomplete cleaning of stubborn stains requiring secondary cleaning. At the same time, the pre-trained models of existing technologies have insufficient generalization ability and cannot quickly adapt to special stain scenarios in different regions and seasons. Moreover, centralized model training poses a risk of leakage of user vehicle data privacy, and cannot meet the core needs of the large-scale and standardized development of the unmanned car wash industry.

[0005] Therefore, this invention provides a deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by the present invention to solve its technical problem is: a deep learning-driven intelligent judgment system for unmanned car washing vehicle cleaning needs, including a multimodal anti-interference data acquisition module, an edge computing processing module, a deep learning model library for car paint-stain dual-branch decoupling, a stain-car paint coupling feature database, a multi-objective Pareto optimal cleaning scheme generation module, a closed-loop execution and feedback module for washing and judging simultaneously, and a federated incremental learning update module. The multimodal anti-interference data acquisition module adopts a ring array layout and integrates a polarization high-definition imaging unit, a 3D lidar, an infrared thermal imaging sensor, a near-infrared spectral analyzer, and a vehicle body micro-motion IMU synchronization unit. It is used to simultaneously acquire all-round polarization differential images, three-dimensional contour data, temperature distribution information, material composition characteristics, and vehicle body micro-vibration data of the vehicle during the dynamic entry of the vehicle, and eliminate the data distortion caused by the mirror reflection of the paint, sudden changes in ambient light, and dynamic vibration of the vehicle body. The edge computing processing module adopts a CPU+GPU+NPU heterogeneous computing architecture and is equipped with an AI acceleration chip. It is used for local real-time processing of multimodal acquired data, running lightweight deep learning models, and performing cleaning decisions and actuator control. The paint-stain dual-branch decoupled deep learning model library has a built-in dual-branch Transformer network with a shared low-level feature encoder, an object detection model, and a semantic segmentation model. It is used to simultaneously complete stain feature extraction and paint state perception, decouple cleanable stains from non-cleanable paint damage, and output full-dimensional features such as stain type, distribution, adhesion strength, paint clear coat thickness, aging degree, damage type, and protective film properties. The stain-car paint coupling feature database stores labeled coupling sample data, covering stain images, physical and chemical properties, car paint state parameters and corresponding optimal cleaning parameter thresholds under different environments, weather, car paint materials and protective film types, providing prior support for model inference and solution generation; The multi-objective Pareto optimal cleaning solution generation module is based on a multi-objective optimization algorithm. It takes maximizing cleanliness, minimizing resource consumption, minimizing the risk of paint damage, and minimizing cleaning time as optimization objectives. It combines the full-dimensional features of stains and paint output by the model with user preferences to generate a Pareto optimal solution set and output a personalized cleaning solution. The simultaneous washing and judgment closed-loop execution and feedback module is an integrated front-mounted micro sensing unit installed at the front end of the nozzle of the cleaning actuator. It is used to collect the status data of the current cleaning area in real time during the cleaning process, calculate the cleanliness in real time through a twin similarity network, and dynamically adjust the water pressure, cleaning agent ratio, residence time and spray angle of the current nozzle to achieve simultaneous washing and judgment and real-time iterative closed-loop control. The federated incremental learning update module adopts a two-level federated learning architecture of cloud and edge. Edge nodes only upload model gradient parameters and do not upload original user data. After the cloud completes global model aggregation, it is distributed to the edge. At the same time, the edge realizes localized incremental learning through a lightweight adapter layer to adapt to the stain characteristics of different regions and seasons.

[0008] Preferably, the polarization high-definition imaging unit of the multimodal anti-interference data acquisition module is composed of polarization industrial cameras, each camera is equipped with four sets of linear polarizers, and the polarization differential imaging algorithm is used to eliminate the reflective spots on the car paint surface; the vehicle body micro-motion IMU synchronization unit has a sampling frequency ≥200Hz; and the end-to-end response delay of the edge computing processing module is ≤30ms.

[0009] Preferably, the dual-branch Transformer network of the paint-stain dual-branch decoupling deep learning model library includes a shared feature encoder, a stain feature branch, a paint state branch, and an orthogonal decoupling head. The shared feature encoder is used to extract the low-level general features of multimodal data, the stain feature branch is used to output the type, location, and adhesion strength features of stains, and the paint state branch is used to output the paint material, clear coat thickness, aging degree, scratch / paint peel damage, and paint protection film / color change film attribute features. The orthogonal decoupling head eliminates the coupling interference between stain features and paint features through feature orthogonal constraints.

[0010] Preferably, the dual-branch Transformer network incorporates a paint safety threshold constraint mechanism. Based on the output of the paint state branch, it automatically generates the upper limit threshold of cleaning parameters, including maximum water pressure, maximum cleaning agent pH, and longest spray dwell time, to prevent secondary damage to the paint and protective film during the cleaning process.

[0011] Preferably, the target detection model is a YOLOv8 model that incorporates a PSFSConv spatial feature separation convolutional layer and small target detection enhancement anchor boxes; the semantic segmentation model is a U-Net semantic segmentation model, used to achieve pixel-level segmentation of the stain region.

[0012] Preferably, the multi-objective Pareto optimal cleaning scheme generation module constructs a four-dimensional multi-objective optimization function, the objective function expression of which is: in, This is a predicted cleanliness value; Values ​​for water resources and cleaning agents; This represents the risk value for paint damage. Total cleaning time; The multi-objective optimization algorithm is the NSGA-III algorithm. It generates a non-dominated Pareto optimal solution set through the algorithm, and then matches and outputs the optimal cleaning solution based on the user's selected fast cleaning mode, paint protection cleaning mode, and deep cleaning mode.

[0013] Preferably, the front-mounted miniature sensing unit of the simultaneous washing and judgment closed-loop execution and feedback module consists of a miniature polarization camera and a miniature near-infrared spectral sensor, which are bound and installed one-to-one with the high-pressure nozzles and move synchronously with the nozzles; the twin similarity network is used to compare the baseline features of the current area before cleaning with the real-time features during the cleaning process in real time, and output the area cleanliness score. When the cleanliness does not reach the preset threshold, the cleaning parameters of the nozzle are adjusted in real time until the cleaning standard is met. The closed-loop execution and feedback module for washing and judging also includes a global re-inspection unit after cleaning. This unit collects data on the entire surface of the vehicle after the entire cleaning process is completed, calculates the overall cleanliness, and automatically generates a targeted secondary cleaning plan when the cleanliness of a local area does not meet the standard.

[0014] Preferably, the federated incremental learning update module includes a cloud-based global model management unit and an edge-based local adaptation unit; the edge-based local adaptation unit adopts an incremental quantization learning mechanism, updating only the lightweight adapter layer parameters of the dual-branch Transformer network without modifying the backbone network weights; the cloud-based global model management unit uses a federated averaging algorithm to securely aggregate the gradient parameters uploaded by multiple edge nodes, update the global model, and simultaneously protect data privacy.

[0015] Preferably, the stain-car paint coupling feature database contains coupling samples of common stain types, car paint materials, and protective film types. The dimensions include at least adhesion strength, ambient temperature, and light conditions. The database supports automatic incremental updates and synchronously stores the actual effect data of the cleaning solution, providing prior data support for model optimization. The edge computing processing module supports offline operation when the network is down. It can store ≥30 days of running data and model parameters locally. After the network is restored, it automatically synchronizes gradient parameters to the cloud, making it suitable for deployment at sites without a stable network.

[0016] Preferably, it also includes a user interaction module, which supports receiving user cleaning preference settings via mobile terminals and on-site touch screens, and simultaneously displays to users the paint condition detection results, stain distribution heat map, cleaning solution details and resource consumption estimates, and supports users to manually adjust the cleaning mode.

[0017] The beneficial effects of this invention are as follows: 1. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs described in this invention solves the problems of high mirror reflection of car paint, sudden changes in ambient light, and vehicle dynamic entry vibration caused by the distortion of collected data and the high rate of missed and false detection of stains in the prior art. This is achieved by using a multimodal anti-interference data acquisition module that integrates a polarization high-definition imaging unit and a vehicle micro-motion IMU synchronization unit, combined with polarization differential imaging algorithm and multi-sensor spatiotemporal synchronous registration technology.

[0018] 2. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs described in this invention constructs a dual-branch decoupled deep learning model library for car paint and stains. It adopts a dual-branch Transformer network with a shared underlying feature encoder and an orthogonal decoupled head structure to simultaneously complete the extraction of full-dimensional features of stains and the perception of car paint status. It decouples cleanable stains from non-cleanable original vehicle damage. Combined with the built-in car paint safety threshold constraint mechanism, it solves the problems of existing technologies that only focus on stain detection, cannot distinguish between stains and original vehicle damage to the car paint / protective film, and are prone to irreversible secondary damage to the car paint due to over-cleaning.

[0019] 3. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs described in this invention constructs a millisecond-level real-time closed-loop control architecture of "cleaning-detection-decision-adjustment" by setting up front-end micro-sensing units that are bound one-to-one with the nozzles of the cleaning execution mechanism, combined with a twin similarity network real-time cleanliness calculation mechanism. This enables simultaneous judgment and targeted dynamic adjustment while washing, solving the problems of "one-size-fits-all" water and cleaning agent waste, incomplete cleaning of stubborn stains, and high secondary cleaning trigger rate caused by the open-loop control mode of "pre-detection-fixed process execution" in existing technologies.

[0020] 4. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs described in this invention employs a multi-objective Pareto optimal cleaning solution generation module based on the NSGA-III algorithm. This module uses four-dimensional optimization objectives—maximizing cleanliness, minimizing resource consumption, minimizing paint damage risk, and minimizing cleaning time—to generate a non-dominated Pareto optimal solution set and match it with user preferences to output personalized solutions. This solves the problem of existing technologies having a single cleaning decision dimension and failing to simultaneously achieve global optimization of multiple objectives, including cleaning effect, resource utilization, paint protection, and cleaning efficiency.

[0021] 5. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs described in this invention constructs a two-level federated incremental learning and update architecture between the cloud and the edge. Edge nodes only upload model gradient parameters and do not upload original user vehicle data. Combined with the localized incremental learning mechanism of the lightweight adapter layer, it solves the problems of insufficient generalization ability of pre-trained models in existing technologies, inability to quickly adapt to special stain scenarios in different regions and seasons, and the risk of user data privacy leakage in centralized model training.

[0022] 6. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs described in this invention, by adopting an edge computing processing module with a CPU+GPU+NPU heterogeneous computing architecture, realizes localized offline execution of the entire process of data processing, model inference and control decision-making, solving the problems of existing technologies that are highly dependent on cloud computing power, cannot be adapted to deployment at remote sites without stable networks, and cannot meet the requirements of 24-hour unattended stable operation. Attached Figure Description

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Figure 1 This is a schematic diagram of the overall business process of the system in this invention; Figure 2 This is a flowchart of the multimodal anti-interference data acquisition and preprocessing process in this invention; Figure 3 This is the deep learning inference flowchart of the dual-branch decoupling of car paint and stain in this invention; Figure 4 This is a flowchart of the multi-objective Pareto optimal cleaning scheme generation process in this invention; Figure 5 This is a flowchart of the real-time closed-loop control execution process for simultaneous washing and judging in this invention; Figure 6 This is a flowchart of the iterative update process of the federated incremental learning model in this invention. Detailed Implementation

[0025] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0026] like Figures 1 to 6 As shown in the embodiment of the present invention, a deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs includes a multimodal anti-interference data acquisition module, an edge computing processing module, a deep learning model library for car paint-stain dual-branch decoupling, a stain-car paint coupling feature database, a multi-objective Pareto optimal cleaning scheme generation module, a closed-loop execution and feedback module for washing and judging simultaneously, and a federated incremental learning update module. The aforementioned multimodal anti-interference data acquisition module adopts a ring array layout and integrates a polarization high-definition imaging unit, a 3D lidar, an infrared thermal imaging sensor, a near-infrared spectral analyzer, and a vehicle micro-motion IMU synchronization unit. It is used to simultaneously acquire all-round polarization differential images, three-dimensional contour data, temperature distribution information, material composition characteristics, and vehicle micro-vibration data of the vehicle during the dynamic entry of the vehicle, and to eliminate the data distortion caused by the mirror reflection of the paint, sudden changes in ambient light, and dynamic vibration of the vehicle body. The aforementioned edge computing processing module adopts a CPU+GPU+NPU heterogeneous computing architecture and is equipped with an AI acceleration chip. It is used for local real-time processing of multimodal acquired data, running lightweight deep learning models, and performing cleaning decisions and actuator control. The aforementioned deep learning model library for decoupling the dual branches of paint and stains incorporates a dual-branch Transformer network with a shared underlying feature encoder, an object detection model, and a semantic segmentation model. This model is used to simultaneously extract stain features and perceive the state of the paint, decoupling cleanable stains from non-cleanable paint damage, and outputting full-dimensional features such as stain type, distribution, adhesion strength, paint clear coat thickness, aging degree, damage type, and protective film properties. The aforementioned stain-car paint coupling feature database stores labeled coupling sample data, covering stain images, physical and chemical properties, car paint state parameters and corresponding optimal cleaning parameter thresholds under different environments, weather, car paint materials and protective film types, providing prior support for model inference and solution generation; The multi-objective Pareto optimal cleaning solution generation module mentioned above is based on a multi-objective optimization algorithm. It takes maximizing cleanliness, minimizing resource consumption, minimizing the risk of paint damage, and minimizing cleaning time as optimization objectives. It combines the full-dimensional features of dirt and paint output by the model with user preferences to generate a Pareto optimal solution set and match and output a personalized cleaning solution. The aforementioned closed-loop execution and feedback module for simultaneous washing and judgment is integrated with a front-mounted micro-sensing unit installed at the front end of the nozzle of the cleaning actuator. It is used to collect the status data of the current cleaning area in real time during the cleaning process, calculate the cleanliness in real time through a twin similarity network, and dynamically adjust the water pressure, cleaning agent ratio, residence time and spray angle of the current nozzle to achieve closed-loop control of simultaneous washing and judgment and real-time iteration. The aforementioned federated incremental learning update module adopts a two-level federated learning architecture of cloud and edge. Edge nodes only upload model gradient parameters and do not upload original user data. After the cloud completes global model aggregation, it is distributed to the edge. At the same time, the edge achieves localized incremental learning through a lightweight adapter layer to adapt to the stain characteristics of different regions and seasons.

[0027] like Figures 1 to 6As shown, the polarization high-definition imaging unit of the aforementioned multimodal anti-interference data acquisition module consists of eight 24-megapixel polarization industrial cameras. Each camera is equipped with four sets of linear polarizers at 0°, 45°, 90°, and 135°. The polarization differential imaging algorithm eliminates the reflective spots on the car paint surface, increasing the effective pixel ratio of the image from ≤35% to ≥98% in high-reflectivity scenarios. The aforementioned vehicle micro-motion IMU synchronization unit has a sampling frequency of ≥200Hz and is used to collect lateral, longitudinal, and vertical micro-shake data during the vehicle's entry process in real time. It performs spatiotemporal synchronization registration and anti-shake correction on multi-sensor data, with a frame alignment error of ≤1ms. The end-to-end response latency of the aforementioned edge computing processing module is ≤30ms.

[0028] like Figures 1 to 6 As shown, the dual-branch Transformer network of the aforementioned paint-stain dual-branch decoupling deep learning model library includes a shared feature encoder, a stain feature branch, a paint state branch, and an orthogonal decoupling head. The shared feature encoder is used to extract the low-level general features of multimodal data. The stain feature branch is used to output the type, location, and adhesion strength features of stains. The paint state branch is used to output the paint material, clear coat thickness, aging degree, scratch / paint peeling damage, and paint protection film / color change film attribute features. The orthogonal decoupling head eliminates the coupling interference between stain features and paint features through feature orthogonal constraints, achieving a stain-paint damage differentiation accuracy of ≥99.2%. The complete expression for the characteristic orthogonal loss function of the above orthogonal decoupling head is: in, Loss due to stain detection; Loss due to stain separation; Classify losses according to the condition of the vehicle paint; This is the orthogonal loss weighting coefficient, with a value of 0.3; The feature orthogonal loss is expressed as follows: In the formula, This is the normalized feature matrix output by the stain feature branch. This is the normalized feature matrix output by the paint state branch. It is the Frobenius norm. During training, it is minimized... This constrains the stain features and paint features to be orthogonal to each other in the feature space, eliminating the coupling interference between the two types of features and avoiding misjudging paint scratches and textures as stains.

[0029] like Figures 1 to 6 As shown, the dual-branch Transformer network has a built-in safety threshold constraint mechanism for vehicle paint. Based on the output of the vehicle paint state branch, it automatically generates the upper limit threshold of the cleaning parameters, including the maximum water pressure, the highest pH of the cleaning agent, and the longest spray dwell time, to prevent secondary damage to the vehicle paint and protective film during the cleaning process.

[0030] like Figures 1 to 6 As shown, the above target detection model is a YOLOv8 model that introduces PSFSConv spatial feature separation convolutional layers and small target detection enhancement anchor boxes. The model size is compressed to less than 7MB, the edge inference speed is ≥65fps, the detection accuracy of tiny stains ≤5×5mm is improved by 15.3%, and the overall stain detection accuracy is ≥98.7%. The above semantic segmentation model is the U-Net semantic segmentation model, which is used to achieve pixel-level segmentation of stain regions. The PSFSConv spatial feature separation convolutional layer described above has a core structure of a dual-branch parallel computation architecture: the input feature map is divided into a foreground feature branch and a background feature branch along the channel dimension. An attention mechanism identifies the foreground region where the stain is located. Standard 3×3 convolutions are performed only on the foreground feature branch, while the background feature branch is downsampled using a 1×1 convolution and then the convolutional layer is skipped. Finally, the feature maps from the two branches are concatenated and fused. This structure can reduce redundant computation by more than 60% without sacrificing detection accuracy, achieving a lightweight model.

[0031] like Figures 1 to 6 As shown, the above multi-objective Pareto optimal cleaning scheme generation module constructs a four-dimensional multi-objective optimization function, the objective function expression of which is: in, This is a predicted cleanliness value; Values ​​for water resources and cleaning agents; This represents the risk value for paint damage. Total cleaning time; The above multi-objective optimization algorithm is the NSGA-III algorithm. It generates a non-dominated Pareto optimal solution set through the algorithm, and then matches and outputs the optimal cleaning solution based on the user's selected fast cleaning mode, paint protection cleaning mode, and deep cleaning mode. The solution adaptation time is ≤20ms. The aforementioned NSGA-III algorithm introduces an adaptive crossover and mutation operator to optimize continuous variables of cleaning parameters. The crossover and mutation probabilities are adaptively adjusted with the number of generations: in the initial evolution stage, the crossover probability is set to 0.9 and the mutation probability to 0.1 to ensure global search capability; in the later evolution stage, the crossover probability decreases to 0.6 and the mutation probability increases to 0.2 to improve local optimization accuracy. Simultaneously, a paint safety threshold is embedded as a hard constraint in the algorithm; solutions exceeding the safety threshold are directly marked as infeasible, ensuring that the generated cleaning solutions always remain within a safe range.

[0032] like Figures 1 to 6 As shown, the front-end miniature sensing unit of the above-mentioned washing and judging closed-loop execution and feedback module consists of a miniature polarization camera and a miniature near-infrared spectral sensor, which are bound and installed one-to-one with the high-pressure nozzles and move synchronously with the nozzles; the above-mentioned twin similarity network is used to compare the baseline features of the current area before cleaning with the real-time features during the cleaning process, and output the area cleanliness score. When the cleanliness does not reach the preset threshold, the cleaning parameters of the nozzle are adjusted in real time until the cleaning standard is met, and the response delay of a single area adjustment is ≤10ms; The aforementioned twin similarity network employs a lightweight Swin-Tiny backbone network with dual-path shared weights. The input is a 256×256 pixel region feature map. The two inputs are the baseline feature map of the region before cleaning and the real-time feature map during the cleaning process, respectively. The network outputs the cosine similarity between the two feature maps, which is mapped to a cleanliness score of 0-100. The network is trained using a contrastive learning approach. Positive samples are pairs of images of the same region before and after cleaning that meet the standards, while negative samples are pairs of images that are not cleaned properly or are over-cleaned. After training, the cleanliness score error is ≤2%. The closed-loop execution and feedback module for washing and judging also includes a global re-inspection unit after cleaning. This unit collects data on the entire surface of the vehicle after the entire cleaning process is completed, calculates the overall cleanliness, and automatically generates a targeted secondary cleaning plan when the cleanliness of a local area does not meet the standard.

[0033] like Figures 1 to 6 As shown, the aforementioned federated incremental learning update module includes a cloud-based global model management unit and an edge-based local adaptation unit. The edge-based local adaptation unit adopts an incremental quantization learning mechanism, updating only the parameters of the lightweight adapter layer of the dual-branch Transformer network without modifying the backbone network weights. This reduces the computational power requirement for incremental training by 92%, and shortens the adaptation cycle for local special stains from 7 days to 4 hours. The cloud-based global model management unit uses a federated averaging algorithm to securely aggregate the gradient parameters uploaded by multiple edge nodes, update the global model, and simultaneously protect data privacy. The aforementioned lightweight adapter layer serves as the bottleneck structure, embedded within each Transformer block of the dual-branch Transformer network. Its input dimension matches the output dimension of the Transformer block, and the bottleneck layer's dimension is 1 / 16th of the input dimension. The overall number of parameters accounts for only 8% of the backbone network. During incremental training, all weights of the backbone network are frozen, and only the adapter layer's parameters are updated. Model adaptation can be completed using a small number of local samples, eliminating the need for high-performance GPUs. A single round of incremental training at the edge takes ≤2 hours.

[0034] like Figures 1 to 6 As shown, the above-mentioned stain-car paint coupling feature database contains coupling samples of at least 25 common stain types, 8 types of car paint materials, and 6 types of protective films. The number of labeled samples for each coupling scenario is ≥6000 sets, and the dimensions include at least adhesion strength, ambient temperature, and lighting conditions. The database supports automatic incremental updates and synchronously stores the actual effect data of cleaning solutions, providing prior data support for model optimization. The aforementioned edge computing processing module supports offline operation when the network is down. It can store ≥30 days of running data and model parameters locally. After the network is restored, it automatically synchronizes gradient parameters to the cloud, making it suitable for deployment at sites without a stable network.

[0035] like Figures 1 to 6 As shown, it also includes a user interaction module, which supports receiving user cleaning preference settings via mobile terminals and on-site touch screens, and simultaneously displays to users the paint condition detection results, stain distribution heat map, cleaning solution details and resource consumption estimates, and supports users to manually adjust the cleaning mode.

[0036] Example 1 Heavy pollution conditions caused by snow melting agents and sediment mixture in northern winters 1. Test conditions Vehicle: Black compact SUV, original black metallic paint, no protective film, slight aging of the clear coat (3 years of use), two minor scratches 3cm long on the front bumper; Stain condition: The vehicle body is covered with a mixture of de-icing agent and mud stains from northern winter roads, with an overall coverage rate of 92%. The adhesion strength of the stains on the wheel hubs and chassis is level 5 (the highest level), and the stains contain corrosive salt components. Environmental conditions: Under direct sunlight on a sunny day, the car paint surface has a large area of ​​mirror-like reflective spots, simulating an outdoor commercial scene.

[0037] 2. System processing workflow The vehicle dynamically enters the detection area at a speed of 3km / h. The IMU synchronization unit collects the micro-shaking data of the vehicle body in real time, completes the spatiotemporal registration of multi-sensor data, and the frame alignment error is ≤1ms. The polarization imaging unit eliminates the mirror reflection of the paint surface through the four polarization difference algorithm. The effective pixel ratio of the high reflective area is 98.7%, and there is no light spot obstruction. The dual-branch Transformer network performs synchronous inference: the stain branch identifies the distribution and adhesion strength of the snow melting agent-mud and sand mixed stains, with a 100% accuracy rate in detecting minute stains; the paint branch detects slight aging of the paint and minor scratches on the front bumper, and outputs safety thresholds: maximum water pressure ≤70 bar, high-pressure direct spraying is prohibited in the scratched area, and alkaline cleaning agents with pH>9 are prohibited. Multi-objective optimization module: When the user selects the "Deep Cleaning for Paint Protection" mode, with paint protection as the highest priority, the Pareto optimal cleaning solution is generated through the NSGA-III algorithm. The core parameters are: neutral cleaning agent for snow melting agent, chassis pre-rinse for 40 seconds, low-pressure immersion of the body for 30 seconds, low-pressure circumferential cleaning of scratched areas using 40 bar, and focused cleaning of wheel hubs. Closed-loop execution with simultaneous washing and judgment: The micro-sensing unit in front of the nozzle moves synchronously with the high-pressure nozzle to collect data on the cleaning area in real time and calculate the cleanliness through a twin similarity network; for stubborn stains on the wheel hub, the dwell time is automatically extended and the concentration of cleaning agent is gradually increased. Once the cleanliness meets the standard, it immediately moves to the next area, with a single adjustment response delay of ≤10ms; Global Re-inspection: After the entire cleaning process is completed, data on the entire surface of the vehicle is collected. The overall cleanliness is 98.8%, with no substandard areas and no secondary cleaning is triggered.

[0038] 3. Implementation Results Cleanliness: 98.8 points, 100% up to standard; Paint gloss change rate: 0.62%, no new damage; Water consumption: 42.3L; Cleaning agent consumption: 185 mL; Total cleaning time: 148 seconds; Secondary cleaning trigger rate: 0; Accuracy in distinguishing between stains and paint damage: 100%.

[0039] Example 2 Coastal salt spray pollution + TPU paint protection film working conditions 1. Test conditions Vehicle: White mid-size sedan, fully wrapped with TPU paint protection film, film has been in use for 1 year, there are multiple seams in the film, no damage to the paint. Stain condition: The vehicle body is covered with a mixture of coastal salt spray and dust stains, with an overall coverage of 65%. There are 4 fresh bird droppings on the hood and roof, and insect remains on the lower edge of the doors. The stains contain corrosive salt components. Environmental conditions: Cloudy, low-light environment, simulating a car wash station scene in an underground parking lot.

[0040] 2. System processing workflow The vehicle entered the inspection area at a speed of 2 km / h. The polarization imaging unit clearly identified the seam of the car cover. There was no glare interference, and data from multiple sensors were collected simultaneously. Dual-branch Transformer network inference: The stain branch identifies the location and adhesion strength of salt spray stains, bird droppings, and insect carcasses; the paint branch accurately identifies the properties of TPU invisible car wrap and the location of the car wrap seams, and outputs safety thresholds: disable strong alkaline / strong solvent-based cleaners, the maximum water pressure in the car wrap seam area is ≤50 bar, and high-pressure direct spraying to the seams is prohibited. Multi-objective optimization module: When the user selects the "Car Cover Protective Wash" mode, the Pareto optimal solution is generated. The core parameters are: neutral detergent for car covers, pre-soaking bird droppings areas for 25 seconds, low-pressure surround cleaning of seam areas, and no excessive high-pressure rinsing. Closed-loop cleaning and assessment: For bird droppings areas, cleanliness is monitored in real time, and soaking and rinsing are stopped immediately once the standard is met, without over-cleaning; low-pressure cleaning is applied to the seams of the car cover throughout the process, without direct high-pressure spraying; Global re-inspection: Overall cleanliness is 99.1%, with no areas failing to meet standards and no secondary cleaning triggered.

[0041] 3. Implementation Results Cleanliness: 99.1 points, 100% up to standard; The gloss change rate of the car wrap was 0.38%, with no corrosion, no curling, and no damage. Water consumption: 28.7L; Cleaning agent consumption: 120mL; Total cleaning time: 112 seconds; Secondary cleaning trigger rate: 0; Accuracy in distinguishing stains from car cover seams: 100%.

[0042] Example 3 Urban commuting with mild dust pollution conditions 1. Test conditions Vehicle: White family sedan, original paint, no damage, no protective film, 2 years old; Stain condition: The vehicle body is covered with light dust from urban roads, with an overall coverage rate of 32%. There are no stubborn stains, and only a few minor mud spots are present on the rearview mirror and the lower edge of the door. Environmental conditions: sunny daytime with strong sunlight, simulating an outdoor commercial car wash station scenario.

[0043] 2. System processing workflow The vehicle dynamically enters the detection area, polarization imaging eliminates reflections, and multi-sensor data acquisition is completed in just 8 seconds. Dual-branch network inference: Identify as light dust stains, the paint is in good condition and undamaged, and output a safety threshold without special constraints; Multi-objective optimization module: When the user selects the "Rapid Energy-Saving Wash" mode, the core optimization objectives are the shortest cleaning time and the least resource consumption. The Pareto optimal solution is generated with the following core parameters: low-pressure pre-rinse 10s, foam cleaning 15s, low-pressure rinse 12s, and rapid air drying, with no redundant cleaning steps. The system operates in a closed-loop manner, washing and judging simultaneously: cleaning is only performed on localized areas with mud spots, while large clean areas are skipped, resulting in no ineffective rinsing. Global re-inspection: Overall cleanliness level 98.2%, no areas failing to meet standards, no secondary cleaning triggered.

[0044] 3. Implementation Results Cleanliness: 98.2 points, 100% up to standard; Paint gloss change rate: 0.21%, no damage; Water consumption: 16.5L; Cleaning agent consumption: 65mL; Total cleaning time: 72 seconds; Secondary cleaning trigger rate: 0; Stain detection accuracy: 99.3%.

[0045] Example 4 Localized extremely stubborn stains 1. Test conditions Vehicle: Black SUV, original paint, no protective film, 1.5 years old, paint in good condition; Stain condition: The vehicle body is covered with light dust. There are 3 tree sap stains that have been cured for 72 hours on the hood, a large number of insect corpse stains left over from high-speed driving on the front bumper, and asphalt splash stains on the lower edge of the door. All of these are stubborn stains with high adhesion strength, and the area of ​​each stain is ≤10cm². Environmental conditions: sunny day with strong sunlight.

[0046] 2. System processing workflow The vehicle entered the testing area, and multimodal data acquisition was completed. The improved YOLOv8 model achieved 100% accuracy in detecting tiny, stubborn stains with no missed detections. Dual-branch network inference: Identify the type and adhesion strength of three stubborn stains: tree sap, insect remains, and asphalt. If the car paint is in good condition, output a safety threshold: Use a special cleaning agent for tree sap / asphalt areas, with a maximum dwell time of ≤60s to avoid car paint corrosion. Multi-target optimization module: Generates targeted cleaning solutions, performing specialized treatment only on areas with stubborn stains, and light cleaning on other areas; Closed-loop execution with simultaneous washing and assessment: For areas with tree sap, a special sap remover is sprayed first, and the dissolution status and cleanliness of the stains are monitored in real time. Once the standard is met, the process is stopped immediately to avoid prolonged contact between the cleaning agent and the paint. For areas with insect remains and asphalt, graded pressure cleaning is used. Once the cleanliness standard is met, the process is moved immediately to avoid over-cleaning. Overall cleanliness: 98.5% overall cleanliness, all stubborn stains were cleaned to the required standard, and no secondary cleaning was triggered.

[0047] 3. Implementation Results Cleanliness score: 98.5 points; 100% success rate in cleaning stubborn stains. Paint gloss change rate: 0.45%, no corrosion, no damage; Water consumption: 22.4L; Cleaning agent consumption: 95mL (including 20mL of special adhesive remover); Total cleaning time: 98 seconds; Secondary cleaning trigger rate: 0; Accuracy in detecting minute and stubborn stains: 100%.

[0048] Comparative Example 1 Traditional fixed-process tunnel-type unmanned car wash system (current mainstream commercial solution) 1. Core features of the system It has no stain detection or intelligent decision-making modules, adopts the industry's standard fixed cleaning process, and executes uniform parameters throughout: pre-rinse 60s → foam spray 60s → roller brush scrub 60s → clean water rinse 60s → air dry 90s, with a fixed water pressure of 80 bar and a fixed alkaline cleaning agent ratio.

[0049] 2. Test Results Cleanliness: 82.3 points. The wheel hubs and chassis have a large amount of residual de-icing agent stains, which does not meet the standard. Paint gloss change rate: 3.87%, with significant expansion of the scratched area on the front bumper, indicating accelerated paint aging; Water consumption: 85.6L; Cleaning agent consumption: 320mL; Total cleaning time: 330s; Secondary cleaning trigger rate: 100%; Ability to distinguish between stains and paint damage: None, completely lacking the ability to distinguish between them.

[0050] Comparative Example 2 Existing single-RGB visual stain detection intelligent car wash system (existing published patent typical solution) 1. Core features of the system It is equipped with only a single RGB industrial camera, lacking polarization imaging, IMU synchronization, paint condition detection, and edge washing and judgment closed-loop; it uses a common YOLOv5 model to detect stains, generates a fixed cleaning plan in advance, executes it in an open loop, and has no dynamic adjustment capability.

[0051] 2. Test Results Cleanliness: 89.7 points. Stains were missed in areas obscured by strong light reflection, and scratched areas were mistakenly identified as excessively cleaned. The result was not up to standard. Paint gloss change rate: 2.15%, with obvious damage in the scratched area; Water consumption: 68.2L; Cleaning agent consumption: 260mL; Total cleaning time: 245 seconds; Secondary cleaning trigger rate: 75%; Accuracy in distinguishing between stains and paint damage: 62.4%, with a large number of false positives and false negatives.

[0052] Comparative Example 3 The simplified system of this invention lacks a closed-loop execution module for simultaneous washing and judging. 1. Core features of the system The invention retains all modules except for the "closed-loop execution and feedback module for simultaneous washing and judgment", and still adopts the process of "pre-detection - generation of fixed scheme - open-loop execution - post-re-inspection", without the ability of nozzle pre-positioned real-time sensing and dynamic adjustment.

[0053] 2. Test Results Cleanliness: 94.2 points. Some stubborn stains on the wheel hubs were not cleaned to the required standard, and the overall performance did not meet the preset standard. The change rate of paint gloss was 0.71%, indicating no obvious damage. Water consumption: 56.8L; Cleaning agent consumption: 220mL; Total cleaning time: 185 seconds; Secondary cleaning trigger rate: 100%; Accuracy in distinguishing between stains and paint damage: 99.1%.

[0054] Core Summary of Tabular Data The complete solution of this invention achieves a dual qualitative leap in cleaning effect and efficiency: the cleanliness of the four examples covering all commercial scenarios is consistently above 98 points, the secondary cleaning trigger rate is reduced to 0, and the industry problem of incomplete cleaning by existing technologies is completely solved; compared with the mainstream traditional fixed process solution (Comparative Example 1), it saves an average of 50.6% of water, saves 42.2% of cleaning agent, shortens the cleaning time by 55.2%, and significantly reduces operating costs.

[0055] Comparing the complete solution of this invention with the simplified solution without closed-loop control (Comparative Example 3), the only missing module is "real-time closed-loop control for simultaneous washing and judgment". The cleaning compliance rate drops from 100% to 0, the secondary cleaning trigger rate increases to 100%, and the water consumption increases by 34.3% and the cleaning time is extended by 25%. This directly proves that this module is the core technical support for achieving precise on-demand cleaning.

[0056] This invention completely overcomes the safety concerns of unmanned car washes: the gloss change rate of car paint / car cover in all embodiments is ≤0.62%, with no new damage; while existing technical solutions (comparative examples 1 and 2) all have varying degrees of car paint corrosion, scratch expansion, and car cover peeling problems, with the traditional solution having a car paint damage rate of 100%. This invention eliminates the service safety hazards of unmanned car washes from the root of the technology.

[0057] The solution has strong adaptability to all scenarios: from heavy pollution from de-icing agents in northern winters and protection from salt spray on coastal car covers, to light cleaning for urban commutes and targeted treatment of stubborn stains, the invention exhibits stable and excellent performance under all typical working conditions and can be directly adapted to the commercial needs of various unmanned car wash businesses.

[0058] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs, characterized in that: It includes a multimodal anti-interference data acquisition module, an edge computing processing module, a deep learning model library for decoupling the dual branches of paint and stains, a stain-paint coupling feature database, a multi-objective Pareto optimal cleaning solution generation module, a closed-loop execution and feedback module for washing and judging simultaneously, and a federated incremental learning update module. The multimodal anti-interference data acquisition module adopts a ring array layout and integrates a polarization high-definition imaging unit, a 3D lidar, an infrared thermal imaging sensor, a near-infrared spectral analyzer, and a vehicle body micro-motion IMU synchronization unit. It is used to simultaneously acquire all-round polarization differential images, three-dimensional contour data, temperature distribution information, material composition characteristics, and vehicle body micro-vibration data of the vehicle during the dynamic entry of the vehicle, and eliminate the data distortion caused by the mirror reflection of the paint, sudden changes in ambient light, and dynamic vibration of the vehicle body. The edge computing processing module adopts a CPU+GPU+NPU heterogeneous computing architecture and is equipped with an AI acceleration chip. It is used for local real-time processing of multimodal acquired data, running lightweight deep learning models, and performing cleaning decisions and actuator control. The paint-stain dual-branch decoupled deep learning model library has a built-in dual-branch Transformer network with a shared low-level feature encoder, an object detection model, and a semantic segmentation model. It is used to simultaneously complete stain feature extraction and paint state perception, decouple cleanable stains from non-cleanable paint damage, and output full-dimensional features such as stain type, distribution, adhesion strength, paint clear coat thickness, aging degree, damage type, and protective film properties. The stain-car paint coupling feature database stores labeled coupling sample data, covering stain images, physical and chemical properties, car paint state parameters and corresponding optimal cleaning parameter thresholds under different environments, weather, car paint materials and protective film types, providing prior support for model inference and solution generation; The multi-objective Pareto optimal cleaning solution generation module is based on a multi-objective optimization algorithm. It takes maximizing cleanliness, minimizing resource consumption, minimizing the risk of paint damage, and minimizing cleaning time as optimization objectives. It combines the full-dimensional features of stains and paint output by the model with user preferences to generate a Pareto optimal solution set and output a personalized cleaning solution. The simultaneous washing and judgment closed-loop execution and feedback module is an integrated front-mounted micro sensing unit installed at the front end of the nozzle of the cleaning actuator. It is used to collect the status data of the current cleaning area in real time during the cleaning process, calculate the cleanliness in real time through a twin similarity network, and dynamically adjust the water pressure, cleaning agent ratio, residence time and spray angle of the current nozzle to achieve simultaneous washing and judgment and real-time iterative closed-loop control. The federated incremental learning update module adopts a two-level federated learning architecture of cloud and edge. Edge nodes only upload model gradient parameters and do not upload original user data. After the cloud completes global model aggregation, it is distributed to the edge. At the same time, the edge realizes localized incremental learning through a lightweight adapter layer to adapt to the stain characteristics of different regions and seasons.

2. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: The polarization high-definition imaging unit of the multimodal anti-interference data acquisition module is composed of polarization industrial cameras. Each camera group is equipped with four sets of linear polarizers and uses polarization differential imaging algorithm to eliminate the reflective spots on the car paint surface. The vehicle body micro-motion IMU synchronization unit has a sampling frequency ≥200Hz. The end-to-end response delay of the edge computing processing module is ≤30ms.

3. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: The dual-branch Transformer network of the paint-stain dual-branch decoupling deep learning model library includes a shared feature encoder, a stain feature branch, a paint state branch, and an orthogonal decoupling head. The shared feature encoder is used to extract the low-level general features of multimodal data, the stain feature branch is used to output the type, location, and adhesion strength features of stains, and the paint state branch is used to output the paint material, clear coat thickness, aging degree, scratch / paint peeling damage, and paint protection film / color change film attribute features. The orthogonal decoupling head eliminates the coupling interference between stain features and paint features through feature orthogonal constraints.

4. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 3, characterized in that: The dual-branch Transformer network incorporates a paint safety threshold constraint mechanism. Based on the output of the paint status branch, it automatically generates the upper limit threshold of cleaning parameters, including maximum water pressure, maximum cleaning agent pH, and longest spray dwell time, to prevent secondary damage to the paint and protective film during the cleaning process.

5. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: The target detection model is a YOLOv8 model that incorporates PSFSConv spatial feature separation convolutional layers and small target detection enhancement anchor boxes; the semantic segmentation model is a U-Net semantic segmentation model, used to achieve pixel-level segmentation of the stain region.

6. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: The multi-objective Pareto optimal cleaning scheme generation module constructs a four-dimensional multi-objective optimization function, the objective function expression of which is: in, This is a predicted cleanliness value; Values ​​for water resources and cleaning agents; This represents the risk value for paint damage. Total cleaning time; The multi-objective optimization algorithm is the NSGA-III algorithm. It generates a non-dominated Pareto optimal solution set through the algorithm, and then matches and outputs the optimal cleaning solution based on the user's selected fast cleaning mode, paint protection cleaning mode, and deep cleaning mode.

7. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: The front-mounted miniature sensing unit of the simultaneous washing and judgment closed-loop execution and feedback module consists of a miniature polarization camera and a miniature near-infrared spectral sensor, which are bound and installed one-to-one with the high-pressure nozzles and move synchronously with the nozzles. The twin similarity network is used to compare the baseline features of the current area before cleaning with the real-time features during the cleaning process, and output the area cleanliness score. When the cleanliness does not reach the preset threshold, the cleaning parameters of the nozzle are adjusted in real time until the cleaning standard is met. The closed-loop execution and feedback module for washing and judging also includes a global re-inspection unit after cleaning. This unit collects data on the entire surface of the vehicle after the entire cleaning process is completed, calculates the overall cleanliness, and automatically generates a targeted secondary cleaning plan when the cleanliness of a local area does not meet the standard.

8. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: The federated incremental learning update module includes a cloud-based global model management unit and an edge-based local adaptation unit. The edge-based local adaptation unit adopts an incremental quantization learning mechanism, updating only the lightweight adapter layer parameters of the dual-branch Transformer network without modifying the backbone network weights. The cloud-based global model management unit uses a federated averaging algorithm to securely aggregate the gradient parameters uploaded by multiple edge nodes, update the global model, and simultaneously protect data privacy.

9. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: The stain-car paint coupling feature database contains coupling samples of common stain types, car paint materials, and protective film types. The dimensions include at least adhesion strength, ambient temperature, and light conditions. The database supports automatic incremental updates and synchronously stores the actual effect data of cleaning solutions, providing prior data support for model optimization. The edge computing processing module supports offline operation when the network is down. It can store ≥30 days of running data and model parameters locally. After the network is restored, it automatically synchronizes gradient parameters to the cloud, making it suitable for deployment at sites without a stable network.

10. The deep learning-driven intelligent judgment system for unmanned car wash vehicle cleaning needs according to claim 1, characterized in that: It also includes a user interaction module, which supports receiving user cleaning preference settings via mobile terminals and on-site touch screens, and simultaneously displays to users the paint condition detection results, stain distribution heat map, cleaning solution details and resource consumption estimates, and supports users to manually adjust the cleaning mode.

Images