An automated cleaning control method and system for a self-propelled cleaning robot

Abstract

The application discloses a self-cleaning robot automatic cleaning control method and system, and relates to the technical field of intelligent cleaning robots. The method comprises the following steps: starting the robot and performing system self-checking, and entering an autonomous operation mode after confirming normality; traveling along a preset path, collecting boundary distance data and obstacle information at a fixed period, and fusing and judging in combination with historical window data to identify the boundary of the operation area; when it is determined that the boundary condition is reached, executing speed reduction control and synchronous speed reduction of the cleaning execution module, and performing path direction re-planning after edge safety parking; when it is determined that an obstacle avoidance action occurs and causes a local area to be uncovered based on the obstacle information, performing secondary coverage; collecting environmental information of the cleaned surface to identify stain types and degrees, and dynamically adjusting cleaning parameters according to the identification results. The application realizes intelligent upgrading of the cleaning robot from action execution to understanding cleaning, and significantly improves the cleaning effect and resource utilization efficiency.

Description

Technical Field

[0001] This application relates to the field of intelligent cleaning robot technology, and in particular to an automated cleaning control method and system for a self-propelled cleaning robot. Background Technology

[0002] With the rapid development of the photovoltaic power generation industry and the continuous growth of photovoltaic power plant installed capacity, the problem of dust and ash accumulation on the surface of photovoltaic modules seriously affects power generation efficiency and causes significant losses in power generation revenue. To replace traditional manual cleaning methods, various photovoltaic panel cleaning robots have emerged. In existing technologies, photovoltaic cleaning robots are typically equipped with a walking mechanism, cleaning brush components, a power supply module, and a basic control unit, enabling them to perform automated cleaning operations on the surface of photovoltaic arrays.

[0003] However, existing technologies still have the following technical shortcomings when applied to large-area photovoltaic panel cleaning scenarios: The cleaning control logic of existing self-propelled cleaning robots is mostly based on preset fixed modes, such as setting a uniform walking speed, constant brush rotation speed, or fixed suction power. The cleaning mechanism is usually independent of the motion control system. Once the system enters the operation mode, it starts and runs with constant parameters until the operation is completed. The cleaning parameters are mostly factory-preset empirical values. The system cannot sense changes in the concentration of dust in the environment, nor can it dynamically adjust the brush frequency and suction power according to the robot's real-time walking speed or the type of dirt.

[0004] This defect leads to two extreme problems: when faced with light dust, the robot still performs full-power cleaning, causing unnecessary consumption of electricity, water resources and mechanical parts, and reducing the working time after a single charge; while when faced with stubborn stains with high adhesion such as bird droppings and oil stains, the fixed cleaning mode has a short action time or weak cleaning power, resulting in stubborn stains remaining and requiring manual rework.

[0005] To improve the intelligence level of cleaning, some existing technologies attempt to introduce deep learning-based visual recognition schemes to perceive the type of stains. However, traditional deep learning models typically have millions or even tens of millions of parameters, and their inference process requires the support of high-performance GPUs or dedicated AI acceleration chips. For photovoltaic cleaning robots, due to limitations in cost, power consumption, and size, their main controllers are usually low-computing-power MCUs such as the STM32 series, which are difficult to handle the real-time inference tasks of traditional deep learning models.

[0006] Furthermore, in actual photovoltaic power plant deployment scenarios, collecting samples of stains on the surface of photovoltaic panels is difficult and costly. For specific types of stains such as bird droppings and oil stains, the number of labeled samples that can be obtained is extremely limited. Traditional deep learning methods are prone to overfitting or convergence problems under conditions with fewer than 10 samples, making it difficult to achieve rapid evaluation and deployment in the early stages of industrial applications.

[0007] When existing cleaning robots encounter obstacles such as foreign objects, cables, and support components on the frame of photovoltaic modules, they typically employ strategies such as directly lifting the brush to avoid the obstacle or emergency shutdown. During the brush-lifting obstacle avoidance process, the brush structure is lifted to pass over the obstacle, and the area along its lifting path becomes an uncovered blind spot because the cleaning mechanism is detached from the surface being cleaned. After the obstacle avoidance action is completed, the robot continues to travel along the original path, and the blind spot is not remembered or covered by subsequent operations.

[0008] This strategy directly leads to an imbalance in the algorithmic game between obstacle avoidance safety and cleaning coverage: the higher the safety and obstacle avoidance requirements, the more frequent the brush lifting actions, resulting in more cleaning blind spots; while reducing obstacle avoidance actions in pursuit of cleaning coverage may cause the robot to get stuck or damage the components. This problem is particularly prominent in scenarios where the surface of photovoltaic modules has a large number of obstacles such as frames, blocks, and cables, and the direct consequences are the need for re-sweeping, increased energy consumption, reduced working time, and reduced overall cleaning efficiency.

[0009] Current self-propelled cleaning robots primarily rely on open-loop control architectures for motion control, commonly employing differential drive via dual-side tracks. The controller drives DC motors by outputting PWM signals with a fixed duty cycle. Path maintenance depends heavily on the consistency of the mechanical structure, lacking real-time attitude feedback. Steering often uses infinite loop commands, where the left / right motors rotate at a fixed differential speed for a specific time, without considering the effects of actual friction and terrain slope on the rotational speed. This leads to deviations between the actual and expected steering angles, and accumulated errors can cause path deviations or even out-of-bounds errors.

[0010] Boundary judgment mainly relies on preset time parameters or simple limit switches, lacking intelligent boundary recognition capabilities based on multi-sensor fusion, resulting in a high false judgment rate at the edges and gaps of photovoltaic modules. Summary of the Invention

[0011] To address the aforementioned issues, this application provides an automated cleaning control method and system for a self-propelled cleaning robot. Through multi-sensor fusion for environmental perception, stain recognition based on lightweight deep learning, obstacle avoidance blind spot recording and compensation, and a collaborative and adaptive adjustment mechanism for cleaning parameters, the cleaning robot achieves an intelligent upgrade from performing actions to understanding cleaning.

[0012] The first technical solution adopted in this application is: providing an automated cleaning control method for a self-propelled cleaning robot, including the following steps: Step S1: Start the robot and perform a system self-test. The system self-test includes power status detection, sensor initialization, motor drive status detection, and cleaning execution module linkage test. After confirming that the system is normal, it enters the autonomous operation mode. Step S2: In autonomous operation mode, the robot travels along a preset path, collects boundary distance data and obstacle information at fixed intervals, and combines it with historical window data containing multiple sampling periods for fusion judgment to identify the boundary of the operation area; Step S3: When the boundary conditions are determined to be met based on the fusion judgment result, the deceleration control and cleaning execution modules are executed to reduce speed synchronously, and the path direction is replanned after the vehicle safely stops at the edge. Step S4: When it is determined that an obstacle avoidance action has occurred based on the obstacle information and a local area is not covered, the location of the local uncovered area is recorded, and a compensation cleaning path is inserted in the subsequent path planning to cover the local uncovered area a second time. Step S5: During the autonomous operation, the environmental information of the surface being cleaned is collected by sensors, the type and degree of stains are identified based on the environmental information, and the cleaning parameters are dynamically adjusted according to the identification results.

[0013] In an optional embodiment, step S5, which identifies the type and extent of stains based on the environmental information, specifically includes: Acquire image data of the surface to be cleaned; Multi-domain feature enhancement is performed on the image data to obtain spatial domain features, gradient features, and frequency domain features, wherein the frequency domain features are obtained through fast Fourier transform; The SimAM attention mechanism is used to fuse the spatial domain features, gradient features and frequency domain features. Attention weights are obtained by calculating the spatial domain amplitude energy and the energy response after frequency domain transformation. The attention weights are then used to enhance the features to obtain enhanced features. Stain identification and classification are performed based on the enhanced features.

[0014] In an optional embodiment, the stain identification and classification based on the enhanced features includes: The enhanced features are extracted using a ResNet50 backbone network, and global average pooling is used at the end of the backbone network to replace the fully connected layer for data compression. The similarity score is calculated by multiplying the feature vector after global average pooling with the reference feature vector of each category of stain in the reference feature matrix, and the feature vector distance is calculated based on the L∞ norm to prevent the curse of high dimensionality. The stain category is determined according to the similarity score and the feature vector distance.

[0015] In an optional embodiment, an incremental learning step is also included: when the classification result is incorrect or a new category appears, the corresponding image sample is extracted using only the ResNet50 skeleton network to obtain a new feature vector, and the new feature vector is appended to the reference feature matrix without retraining the entire deep learning network.

[0016] In an optional embodiment, the cleaning parameters in step S5 include the roller brush speed, the vacuum fan power, and the travel speed; the cleaning parameters are adjusted in conjunction with a collaborative control algorithm, which synchronously adjusts the roller brush speed, the vacuum fan power, and the travel speed according to the stain recognition results to ensure a dynamic balance between cleaning intensity and travel efficiency.

[0017] In an optional embodiment, step S2 further includes performing a mileage estimation task; the acquisition of the boundary distance data and the mileage estimation task are executed in parallel through multi-threading or interrupt scheduling; the historical window data is used to suppress transient noise interference.

[0018] In an optional embodiment, the method is applied to the surface cleaning of photovoltaic modules, and the robot is a tracked mobile structure; the dynamic adjustment of cleaning parameters includes: When the identification result is floating dust, the cleaning execution module is controlled to run at the first power level and the robot is controlled to move at the first speed. When the identification result is an adhesive stain, the cleaning execution module is controlled to run at the second power level and the robot is controlled to move at the second speed. Wherein, the first power level is lower than the second power level, and the first speed is greater than the second speed.

[0019] The second technical solution adopted in this application is: providing an automated cleaning control system for a self-propelled cleaning robot, including: The motion control module is used to control the robot to move along a preset path and to perform deceleration control, edge safety stopping, and path replanning when boundary conditions are determined to be reached. The environmental perception module is used to collect boundary distance data, obstacle information, and stain information on the surface being cleaned at fixed intervals. The cleaning execution module, including a roller brush mechanism and a vacuum fan, is used to perform cleaning operations; The main control module is connected to the motion control module, the environmental perception module, and the cleaning execution module, respectively, and is used to dynamically adjust the cleaning parameters of the cleaning execution module according to the data collected by the environmental perception module. The main control module is also used to record the partially uncovered area caused by obstacle avoidance when it is determined that an obstacle avoidance action has occurred based on obstacle information and the local area is not covered, and to insert a compensation cleaning path in the subsequent path planning.

[0020] In an optional embodiment, the main control module is built on an STM32F103 series MCU and is a low-computing-power edge control unit; the environmental perception module includes a dust concentration sensor for detecting dust concentration, an image acquisition unit for acquiring image data, and a light intensity detection unit; the main control module is configured to: prioritize the dust concentration sensor when the light intensity is higher than a preset threshold, and prioritize visual recognition when the light intensity is lower than the preset threshold or in a shadow area.

[0021] In an optional embodiment, a communication module is also included, which is used to receive remote control commands and transmit back operating status, cleaning progress, battery information, fault alarms and coverage statistics. The communication module has an command cache queue and a priority scheduling mechanism. When multiple commands are received at the same time, safety commands such as emergency stop and steering are executed first. The main control module is also configured to switch between remote control mode and autonomous operation mode using a smooth transition logic.

[0022] Due to the adoption of the above technical solution, this application has at least one of the following beneficial effects compared with the prior art: 1. By using multi-sensor fusion to perceive the actual type and degree of dirt on the surface being cleaned, and based on the identification results, the roller brush speed, vacuum fan power and travel speed are adjusted synchronously through a collaborative control algorithm. This achieves an upgrade in cleaning logic from an open-loop fixed mode to a closed-loop adaptive adjustment mode, improving cleaning effect and resource utilization.

[0023] 2. By recording the location of partially uncovered areas caused by obstacle avoidance and automatically inserting a compensation cleaning path for secondary coverage in subsequent path planning, the contradiction between obstacle avoidance and cleaning coverage is resolved, effectively reducing the probability of missed sweeps and improving the overall cleaning coverage rate.

[0024] 3. By using feature dot product similarity metric combined with L∞ norm distance calculation, the scores of unknown categories have obvious statistical outlier characteristics, thus improving the accuracy of excluding non-target categories. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. Wherein: Figure 1 This is a flowchart illustrating an automated cleaning control method for a self-propelled cleaning robot provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only for explaining this application and not for limiting it. Furthermore, it should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all structures. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0027] The terms "first," "second," etc., used in this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] Existing photovoltaic cleaning robot cleaning control schemes typically employ an open-loop fixed mode, relying on preset cleaning parameters without fully considering real-time perception and adaptive adjustment of the type and degree of dirt on the surface being cleaned. Simultaneously, traditional obstacle avoidance schemes lack mechanisms for recording and compensating for blind spots, resulting in a trade-off between cleaning coverage and obstacle avoidance safety. Furthermore, existing deep learning-based stain recognition schemes consume significant computational resources, making them difficult to deploy on low-computing-power edge control platforms, exhibiting low recognition accuracy under small sample conditions, and facing difficulties in expanding to new categories.

[0030] In view of this, this application provides an automated cleaning control method for a self-propelled cleaning robot. By introducing a lightweight stain recognition algorithm based on multi-domain feature enhancement and SimAM attention mechanism fusion, an obstacle avoidance blind spot recording and compensation path planning mechanism, and a cleaning parameter linkage adjustment strategy based on a cooperative control algorithm, the method achieves an intelligent upgrade of the cleaning robot from simply executing actions to understanding cleaning. Figure 1 As shown, Figure 1 A flowchart illustrating an automated cleaning control method for a self-propelled cleaning robot according to an embodiment of this application includes the following steps: Step S1: Start the robot and perform a system self-test. The system self-test includes power status detection, sensor initialization, motor drive status detection, and cleaning execution module linkage test. After confirming that the system is normal, it enters the autonomous operation mode. The power status detection includes checking whether the battery voltage is within the normal range and whether the remaining power is sufficient for the operation. The sensor initialization includes self-testing and calibration of the boundary distance sensor, obstacle detection sensor, dust concentration sensor, image acquisition unit, and light intensity detection unit. The motor drive status detection includes detection of the drive circuits and encoder feedback of the left travel motor, right travel motor, roller brush motor, and vacuum fan. The cleaning execution module linkage test includes confirming that the roller brush mechanism's lifting, rotation, and vacuum fan start / stop functions are normal. After confirming that all states are normal, the robot enters the autonomous operation mode. If the self-test detects an abnormality, the robot sends a fault alarm to the maintenance terminal through the communication module, waiting for manual intervention.

[0031] Step S2: In autonomous operation mode, the robot travels along a preset path, collecting boundary distance data and obstacle information at fixed intervals. This data is then fused with historical window data from multiple sampling periods to identify the boundaries of the work area. In this embodiment, the robot employs a tracked mobile structure, and the main controller is built on an STM32F103 series MCU. The environmental perception module includes a ranging sensor, an obstacle detection sensor, a dust concentration sensor, an image acquisition unit, and a light intensity detection unit. The ranging sensor is used for boundary distance detection, and the obstacle detection sensor is used for obstacle detection.

[0032] The environmental perception module collects boundary distance data and obstacle information at fixed intervals; the main control module combines historical window data containing data from multiple recent sampling periods for fusion judgment. Specifically, a sliding window filtering method is used: the mean of the data within the historical window is calculated, and the fused boundary distance value at the current moment is obtained by weighting the current measurement value with the mean of the historical window; when the fused boundary distance value is less than a preset boundary threshold, it is determined that the robot has approached the boundary of the work area.

[0033] The introduction of historical window data effectively suppresses transient noise interference caused by factors such as reflections from the photovoltaic panel surface, water droplets, and dust obstruction, thus avoiding boundary misjudgments. For example, when water droplets briefly obstruct the ranging sensor, a single ranging value may jump, but the mean filtering within the historical window can effectively smooth this jump, preventing the robot from turning prematurely due to transient misjudgments.

[0034] In one embodiment, step S2 further includes performing a mileage estimation task; the acquisition of boundary distance data and the mileage estimation task are executed in parallel through multi-threading or interrupt scheduling; historical window data is used to suppress transient noise interference.

[0035] The robot performs the task of odometer estimation. The wheeled odometer obtains the rotation angle of the walking wheels through a high-resolution encoder and calculates the theoretical walking distance by combining it with the wheel diameter; the inertial measurement unit provides heading angle and acceleration information; the data from the wheeled odometer and the inertial measurement unit are fused through extended Kalman filtering to suppress the cumulative error caused by wheel slippage at the edge of the photovoltaic panel and the gaps between the arrays.

[0036] There is no strict sequential dependency between the acquisition of boundary distance data and the odometry estimation task. By using multi-threading or interrupt scheduling to execute in parallel, the time consumed in a single control cycle is reduced, and the response speed of the robot in high-speed travel or complex terrain is improved.

[0037] Step S3: When the boundary condition is determined to be reached based on the fusion judgment result, the deceleration control and cleaning execution modules are executed to reduce speed synchronously, and the path direction is replanned after the vehicle stops safely at the edge. In this embodiment, the boundary condition is determined to be reached when the fusion judgment result is less than the preset boundary threshold and continues to exceed the preset time.

[0038] The main control module immediately executes deceleration control: the walking motor gradually decelerates from its current speed to zero. At the same time, the cleaning execution module synchronously reduces its speed: the roller brush speed is reduced synchronously, and the power of the vacuum fan is reduced to standby power, in order to avoid scratching damage to the edges of the photovoltaic modules caused by the high-speed rotation of the cleaning mechanism when stopping at the edge.

[0039] After the robot comes to a complete stop, it performs path replanning: at the current position, it rotates 180° in the direction of travel, enters the next adjacent cleaning path, and prepares to start reverse cleaning.

[0040] Step S4: When it is determined that an obstacle avoidance action has occurred based on obstacle information and that a local area is not covered, the location of the local uncovered area is recorded, and a compensation cleaning path is inserted in the subsequent path planning to cover the local uncovered area a second time; In this embodiment, when the robot encounters obstacles such as pressure blocks, jumper cables, and support components on the frame of the photovoltaic module during operation, the obstacle detection sensor of the environmental perception module triggers an obstacle avoidance signal.

[0041] After the main control module detects the obstacle avoidance signal, it records the current position coordinates of the robot. During the obstacle avoidance process, the robot performs a brush lifting action, and the roller brush mechanism is raised to cross the obstacle. Due to the cleaning mechanism detaching from the surface being cleaned, some uncovered areas are formed on the brush lifting path.

[0042] Unlike traditional solutions that directly ignore obstacle avoidance blind spots, in step S4, the main control module records the location and extent of locally uncovered areas caused by obstacle avoidance and automatically inserts a compensation cleaning path in subsequent path planning. Specifically: After the obstacle avoidance maneuver is completed, the main control module calculates the geometric range of the uncovered area based on the robot's current heading and the brush lifting time. This geometric range is a rectangular area centered on the obstacle avoidance position, with a width equal to the width of the brush and a length equal to the distance the robot travels during the brush lifting duration. The coordinate information of this area is stored in the compensation cleaning task queue.

[0043] After the robot completes the main cleaning task of the current path, the main control module reads the coordinate information of all areas to be compensated from the compensation cleaning task queue, inserts a dedicated compensation cleaning path into the subsequent path planning, and performs secondary cleaning on each uncovered area. During compensation cleaning, the robot passes through the area at a low speed to ensure cleaning effectiveness.

[0044] Step S5: During autonomous operation, environmental information of the surface being cleaned is collected through sensors. Based on this environmental information, the type and extent of stains are identified, and cleaning parameters are dynamically adjusted according to the identification results. Identifying the type and extent of stains based on environmental information specifically includes: The system acquires image data of the surface to be cleaned. In direct sunlight conditions at a photovoltaic power station, the main control module prioritizes dust concentration sensors for stain detection, supplemented by image acquisition; in low-light or shaded areas, visual recognition is the primary method. Lighting conditions are monitored in real-time by the light intensity detection unit.

[0045] Multi-domain feature enhancement is performed on the image data to obtain spatial domain features, gradient features, and frequency domain features. Frequency domain features are obtained through Fast Fourier Transform (FFT). Spatial domain features include texture features extracted directly from the image pixel grayscale values, such as gray-level co-occurrence matrix (GLCM) features and local binary pattern histogram (LCM) features. Gradient features are calculated using edge detection operators to determine the magnitude and direction of image gradients, capturing abrupt grayscale changes in the edge regions of stains. For highly adhesive stains such as bird droppings, the gradient magnitude in the edge regions is typically significantly higher than that in dusty areas. Frequency domain features are transformed from the spatial domain to the frequency domain using FFT.

[0046] In the frequency domain, different types of stains exhibit different frequency energy distribution characteristics. For example, a uniform thin layer of dust forming on the surface of a photovoltaic panel corresponds to low-frequency enhancement components; while highly adhesive stains such as bird droppings exhibit specific mid-to-high frequency texture features in the frequency domain. Obtaining frequency domain features through FFT transformation effectively compensates for the insufficient contrast of spatial domain features under strong reflective conditions.

[0047] A SimAM attention mechanism is employed to fuse spatial domain features, gradient features, and frequency domain features. Attention weights are obtained by calculating the spatial domain amplitude energy and the energy response after frequency domain transformation. These attention weights are then used to enhance the features, resulting in enhanced features. SimAM calculates the attention weights by calculating the spatial domain amplitude energy and the frequency domain energy. The enhanced features are the product of the input features and the attention weights. This fusion strategy achieves a reasonable integration of spatial and frequency domain features without increasing the number of model parameters, effectively highlighting the key discriminative features of the stained area while suppressing interference from photovoltaic panel surface texture noise and background noise from uneven illumination.

[0048] Stain identification and classification based on enhanced features, including: Feature extraction for enhancement is performed using the ResNet50 backbone network, and global average pooling is used at the end of the backbone network to replace the fully connected layer for data compression. Global average pooling at the end of the ResNet50 backbone network replaces the traditional fully connected layer for data compression. The global average pooling operation condenses the two-dimensional feature map of each channel into a single mean statistic, and the output feature vector dimension is equal to the number of channels, i.e., 2048 dimensions at the end of ResNet50, rather than the tens of thousands of parameters of the traditional fully connected layer.

[0049] The similarity score is calculated by multiplying the feature vectors after global average pooling with the reference feature vectors of each category of stains in the reference feature matrix. The feature vector distance is calculated based on the L∞ norm to prevent the curse of high dimensionality. The stain category is determined based on the similarity score and the feature vector distance.

[0050] The similarity score is calculated by multiplying the feature vector after global average pooling with the reference feature vector of each category of stain in the reference feature matrix. The closer the score is to 1, the more similar the sample to be identified is to the reference feature of that category.

[0051] The L∞ norm is used to calculate the distance between eigenvectors to prevent the curse of high dimensionality. The L∞ norm focuses on the maximum difference in each dimension of the eigenvector. Compared with Euclidean distance, it can better preserve the outlier information of features in high-dimensional space and effectively prevent the problem of the distance between points tending to be equal in high-dimensional space.

[0052] The stain category is determined based on the similarity score and the feature vector distance. If the similarity score is greater than a preset threshold and the feature vector distance is less than a preset threshold, it is determined to be the corresponding stain category; otherwise, it is determined to be an unknown category, triggering an incremental learning process.

[0053] The traditional Softmax method requires the sum of the output probabilities of all categories to be 1, and will give a category determination even for completely unknown inputs. The similarity measure plus L∞ norm distance scheme adopted in this application makes the similarity scores of unknown category samples generally low, with obvious statistical outlier characteristics, so as to accurately identify unknown categories.

[0054] It also includes an incremental learning step: when the classification result is incorrect or a new category appears, only the ResNet50 skeleton network is used to extract features from the corresponding image samples to obtain a new feature vector, and the new feature vector is appended to the reference feature matrix without retraining the entire deep learning network.

[0055] In this embodiment, when the classification result is incorrect or a new category appears, the incremental learning process is initiated: New feature vectors are obtained by extracting features from the corresponding image samples using only the ResNet50 skeleton network, and these new feature vectors are directly appended to the reference feature matrix, thus updating the number of categories accordingly.

[0056] This process eliminates the need to retrain the entire deep learning network, perform backpropagation and gradient updates, or adjust any network weight parameters. The deployment time for adding a new category is reduced from several hours required for retraining to seconds for feature appending, significantly reducing the time cost of later maintenance.

[0057] In step S5, the cleaning parameters include the roller brush speed, the vacuum fan power, and the travel speed. These parameters are adjusted in conjunction with a collaborative control algorithm. The algorithm synchronously adjusts the roller brush speed, vacuum fan power, and travel speed based on the stain recognition results, ensuring a dynamic balance between cleaning intensity and travel efficiency. Based on the stain recognition results, the main control module adjusts the cleaning parameters in conjunction with the collaborative control algorithm. The cleaning parameters include the roller brush speed, vacuum fan power, and travel speed. The collaborative control algorithm synchronously adjusts these three parameters based on the stain recognition results.

[0058] The method is applied to the surface cleaning of photovoltaic modules, and the robot is a tracked mobile structure; the cleaning parameters are dynamically adjusted, including: When the identification result is dust, the cleaning execution module is controlled to operate at the first power level and the robot is controlled to move at the first speed; thus achieving rapid passage and energy saving.

[0059] When the identification result is an adhesive stain, the cleaning execution module is controlled to run at the second power level and the robot is controlled to move at the second speed; to achieve deep cleaning, the cleaning action time is extended by moving at a low speed, and the mechanical peeling force of the high-speed roller brush and the suction force of the high-suction fan are combined to remove stubborn stains.

[0060] The first power level is lower than the second power level, and the first speed is higher than the second speed. The cleaning parameters are adjusted using a collaborative control algorithm, meaning that the three parameters are synchronized rather than adjusted independently in steps, ensuring a dynamic balance between cleaning power and travel efficiency.

[0061] This application also provides an automated cleaning control system for a self-propelled cleaning robot, including: The motion control module is used to control the robot to move along a preset path and to perform deceleration control, edge safety stopping, and path replanning when boundary conditions are determined to be reached. The environmental perception module is used to collect boundary distance data, obstacle information, and stain information on the surface being cleaned at fixed intervals. The cleaning execution module, including a roller brush mechanism and a vacuum fan, is used to perform cleaning operations; The main control module is connected to the motion control module, the environmental perception module, and the cleaning execution module, respectively, and is used to dynamically adjust the cleaning parameters of the cleaning execution module based on the data collected by the environmental perception module. The main control module is also used to record the uncovered areas caused by obstacle avoidance when it is determined that an obstacle avoidance action has occurred based on obstacle information and a local area is not covered, and to insert a compensation cleaning path in the subsequent path planning.

[0062] The main control module is built on the STM32F103 series MCU and is a low-computing-power edge control unit; the environmental perception module includes a dust concentration sensor for detecting dust concentration, an image acquisition unit for collecting image data, and a light intensity detection unit; the main control module is configured to: prioritize the dust concentration sensor when the light intensity is higher than the preset threshold, and prioritize visual recognition when the light intensity is lower than the preset threshold or in the shadow area.

[0063] It also includes a communication module, which receives remote control commands and transmits back operating status, cleaning progress, battery information, fault alarms and coverage statistics. The communication module has a command cache queue and priority scheduling mechanism. When multiple commands are received at the same time, safety commands such as emergency stop and steering are executed first. The main control module is also configured to switch between remote control mode and autonomous operation mode using smooth transition logic.

[0064] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0065] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0066] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0067] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. An automated cleaning control method for a self-propelled cleaning robot, characterized in that, Includes the following steps: Step S1: Start the robot and perform a system self-test. The system self-test includes power status detection, sensor initialization, motor drive status detection, and cleaning execution module linkage test. After confirming that the system is normal, it enters the autonomous operation mode. Step S2: In autonomous operation mode, the robot travels along a preset path, collects boundary distance data and obstacle information at fixed intervals, and combines it with historical window data containing multiple sampling periods for fusion judgment to identify the boundary of the operation area; Step S3: When the boundary conditions are determined to be met based on the fusion judgment result, the deceleration control and cleaning execution modules are executed to reduce speed synchronously, and the path direction is replanned after the vehicle safely stops at the edge. Step S4: When it is determined that an obstacle avoidance action has occurred based on the obstacle information and a local area is not covered, the location of the local uncovered area is recorded, and a compensation cleaning path is inserted in the subsequent path planning to cover the local uncovered area a second time. Step S5: During the autonomous operation, the environmental information of the surface being cleaned is collected by sensors, the type and degree of stains are identified based on the environmental information, and the cleaning parameters are dynamically adjusted according to the identification results.

2. The method according to claim 1, characterized in that, Step S5, which involves identifying the type and extent of stains based on the environmental information, specifically includes: Acquire image data of the surface to be cleaned; Multi-domain feature enhancement is performed on the image data to obtain spatial domain features, gradient features, and frequency domain features, wherein the frequency domain features are obtained through fast Fourier transform; The SimAM attention mechanism is used to fuse the spatial domain features, gradient features and frequency domain features. Attention weights are obtained by calculating the spatial domain amplitude energy and the energy response after frequency domain transformation. The attention weights are then used to enhance the features to obtain enhanced features. Stain identification and classification are performed based on the enhanced features.

3. The method according to claim 2, characterized in that, The stain identification and classification based on the enhanced features includes: The enhanced features are extracted using a ResNet50 backbone network, and global average pooling is used at the end of the backbone network to replace the fully connected layer for data compression. The similarity score is calculated by multiplying the feature vector after global average pooling with the reference feature vector of each category of stain in the reference feature matrix, and the feature vector distance is calculated based on the L∞ norm to prevent the curse of high dimensionality. The stain category is determined according to the similarity score and the feature vector distance.

4. The method according to claim 3, characterized in that, It also includes an incremental learning step: when the classification result is incorrect or a new category appears, the new feature vector is obtained by extracting features from the corresponding image sample through the ResNet50 skeleton network, and the new feature vector is appended to the reference feature matrix without retraining the entire deep learning network.

5. The method according to claim 1, characterized in that, The cleaning parameters in step S5 include the roller brush speed, the vacuum fan power, and the travel speed. The cleaning parameters are adjusted in conjunction with a collaborative control algorithm. The collaborative control algorithm adjusts the roller brush speed, the vacuum fan power, and the travel speed synchronously according to the stain recognition results to ensure a dynamic balance between cleaning intensity and travel efficiency.

6. The method according to claim 1, characterized in that, Step S2 further includes performing a mileage estimation task; the acquisition of the boundary distance data and the mileage estimation task are executed in parallel through multi-threading or interrupt scheduling; the historical window data is used to suppress transient noise interference.

7. The method according to any one of claims 1 to 6, characterized in that, The method is applied to the surface cleaning of photovoltaic modules, and the robot is a tracked mobile structure; the dynamic adjustment of cleaning parameters includes: When the identification result is floating dust, the cleaning execution module is controlled to run at the first power level and the robot is controlled to move at the first speed. When the identification result is an adhesive stain, the cleaning execution module is controlled to run at the second power level and the robot is controlled to move at the second speed. Wherein, the first power level is lower than the second power level, and the first speed is greater than the second speed.

8. An automated cleaning control system for a self-propelled cleaning robot, characterized in that, include: The motion control module is used to control the robot to move along a preset path and to perform deceleration control, edge safety stopping, and path replanning when boundary conditions are determined to be reached. The environmental perception module is used to collect boundary distance data, obstacle information, and stain information on the surface being cleaned at fixed intervals. The cleaning execution module, including a roller brush mechanism and a vacuum fan, is used to perform cleaning operations; The main control module is connected to the motion control module, the environmental perception module, and the cleaning execution module, respectively, and is used to dynamically adjust the cleaning parameters of the cleaning execution module according to the data collected by the environmental perception module. The main control module is also used to record the partially uncovered area caused by obstacle avoidance when it is determined that an obstacle avoidance action has occurred based on obstacle information and the local area is not covered, and to insert a compensation cleaning path in the subsequent path planning.

9. The system according to claim 8, characterized in that, The main control module is built on the STM32F103 series MCU and is a low-computing-power edge control unit; the environmental perception module includes a dust concentration sensor for detecting dust concentration, an image acquisition unit for acquiring image data, and a light intensity detection unit; the main control module is configured to: prioritize the dust concentration sensor when the light intensity is higher than a preset threshold, and prioritize visual recognition when the light intensity is lower than the preset threshold or in a shadow area.

10. The system according to claim 8, characterized in that, It also includes a communication module for receiving remote control commands and transmitting back operating status, cleaning progress, battery information, fault alarms, and coverage statistics. The communication module has an command cache queue and a priority scheduling mechanism. When multiple commands are received simultaneously, safety commands such as emergency stop and steering are executed first. The main control module is also configured to switch between remote control mode and autonomous operation mode using a smooth transition logic.

Images