Photovoltaic cleaning robot path planning method and system based on pollution degree

Abstract

The application provides a photovoltaic cleaning robot path planning method and system based on pollution degree, and relates to the field of robot path planning. The method comprises: collecting the reflectivity change event stream of the surface of the photovoltaic module, and generating an asynchronous data packet containing event coordinates, a timestamp and polarity. The asynchronous data packet is converted into a dynamic pollution level map, and the pollution level value is positively correlated with the event occurrence frequency per unit time. A distributed dynamic pollution map is constructed by using a pulse neural network, and the pollution diffusion and cleaning recovery process is simulated by neuron membrane potential evolution. An action selection competition mechanism is established in the pulse neural network, and cleaning path instructions are generated according to the spatio-temporal distribution characteristics of the dynamic pollution map. The cleaning action is executed and the environmental change event is fed back, and the dynamic pollution map and the action selection competition mechanism are dynamically updated. The technical problems of high energy consumption, slow response to sudden pollution and dynamic environment, poor adaptability and weak anti-interference ability are solved.

Description

Technical Field

[0001] This application relates to the field of robot path planning, and in particular to a method and system for path planning of photovoltaic cleaning robots based on the degree of contamination. Background Technology

[0002] As photovoltaic power plants expand to larger, more centralized locations and in more complex terrains, the problem of power generation efficiency degradation caused by surface contamination of modules is becoming increasingly prominent. Modules that are not cleaned in time can suffer an annual power generation efficiency loss of up to 30%. Stubborn contamination such as bird droppings can easily trigger hot spot effects, causing permanent damage. Cleaning robots have become an essential piece of equipment for the intelligent operation and maintenance of power plants. Most existing photovoltaic cleaning robots rely on traditional vision sensors such as CCD / CMOS, using periodic full-frame image acquisition to detect contamination. This requires processing a large amount of redundant image data, resulting in high energy consumption. Furthermore, the failure rate can exceed 15% in harsh environments such as high temperatures and high dust levels. The full-frame image transmission and processing process causes a delay of up to seconds in contamination perception, making it impossible to detect sudden contamination such as bird droppings in time, affecting the timeliness of cleaning. At the same time, when planning cleaning routes, static grid maps or preset path planning are used based on the acquired data, resulting in a "one-size-fits-all" cleaning strategy that is difficult to target highly contaminated areas. This leads to waste of water and energy resources and an inability to cope with the dynamic changes in contamination distribution in complex scenarios. Summary of the Invention

[0003] This application provides a path planning method and system for photovoltaic cleaning robots based on the degree of pollution, which solves the technical problems of existing technologies that rely on full-frame synchronous acquisition of traditional vision sensors, static map modeling and complex global path planning algorithms, resulting in high energy consumption, slow response to sudden pollution and dynamic environment, poor adaptability and weak anti-interference ability during use.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a path planning method for photovoltaic cleaning robots based on contamination levels includes: collecting event streams of reflectivity changes on the surface of photovoltaic modules and generating asynchronous data packets containing event coordinates, timestamps, and polarities. These asynchronous data packets are then converted into a dynamic contamination level map, where the contamination level value is positively correlated with the frequency of events occurring per unit time. A distributed dynamic contamination map is constructed using a spiking neural network, and the contamination diffusion and cleaning recovery process is simulated through the evolution of neuronal membrane potentials. An action selection competition mechanism is established within the spiking neural network to generate cleaning path instructions based on the spatiotemporal distribution characteristics of the dynamic contamination map. Cleaning actions are executed, and environmental change events are fed back, dynamically updating the dynamic contamination map and the action selection competition mechanism.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the construction process of the spiking neural network includes: constructing a spiking neural network structure comprising an event input layer, a contamination processing layer, and an action output layer. The event input layer consists of multiple event response neurons, each corresponding to a specific coordinate position of the photovoltaic module, used to receive asynchronous data packets from the event camera and generate pulse signals. The contamination processing layer is composed of a dynamic contamination simulation neural network, where neurons simulate the accumulation and diffusion of contamination through membrane potential evolution. The action output layer includes competitive action neurons, each encoding a basic cleaning action, and selecting the optimal action through a pulse competition mechanism. This competition mechanism is based on the dynamic evolution of membrane potential; when the membrane potential of a certain action neuron first reaches the firing threshold, the corresponding cleaning instruction is triggered. A dedicated neuromorphic computing chip is integrated to process the pulse stream in parallel and embed a dynamic reconfiguration interface to adapt to the spatial mapping of photovoltaic arrays of different sizes. The accuracy of the contamination simulation is optimized by adjusting the connection weights between neurons through pulse timing.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves constructing a distributed dynamic pollution map using a spiking neural network and simulating the pollution diffusion and cleaning recovery process through neuron membrane potential evolution. This includes: mapping each event response neuron in the event input layer to the spatial coordinates of the photovoltaic module surface based on the physical dimensions of the photovoltaic module, establishing a fixed mapping relationship between physical coordinates and virtual neurons, and forming a lookup table from physical coordinates to neurons. Asynchronous data packets are input into the spiking neural network, and the corresponding event response neuron is located based on the event coordinates. The membrane potential increment is calculated and used to update the membrane potential of the pollution treatment layer. The reflectivity of the photovoltaic module after cleaning is monitored, a reset event is generated and sent to the pollution treatment layer, updating the dynamic pollution simulation neuron potential at the corresponding coordinates in the pollution treatment layer. The neural network in the pollution treatment layer is extracted to form a pollution state distribution map.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves establishing an action selection competition mechanism in a spiking neural network to generate cleaning path instructions based on the spatiotemporal distribution characteristics of a dynamic pollution map. This process includes: constructing a spatiotemporal feature encoding layer, which includes feature extraction neurons connected to the neuron cluster in the pollution treatment layer. These neurons receive membrane potential pulse streams and calculate pollution accumulation intensity and pollution diffusion trend feature values. The pollution accumulation intensity and pollution diffusion trend feature values ​​are then converted into excitation pulse sequences for different action neurons in the action neuron array according to a preset action mapping rule. All action neurons in the action neuron array compete by firing parallel asynchronous pulses based on their current membrane potential, forming cleaning path instructions.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves collecting a stream of reflectivity change events on the surface of photovoltaic modules and generating asynchronous data packets containing event coordinates, timestamps, and polarities. This process includes: deploying a dynamic vision sensor operating in an asynchronous mode, where each pixel unit independently responds to changes in light intensity. When the logarithmic change in light intensity in a local area exceeds a preset intensity threshold, event acquisition is triggered to obtain the reflectivity change event stream. The reflectivity change event stream is then filtered in real-time to remove globally homogeneous events caused by ambient light fluctuations, generating a pre-processed event stream. A mapping relationship is established between the physical coordinate system of the photovoltaic modules and the pixel coordinate system of the event camera, and pollution events are dynamically prioritized and weighted according to event polarity parameters. The weighted pre-processed event stream is then sorted by priority and output to form an asynchronous data packet sequence for tracking pollution hotspots.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves converting asynchronous data packets into a dynamic pollution level map, which includes: parsing the asynchronous data packets and calculating the pollution level to establish a pollution level mapping rule; using a sliding time window to perform frame processing on the asynchronous data packets and extracting the event frequency for each spatial coordinate; and integrating an FPGA parallel computing platform to fuse historical pollution accumulation values ​​with real-time event frequencies to generate a dynamic pollution level map.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the process of dynamically updating the dynamic contamination map and the action selection competition mechanism includes: real-time monitoring of changes in light intensity in the cleaning area to trigger the generation of reset events, and, based on the polarity parameters of the reset events, performing membrane potential reset operations on the dynamic contamination simulation neurons corresponding to the coordinates of the contamination treatment layer. Based on the frequency changes of contamination events before and after cleaning, adjusting the connection weights between neurons using pulse time-dependent plasticity rules. Based on the cleaning action execution effect, dynamically adjusting the inhibition weight coefficients in the competition mechanism.

[0011] Secondly, a path planning system for a photovoltaic cleaning robot based on contamination levels is provided, comprising an event perception module, a pulse processing module, and an execution module. The event perception module collects and preprocesses the independent responses of each pixel unit to changes in light intensity, generating asynchronous data packet sequences and establishing a mapping between the physical coordinate system of the photovoltaic module and the sensor pixel coordinate system, assigning dynamic priorities to contamination events. The pulse processing module constructs a three-level spiking neural network, outputting a distributed dynamic contamination map and an excitation pulse sequence for an action selection competition mechanism through the asynchronous data packet sequences. The execution module performs cleaning actions based on the excitation pulse sequence of the action selection competition mechanism and executes state feedback and environmental change event streams.

[0012] In conjunction with the second aspect mentioned above, in one possible implementation, the three-level spiking neural network includes an event input layer, a contamination processing layer, and an action output layer: The event input layer locates the neuron corresponding to the event through a fixed mapping relationship from physical coordinates to virtual neurons, calculates the membrane potential increment, and drives the contamination processing layer to update the membrane potential. The contamination processing layer calculates the potential increments of positive and negative events to generate a dynamic contamination level map. The action output layer selects the optimal action through a spiking competition mechanism; when the membrane potential of a certain action neuron reaches the threshold first, a cleaning instruction is triggered.

[0013] In conjunction with the second aspect mentioned above, one possible implementation involves constructing the event-aware module, which includes: deploying a dynamic visual sensor array, where each pixel unit independently responds to changes in light intensity; establishing a dynamic mapping relationship between the physical coordinate system of the photovoltaic module and the sensor pixel coordinate system, accelerating coordinate transformation via an FPGA; and integrating an event filter, employing a sliding time window and polarity weighting strategy.

[0014] This application provides a path planning method and system for photovoltaic cleaning robots based on contamination levels. By using an event camera in an asynchronous acquisition mode, it captures only localized events of abrupt changes in the reflectivity of photovoltaic module surfaces and generates data packets containing coordinates, timestamps, and polarity. This eliminates the redundant acquisition of traditional full-frame images, significantly reducing energy consumption. Simultaneously, it achieves instantaneous capture of sudden contamination with microsecond-level response, solving the problem of high latency in traditional sensing. Furthermore, based on a neural network-constructed neuronal map, it accurately simulates the contamination diffusion and cleaning recovery process through the dynamic evolution of membrane potential charging, leakage, and reset, replacing static map modeling and solving the problem of poor adaptability of traditional planning to dynamic environments. The event-driven nature and action competition mechanism of spiking neural networks allow the optimal cleaning path to emerge naturally through neuronal potential competition, eliminating the need for complex global calculations and achieving millisecond-level decision response. This further solves the core problems of high latency and weak dynamic adaptation in traditional serial planning, enabling the overall system to effectively form a closed-loop progressive architecture of "perception, transformation, modeling, decision-making, and feedback." This not only addresses the shortcomings of traditional planning's one-time decision-making, which cannot be adjusted, but also achieves decentralized work area division through the suppression and coordination of pulse signals between multiple machines, avoiding overlapping operations. At the same time, global event flow monitoring distinguishes between light and shadow changes and local pollution signals, improving anti-interference capabilities. Ultimately, this achieves high-efficiency photovoltaic cleaning with low energy consumption, fast response, high adaptability, and strong collaboration, improving the cleaning efficiency and power generation benefits of photovoltaic modules.

[0015] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0016] Figure 1 A system architecture diagram of the photovoltaic cleaning robot path planning method based on contamination level provided in the embodiments of this application; Figure 2 A flowchart illustrating the path planning method for a photovoltaic cleaning robot based on the degree of contamination provided in this application embodiment; Figure 3 A flowchart illustrating the path planning method for a photovoltaic cleaning robot based on the degree of contamination provided in this application embodiment; Figure 4 A flowchart illustrating the path planning method for a photovoltaic cleaning robot based on the degree of contamination provided in this application embodiment; Figure 5 A flowchart illustrating the path planning method for a photovoltaic cleaning robot based on the degree of contamination provided in this application embodiment; Figure 6 A flowchart illustrating the path planning method for a photovoltaic cleaning robot based on the degree of contamination provided in this application embodiment; Figure 7 A flowchart illustrating the path planning method for a photovoltaic cleaning robot based on the degree of contamination provided in this application embodiment; Figure 8 This is a schematic diagram of the path planning system for a photovoltaic cleaning robot based on the degree of contamination provided in an embodiment of this application. Detailed Implementation

[0017] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0018] like Figure 1 As shown in the embodiments of this application, the photovoltaic cleaning robot path planning method based on the degree of contamination provides the following: Step 101: Collect the event stream of reflectivity change on the surface of the photovoltaic module and generate an asynchronous data packet containing event coordinates, timestamps and polarity.

[0019] In this context, the photovoltaic module surface refers to the exposed surface of the photovoltaic panel, and its reflective properties can change due to the accumulation or removal of contaminants. The reflectivity change event stream is a sequence of asynchronous events triggered by instantaneous changes in surface reflectivity; each event represents a sudden change in local light intensity. Event coordinates are the spatial location identifiers of events within the sensor array, typically represented as two-dimensional pixel coordinates (x, y). Polarity indicates the direction of brightness change; positive polarity corresponds to increased brightness, and negative polarity corresponds to decreased brightness.

[0020] In some implementations, dynamic vision sensors continuously monitor changes in the reflectivity of photovoltaic module surfaces. When the light intensity in a local area undergoes a logarithmic change and exceeds a preset threshold, the sensor immediately triggers event acquisition. Each independent event is then encapsulated in real-time into an asynchronous data packet. Each data packet precisely records the pixel coordinates (x, y) of the event, a microsecond-level timestamp (t), and the polarity parameter of the brightness change, ultimately forming a continuous event stream output. The entire process employs a completely asynchronous working mechanism, with each pixel responding independently to changes, eliminating the need for global synchronous sampling. This enables instantaneous capture of surface contamination events and rapid data generation.

[0021] It should be noted that the asynchronous data packet generation mechanism can reduce system latency and energy consumption because the event camera only reports change information rather than the full-frame image, avoiding data redundancy caused by traditional periodic acquisition. During its use, the sensor threshold setting can be adaptively adjusted according to the light fluctuations in the installation environment of different photovoltaic panels.

[0022] Step 102: Convert asynchronous data packets into a dynamic pollution level graph. The pollution level value is positively correlated with the frequency of events occurring per unit time.

[0023] The dynamic pollution level map is a spatial distribution map that represents the real-time evolution of pollution levels in various regions in numerical form. The pollution level value is a scale value that quantifies the pollution level in the dynamic pollution level map.

[0024] In some implementations, as the system continuously receives asynchronous data packet streams from the event camera, each packet can be directly parsed to extract the event coordinates, microsecond-level timestamps, and polarity information. Then, a sliding time window (e.g., 100 milliseconds) is set for each spatial location, and the event frequency within this window is counted in real time. The frequency values ​​can then be directly converted into pollution level values ​​using a linear or nonlinear mapping function. This ensures that high-frequency areas correspond to high pollution levels, and low-frequency areas correspond to low pollution levels. Finally, the level values ​​for all locations are integrated into a dynamically updated grid map, forming a dynamic pollution level map that reflects the spatial distribution of pollution in real time.

[0025] It should be noted that the frequency mapping mechanism requires only simple statistics to achieve low-latency conversion during use, effectively avoiding complex image processing algorithms. In practical application, its time window size can be adaptively adjusted directly according to the pollution diffusion rate, thus balancing response speed and stability. Simultaneously, the positive correlation ensures rapid identification of pollution hotspots, and during practical application, it can be directly combined with polarity filtering optimization to avoid interference from environmental noise events.

[0026] Step 103: Construct a distributed dynamic pollution map using a spiking neural network, and simulate the pollution diffusion and cleanup process through the evolution of neuronal membrane potential.

[0027] Among them, the spiking neural network is a computational model simulating biological neuron information processing, using the spatiotemporal pattern of pulse sequences for asynchronous event-driven computation. The distributed dynamic pollution map is a virtual spatial network composed of multiple neurons, each corresponding to a specific location on the photovoltaic surface, collectively forming a dynamically evolving representation of pollution distribution. Neuron membrane potential evolution refers to the dynamic change process of the potential within the neuron, mainly simulating pollution accumulation through potential increase, diffusion through potential leakage, and cleaning recovery through potential reset. Pollution diffusion is the physical process of pollutants spreading across the photovoltaic surface over time. The cleaning recovery process is the gradual return of surface reflectivity to normal after cleaning.

[0028] In some implementations, a distributed neural network is initialized based on a spiking neural network architecture, where each neuron corresponds one-to-one with the spatial coordinates of a photovoltaic module, forming a virtual pollution map skeleton. Asynchronous data packets from event cameras can then be directly converted into input pulse streams, causing each event within the network to trigger an increase in the membrane potential of the corresponding neuron. This potential value directly maps to the pollution level at that point. The pollution diffusion process can then be simulated directly through the natural leakage mechanism of membrane potential, allowing high-potential neurons to transmit some potential to adjacent neurons, thus achieving spatial spread modeling of pollution. Simultaneously, when the cleaning robot performs cleaning actions, the membrane potential of neurons in the corresponding area is actively reset or rapidly decays, and the collector can simulate the cleaning recovery effect. Ultimately, the entire network continuously evolves, forming a distributed map that reflects the real-time dynamics of pollution.

[0029] Step 104: Establish an action selection competition mechanism in the spiking neural network to generate cleaning path instructions based on the spatiotemporal distribution characteristics of the dynamic pollution map.

[0030] The action selection competition mechanism involves action neurons that receive excitatory and inhibitory pulses, enabling dynamic competition of membrane potentials to reach a firing threshold. Spatiotemporal distribution characteristics refer to the evolutionary patterns of the contamination map in spatial coordinates and time dimensions, such as the aggregation and diffusion of high-potential areas. Cleaning path instructions are sequences of commands controlling the robot's movement and cleaning actions, directly generated by the output pulses of the action neurons.

[0031] In some implementations, a dedicated set of action neurons is initialized in a spiking neural network, with each neuron encoding a basic robot action (such as forward, left turn, right turn, or cleaning), forming an action selection layer. This allows high-potential neurons (corresponding to highly polluted areas) to continuously send excitation pulses to relevant action neurons within a dynamic pollution map, with the excitation intensity positively correlated with the pollution level. Simultaneously, inhibitory pulses are generated within the robot's internal state (such as battery level and current position), introducing decision constraints.

[0032] All action neurons undergo membrane potential evolution under the continuous stimulation and inhibition pulse flow, dynamically adjusting their potentials through a competitive and cooperative mechanism. That is, when the membrane potential of a certain action neuron first reaches the preset firing threshold, the pulse sequence it triggers is interpreted in real time as a cleaning path instruction to guide the robot's action, while the potentials of other neurons are temporarily suppressed to ensure the uniqueness of the action.

[0033] It should be noted that the competition mechanism is entirely based on asynchronous event-driven pulse flow, eliminating the need for a central processing unit to perform global path search and achieving millisecond-level low-latency decision-making. However, during actual use, the pollution priority and robot operation constraints can be pre-balanced, and the weights of attention incentives and inhibition pulses can be configured to avoid local optimum traps. Meanwhile, continuous updates to spatiotemporal characteristics ensure the dynamic adaptability of instructions, but the membrane potential leakage rate needs to be matched with the pollution diffusion rate to prevent decision oscillations.

[0034] Step 105: Perform the cleaning action and report environmental change events, dynamically update the dynamic pollution map and the action selection competition mechanism.

[0035] Specifically, "executing the cleaning action" refers to the photovoltaic cleaning robot actually performing physical cleaning operations based on path instructions generated by the spiking neural network. "Feedback on environmental changes" refers to the camera detecting changes in surface reflectivity after cleaning and generating new asynchronous data packets. "Dynamic updates" refers to adjusting the parameters of the neuron membrane potential and action selection competition mechanism of the dynamic contamination map in real time based on feedback data.

[0036] In some implementations, after the robot executes a cleaning action (such as driving a robotic arm to physically clean a highly contaminated area) based on the neuron instruction that wins the action selection competition mechanism, an event camera monitors surface changes in real time. Any change in reflectivity (such as increased brightness due to contaminant removal) triggers a new event, generating an asynchronous data packet containing coordinates, timestamps, and polarity as feedback input. When the feedback data packet is input into the spiking neural network, the membrane potential of the neuron at the corresponding spatial location is dynamically adjusted according to the event frequency (such as the neuron potential in the cleaned area actively resetting or leaking to simulate cleaning recovery, while the potential of newly contaminated points increases). Simultaneously, the action selection competition mechanism recalculates the excitation pulse flow (such as reinforcing the excitation of neurons related to actions in high-potential areas) and inhibition pulses (such as robot state constraints) based on the updated contamination map, optimizing subsequent path decisions. This enables the entire system to form a closed-loop control, achieving continuous co-evolution of the map and the mechanism.

[0037] It should be noted that the event detection threshold needs to be matched with the cleaning effect in order to effectively avoid over-response.

[0038] Based on the above technical solution, by capturing local information of sudden changes in light intensity and generating asynchronous data packets, the traditional periodic global image acquisition mode is abandoned. This reduces redundant data transmission and processing burden from the source, solving the high energy consumption problem caused by full-frame image acquisition in traditional solutions. Simultaneously, microsecond-level timestamps and independent pixel response mechanisms compress the pollution perception delay from seconds to hundreds of milliseconds, enabling instantaneous capture of sudden pollution such as bird droppings. Furthermore, by statistically analyzing the frequency of events per unit time, asynchronous data packets are mapped to a dynamic pollution level map, making the pollution level positively correlated with the event frequency, constructing a spatial distribution of pollution. This solves the problem that traditional static assessments cannot accurately locate pollution hotspots, allowing the robot to focus on highly polluted areas and avoid the inefficiency caused by indiscriminate cleaning. A distributed dynamic pollution map is then constructed using a spiking neural network. The "charging" of neuronal membrane potential represents pollution accumulation, the "leakage" simulates pollution diffusion, and the corresponding cleaning and recovery evolution process is reset, replacing traditional static map modeling. This solves the problem of the inability to update pollution diffusion and post-cleaning status in real time in dynamic environments, enabling the robot to automatically track the "pollution front" and achieve proactive cleaning of diffused pollution. Finally, the action selection competition mechanism combines the excitation pulses of high-potential areas on the contamination map with the inhibition pulses of the robot's power, position, and other states. This allows the action neurons to naturally emerge with optimal cleaning path instructions through membrane potential competition, eliminating the need for complex global path search. This solves the problems of high latency and poor dynamic adaptability in traditional serial decision-making, achieving millisecond-level response to environmental fluctuations such as the appearance of new contamination and changes in light and shadow. In addition, the event camera provides real-time feedback on environmental changes, dynamically updating the contamination map and competition mechanism parameters. This effectively forms a closed-loop progressive architecture of "perception, transformation, modeling, decision-making, and feedback." It not only solves the shortcomings of traditional planning's one-time decision-making that cannot be adjusted, but also achieves decentralized work area division through the suppression and coordination of pulse signals between multiple machines, avoiding overlapping operations. At the same time, global event flow monitoring distinguishes between changes in light and shadow and local contamination signals, improving anti-interference capabilities. Ultimately, this achieves high-efficiency photovoltaic cleaning with low energy consumption, fast response, high adaptability, and strong collaboration, improving the cleaning efficiency and power generation benefits of photovoltaic modules.

[0039] In one possible implementation of the embodiments of this application, combined with Figure 1-2 As shown, the construction process of a spiking neural network can be achieved through the following steps 201 to 203, which are explained in detail below: Step 201: Construct a spiking neural network structure including an event input layer, a contamination processing layer, and an action output layer. The event input layer consists of multiple event response neurons, each corresponding to a specific coordinate position of the photovoltaic module, used to receive asynchronous data packets from the event camera and generate pulse signals. The contamination processing layer consists of a dynamic contamination simulation neuron network, where neurons simulate the accumulation and diffusion process of contamination through membrane potential evolution. The action output layer includes competitive action neurons, each encoding a basic cleaning action, and selecting the optimal action through a pulse competition mechanism. The competition mechanism is based on the dynamic evolution of membrane potential; when the membrane potential of a certain action neuron first reaches the firing threshold, the corresponding cleaning instruction is triggered.

[0040] The event input layer is the first layer in the spiking neural network responsible for receiving event camera data, and it consists of event response neurons. These event response neurons are specialized neuronal units that process asynchronous data packets, with each neuron corresponding to a specific coordinate position on the photovoltaic module. The contamination processing layer is an intermediate layer of the spiking neural network, composed of a dynamic contamination simulation neuron network. These dynamic contamination simulation neurons are neuronal models that simulate the accumulation and diffusion of contamination through membrane potential evolution. The action output layer is the output layer of the spiking neural network, including competitive action neurons. These competitive action neurons select the cleaning action through a spiking competition mechanism, with each neuron encoding a basic action (such as moving forward or cleaning).

[0041] When constructing the spiking neural network structure, the event input layer is directly initialized, mapping each event response neuron to a specific coordinate position on the photovoltaic module surface. This allows the neuron to directly receive asynchronous data packets output by the event camera and generate pulse signals in real time based on the event frequency in the data packets. Next, a contamination processing layer is configured, using a dynamic contamination simulation neural network distributed in a grid pattern. Each neuron simulates contamination accumulation through an increase in membrane potential and contamination diffusion through a leak in membrane potential, thus forming a continuous dynamic evolution process. Finally, an action output layer is set up, using competing action neurons to encode the robot's basic cleaning actions. Continuous pulse input stimulates the membrane potential. When the membrane potential of a certain action neuron reaches a preset firing threshold first in the competition, the corresponding cleaning command is immediately triggered, thus obtaining the complete spiking neural network structure.

[0042] The three layers of the spiking neural network are tightly interconnected. The neurons in the event input layer are mapped to accurately match the physical layout of the photovoltaic module before use, thus avoiding coordinate deviations that could affect perception accuracy. This allows the event input layer to drive the evolution of the contamination treatment layer via pulse signals. The membrane potential evolution parameters of the contamination treatment layer, such as the leakage rate, need to be dynamically adjusted according to the environment (e.g., to accelerate response speed in highly polluted environments). During its use, the dynamic contamination simulation neurons in the contamination treatment layer continuously update their membrane potentials based on the pulse flow from the event input layer. Increased membrane potential simulates contamination accumulation, and leakage simulates diffusion, forming a real-time state output that is transmitted to the action output layer in the form of a pulse flow as the excitation input for competing action neurons. In the action output layer, each action neuron encodes a basic cleaning action (such as moving forward or cleaning) and receives excitation pulses that gradually increase its membrane potential. Through a competitive mechanism, the membrane potentials of all action neurons dynamically evolve. When the membrane potential of a neuron first reaches a preset firing threshold, the corresponding cleaning instruction is immediately triggered, ensuring a direct mapping from state output to action selection, thereby achieving a low-latency, smooth transition from perception to decision-making.

[0043] For example, when the event camera detects bird droppings at a certain coordinate location on the surface of a photovoltaic module, the event response neuron at that location in the event input layer receives asynchronous data packets and generates a high-frequency pulse signal. In the pollution processing layer, the membrane potential of the relevant dynamic pollution simulation neuron rapidly increases to simulate pollution accumulation and spreads to the surrounding area through leakage (e.g., the membrane potential of the corresponding neuron rapidly increases to 30mV as a state output). Meanwhile, in the evolution of the membrane potential of the competing action neurons in the action output layer, the "cleaning" action neuron, receiving a high-excitation pulse, reaches the threshold first, triggering a local cleaning command (e.g., a membrane potential that reaches the 50mV threshold first in the competition triggers an immediate cleaning command). After the robot performs cleaning, the event camera reports a change, and the entire network continues to process new events in a loop, forming a complete response closed loop.

[0044] Step 202: Integrate a dedicated neuromorphic computing chip to process pulse streams in parallel and embed a dynamic reconfiguration interface to adapt to the spatial mapping of photovoltaic arrays of different sizes.

[0045] Among them, the dedicated neuromorphic computing chip is a special-purpose hardware integrated circuit that simulates biological neural network processing. Dynamic reconfiguration allows the chip's logic resources to be reconfigured at runtime to adapt to changing needs. Spatial mapping of the photovoltaic array is the process of mapping the physical layout coordinates of the photovoltaic modules to the chip's logic structure.

[0046] In some implementations, a dedicated neuromorphic computing chip is first integrated onto the system motherboard, allowing direct connection between the event input layer and the action output layer via a high-speed bus. An event-driven architecture is configured on the chip, enabling it to activate processing units only upon receiving a pulse stream, avoiding continuous power consumption. Simultaneously, the chip's parallel processing capabilities are utilized to process multiple pulse streams from the contamination processing layer, ensuring low-latency response. Furthermore, a dynamic reconfiguration function adjusts the chip's neuron mapping relationships online according to the actual size of the photovoltaic array (e.g., reallocating logic resources when the array expands). Ultimately, adaptive spatial mapping is achieved, supporting seamless deployment of photovoltaic modules from small to large sizes.

[0047] For example, when the photovoltaic array is expanded from 10x10 to 20x20, the dedicated neuromorphic computing chip automatically adjusts the logic mapping through dynamic reconfiguration, mapping the new coordinates to idle neurons; the event-driven architecture ensures that processing is activated only when a contamination event occurs, and the parallel processing pulse stream processes the input of multiple highly contaminated areas simultaneously, ultimately maintaining an efficient cleaning response.

[0048] Step 203: Adjust the connection weights between neurons by adjusting the pulse time to optimize the accuracy of the contamination simulation.

[0049] In some implementations, immediately after the cleaning action is performed, an event camera monitors changes in the reflectivity of the photovoltaic module surface and generates a feedback event stream containing new timestamps. The firing times of presynaptic neurons can then be directly extracted from this feedback event stream. and the pulse firing time of postsynaptic neurons Then, according to the pulse time-dependent plasticity rule, through The weight change ΔW is calculated, with the learning rate η preset to 0.01 to balance adjustment speed and stability. Finally, the connection weights between neurons can be updated online. By optimizing the weight values ​​in real time, the accuracy of pollution simulation is improved, allowing the spiking neural network to more accurately map the dynamics of actual pollution.

[0050] It should be noted that the learning rate η needs to be dynamically adjusted according to the environment. For example, in scenarios with high pollution fluctuations, η can be appropriately increased to accelerate adaptation.

[0051] For example, after a certain area of ​​a photovoltaic module is cleaned, an event camera detects an increase in reflectivity and generates a feedback event stream, in which the presynaptic neuron pulse firing time... It is 100.5 ms, postsynaptic neuron The time was 100.8ms; system calculation This allows for a direct reduction in connection weights online, resulting in a more accurate match between the optimized contamination simulation and the actual cleanliness.

[0052] Based on the above technical solution, by constructing a three-level spiking neural network structure consisting of an event input layer, a pollution processing layer, and an action output layer, the event response neurons in the event input layer, corresponding to specific coordinates of the photovoltaic module, can accurately receive asynchronous data packets from the event camera and generate pulse signals. This eliminates the redundant acquisition and processing of traditional full-frame images, reduces the data burden from the source of perception, and solves the problems of high energy consumption and second-level pollution perception delay of traditional visual sensors. It enables microsecond-level capture of sudden pollution such as bird droppings, making pollution detection more accurate and efficient. Then, through the dynamic pollution simulation neural network in the pollution processing layer, the evolution process of membrane potential increase, leakage, and reset is used to realistically simulate pollution accumulation, diffusion, and cleaning recovery, replacing traditional static grid map modeling. This solves the problem that static planning cannot cope with dynamic changes in pollution and the "one-size-fits-all" cleaning strategy, enabling the robot to focus on highly polluted areas and avoid waste of water resources and energy. Finally, by using competitive action neurons that encode basic cleaning actions in the action output layer, and relying on the competitive mechanism of dynamic evolution of membrane potential, the optimal cleaning path emerges naturally. This solves the problems of high decision latency and weak dynamic adaptability of traditional path planning algorithms without complex global calculations, achieving millisecond-level decision response.

[0053] The integration of a dedicated neuromorphic computing chip further reduces processing latency by leveraging its ability to process pulse streams in parallel. Simultaneously, a dynamic reconfiguration interface enables flexible spatial mapping of photovoltaic arrays of different sizes, addressing the poor adaptability of traditional equipment and its inability to meet the needs of large-scale and diverse photovoltaic power plants. This effectively improves the equipment's versatility and deployment flexibility. Finally, by adjusting the connection weights between neurons using pulse time-dependent plasticity rules, the accuracy of contamination simulation is optimized in real time based on cleaning feedback. This solves the problem of large deviations between traditional model contamination simulations and actual conditions, which can lead to incomplete cleaning or repeated cleaning. This allows path planning to better align with actual contamination dynamics, ultimately forming a progressive structure characterized by precise perception, efficient processing, flexible adaptation, and accurate simulation. This improves the cleaning efficiency and power generation benefits of photovoltaic modules and reduces power plant operation and maintenance costs.

[0054] In one possible implementation of the embodiments of this application, combined with Figure 1-3 As shown, the process of constructing a distributed dynamic pollution map using a spiking neural network and simulating pollution diffusion and cleanup recovery through neuronal membrane potential evolution can be achieved through the following steps 301 to 304, which are explained in detail below: Step 301: Based on the physical dimensions of the photovoltaic module, map each event response neuron in the event input layer to the spatial coordinates of the photovoltaic module surface, establish a fixed mapping relationship from physical coordinates to virtual neurons, and form a lookup table from physical coordinates to neurons.

[0055] In this context, the event response neuron refers to the computational unit in the event input layer responsible for receiving and processing events related to changes in the reflectivity of the photovoltaic module surface. Each neuron is bound to a specific coordinate position in physical space. The virtual neuron is a computational node in the spiking neural network used to simulate the pollution diffusion and cleanup process, forming a spatial correspondence with the photovoltaic module through physical coordinate mapping. The lookup table is a pre-established mapping table between physical coordinates and virtual neuron numbers, used to achieve rapid indexing and location of event data to neuron activation signals.

[0056] In some implementations, based on the physical size parameters of the photovoltaic module (such as length and width in pixels), a uniform grid partitioning method is used to divide the surface of the photovoltaic module into rectangular regions equal to the number of neurons. The center point of each region is defined as the response coordinate of the neuron (e.g., for an M×N photovoltaic module, M×N event response neurons are constructed, and the coordinate of each neuron i corresponds to...). This allows each event response neuron in the event input layer to be bound to the physical coordinates on the surface of the photovoltaic module, forming a fixed one-to-one correspondence. This enables the direct generation of a lookup table from physical coordinates to virtual neurons, storing the neuron number and activation threshold parameter corresponding to each physical coordinate (x, y).

[0057] When the event camera detects a change in reflectivity, it can directly query the lookup table to immediately locate the virtual neuron corresponding to the physical location that triggered the event, and generate a pulse signal of the corresponding polarity according to the event polarity (increase / decrease in brightness) to drive the neurons in the contamination treatment layer to perform membrane potential evolution calculations.

[0058] It should be noted that the mapping mechanism adopts a fixed coordinate binding strategy to ensure the long-term consistency between the event response neuron and the physical location, and to avoid coordinate deviation caused by dynamic reconfiguration.

[0059] Taking a 10×10 photovoltaic module as an example, 100 event response neurons are constructed during system initialization, with each neuron corresponding to physical coordinates from (1,1) to (10,10). A lookup table records the mapping relationship between each coordinate (x,y) and the neuron number (e.g., (5,3) corresponds to neuron number 43). When the event camera detects a sudden pollution event (a sharp drop in reflectivity) at coordinate (7,8), it immediately locates neuron number 68 through the lookup table, triggering this neuron to generate a high-frequency negative polarity pulse signal. After receiving the pulse signal, the membrane potential of the corresponding neuron rapidly increases, simulating the accumulation process of pollutants in that area. Simultaneously, the pollution signal is transmitted to adjacent neurons through a pulse diffusion mechanism, ultimately forming a high-pollution level area centered on (7,8) in the dynamic pollution map. This process, from event triggering to neuron activation, takes only 50 microseconds, achieving millisecond-level response to pollution events.

[0060] Step 302: Input the asynchronous data packet into the spiking neural network, find the corresponding event response neuron according to the event coordinates, calculate the membrane potential increment, and drive the contamination treatment layer to update the membrane potential.

[0061] Among them, membrane potential increment refers to the numerical value of the change in membrane potential caused by neurons in a spiking neural network receiving pulse signals. A positive increment indicates pollution accumulation, while a negative increment indicates cleanliness and recovery.

[0062] In some implementations, after the event camera captures a change in reflectivity on the surface of the photovoltaic module, it can directly parse the asynchronous data packet to extract the event coordinates (x, y), timestamp (t), and polarity (p). Then, using a pre-built lookup table of physical coordinates to virtual neurons, it quickly locates the event response neuron corresponding to the event coordinates. If the event polarity is positive (increased contamination), then... Calculate the membrane potential increment ,in Here, T is the sensitivity coefficient, and T is the current timestamp. This is the timestamp of the neuron's last activation; if the polarity is negative (contamination removal), then... Calculate the membrane potential increment ,in For cleaning sensitivity coefficient, This represents the pollution decay time constant.

[0063] At this point, the contamination treatment layer can be directly adjusted based on the membrane potential increment. ,pass Calculate the updated membrane potential value of the corresponding neuron. ,in For membrane potential leakage rate, The time step is defined as θ. The updated membrane potential then propagates to neighboring neurons through pulse firing rules (such as triggering a pulse when a threshold θ is reached), forming a spatiotemporal simulation of contamination diffusion.

[0064] It should be noted that the local response mechanism only updates the membrane potential of neurons around the event coordinates, avoiding global computation. Additionally, the time decay factor in the membrane potential increment formula... This ensures that the rate of contamination accumulation is positively correlated with the frequency of events, while the exponential decay model of cleaning events simulates the gradual effect of cleaning actions. Finally, the neuron connection weights of the contamination treatment layer are dynamically adjusted through the pulse time-dependent plasticity (STDP) rule to enhance the spatiotemporal correlation of contamination propagation.

[0065] For example, in a 10×10 photovoltaic module, if the event camera detects a sudden contamination (positive polarity) at coordinates (7,8), the corresponding neuron N68 can be located using a lookup table. Further assuming the current timestamp T = 100ms and the last activation time was... =50ms, sensitivity coefficient =0.5 / mV·ms, then the membrane potential increment =0.5×(100-50)=25mV. The contamination treatment layer then updates the membrane potential of N68 from the initial value of 20mV to 45mV (below the threshold of 50mV). Subsequently, N68 transmits the signal to its four adjacent neurons (N67, N69, N77, N79) according to the distance decay rule. (d is the spatial distance, λ=2 pixels), increasing the membrane potential of adjacent neurons by 3mV. When N68 is triggered again in the next cycle, the cumulative membrane potential reaches a trigger pulse of 55mV, driving the action neuron to execute the local cleaning command, while resetting the membrane potential of the region to its initial value, completing the complete simulated closed loop from contamination to cleaning.

[0066] Step 303: Monitor the reflectivity of the photovoltaic module after cleaning, generate a reset event and send it to the contamination treatment layer, and update the dynamic contamination simulation neuron potential at the corresponding coordinates in the contamination treatment layer.

[0067] Among them, the reset event refers to a special event triggered by the event camera after detecting the completion of the cleaning action on the surface of the photovoltaic module, used to indicate that the contaminated area has returned to a clean state. The dynamic contamination simulation neuron potential is the internal potential value of the neuron representing the degree of contamination. The potential level is positively correlated with the contamination level and is dynamically evolved and updated through pulse signals.

[0068] In some implementations, after the cleaning robot completes cleaning a certain area, the event camera continuously monitors changes in the reflectivity of that area. When a sudden brightness change caused by contaminant removal is detected (e.g., a reflectivity threshold increase exceeding 20%), a reset event containing event coordinates, a timestamp, and positive polarity is directly generated. This reset event is then transmitted to the contamination processing layer via a high-speed communication bus, allowing the layer to look up the physical coordinates in the virtual neuron's lookup table based on the reset event's coordinates and locate the corresponding neuron.

[0069] At this point, the contamination treatment layer can perform a membrane potential reset operation on the neuron: if the cleaning action completely removes the contamination, the membrane potential is directly reset to the initial baseline value (e.g., 20mV); if residual contamination exists, a time-based exponential decay model is initiated. ),in The pollution attenuation coefficient, (The duration after cleaning) is used to gradually reduce the membrane potential to the environmental baseline level. Simultaneously, the updated membrane potential state will diffuse to adjacent neurons via a pulse transmission mechanism, correcting the pollution diffusion simulation results in the dynamic pollution map.

[0070] It should be noted that by differentiating the signal polarity between positive polarity reset events and negative polarity events caused by contamination accumulation, misjudgment can be effectively avoided. Simultaneously, the tiered membrane potential reset strategy, which triggers a hard reset in completely clean areas and a soft reset in areas with residual contamination, balances cleaning effectiveness tolerance with energy consumption optimization.

[0071] For example, in a 10×10 photovoltaic module, after the cleaning robot completes local cleaning at coordinates (7,8), the event camera detects that the reflectivity of this area has recovered from 30% to 85%, generating a reset event (coordinates (7,8), timestamp 200ms, positive polarity). The system then locates neuron N68 using a lookup table, finding its current membrane potential to be 45mV (not yet reaching the cleaning threshold). Since the cleaning action completely removes the contamination, the system hard resets the membrane potential of N68 to 20mV and triggers a reverse pulse transmission to adjacent neurons (N67, N69, N77, N79), reducing the membrane potential of each adjacent neuron by 5mV. This update eliminates the high pollution level in the (7,8) area of ​​the dynamic pollution map and corrects the pollution diffusion simulation results of the surrounding area, providing accurate environmental status information for subsequent path planning.

[0072] Step 304: Extract the neural network in the pollution treatment layer to form a pollution state distribution map.

[0073] The neural network refers to a three-dimensional topological structure composed of all neurons in the contamination treatment layer and their connections, including parameters such as spatial coordinates, membrane potential values, and connection weights. The contamination state distribution map is a visualized two-dimensional image generated by mapping the membrane potential values ​​of each neuron in the contamination treatment layer according to physical coordinates, with color gradients representing the contamination level.

[0074] In some implementations, all neurons in the contamination treatment layer are traversed, and the physical coordinates (x, y) and current membrane potential value V of each neuron are extracted to form a structured dataset containing spatial location and contamination level. Then, based on preset contamination levels and color mapping rules (e.g., red represents high contamination, green represents cleanliness), the membrane potential values ​​are normalized to the 0-255 range, generating a rasterized image matrix matching the physical dimensions of the photovoltaic module. The matrix is ​​then converted into a pseudo-color heatmap by an image rendering engine, where the color value of each pixel corresponds to the contamination level at a specific coordinate, ultimately forming a real-time updated contamination status distribution map.

[0075] It should be noted that the spatial distribution characteristics of contamination can be intuitively reflected through color gradients during use. Red areas (membrane potential > 80% of the maximum value) mark highly contaminated hotspots that require priority cleaning, while blue areas (membrane potential < 20% of the maximum value) represent clean areas. To reduce computational load, the system employs a layered rendering strategy: when the frequency of contamination events is below a threshold, only the local image of the changed area is updated; when a sudden contamination occurs, a global redraw is triggered. Furthermore, the distribution map can support multi-scale display, showing both the microscopic contamination details of individual components and synthesizing macroscopic contamination status maps of large photovoltaic arrays through coordinate mapping.

[0076] For example, in a 10×10 photovoltaic module, after the membrane potential of neurons in the pollution treatment layer evolves, a locally highly polluted area can be formed (e.g., the membrane potential of neurons near coordinates (5,5) reaches 90%). This process involves traversing all neurons, extracting their coordinates and membrane potential values, and generating a 100×100 pixel heatmap matrix after normalization. The rendering engine marks areas with membrane potential >80% as red, 40%–80% as yellow, and <40% as green, ultimately forming a radial red pollution area centered at (5,5). This distribution map is displayed in real-time on the monitoring interface, guiding the robot to prioritize the cleaning path to (5,5), while the surrounding green areas indicate that no intervention is currently needed, achieving precise operation scheduling.

[0077] Based on the above technical solution, each neuron is bound to the specific physical coordinates of the photovoltaic module through a grid, forming a fast index lookup table. This allows for precise location of pollution events, solving the problems of coordinate deviation and event response delay in traditional pollution location. Events involving reflectivity changes captured by the event camera can be instantly matched to the corresponding neurons through the lookup table. Simultaneously, after inputting asynchronous data packets into the spiking neural network, the membrane potential increment of the corresponding event response neuron is calculated, driving the pollution treatment layer to update its potential. This allows for the simulation of pollution accumulation through positive increments and cleaning recovery through negative increments. Furthermore, the dynamic evolution of pollution diffusion is achieved through membrane potential leakage, solving the problem that traditional static maps cannot reflect real-time dynamic changes in pollution. This makes the simulation of pollution states more closely resemble the actual pollution development process, ensuring the real-time accuracy of pollution degree judgment and resetting event generation. A graded strategy of hard reset during complete cleaning and soft reset during residual pollution can be used to update the potential of the dynamic pollution simulation neurons, forming a closed-loop feedback structure. This solves the problem of delayed pollution state updates after traditional cleaning, avoiding repeated cleaning or omission of residual pollution due to outdated data in subsequent path planning, and ensuring that the pollution map always remains consistent with the actual cleanliness state. Finally, by generating a visualized pollution distribution map through color gradient mapping, the abstract neuronal membrane potential can be transformed into an intuitive heat map, clearly identifying high-pollution hotspots and clean areas. This solves the problem of blindness in traditional cleaning path planning, allowing robots to accurately focus on high-pollution areas, avoiding waste of water and energy. At the same time, it provides a clear basis for pollution distribution for multi-robot collaborative operations, helps decentralized work area division, improves the overall efficiency and operation and maintenance management level of photovoltaic cleaning, and forms a progressive technical advantage from precise positioning, dynamic simulation, closed-loop feedback to precise scheduling, comprehensively optimizing the practicality and economy of photovoltaic cleaning.

[0078] In one possible implementation of the embodiments of this application, combined with Figure 1-4 As shown, the process of establishing an action selection competition mechanism in a spiking neural network and generating cleaning path instructions based on the spatiotemporal distribution characteristics of a dynamic pollution map can be achieved through the following steps 401 to 403, which are explained in detail below: Step 401: Construct a spatiotemporal feature coding layer. The spatiotemporal feature coding layer includes feature extraction neurons connected to the neuron cluster in the pollution treatment layer. These neurons are used to receive membrane potential pulses and calculate the feature values ​​of pollution accumulation intensity and pollution diffusion trend.

[0079] The spatiotemporal feature encoding layer is a key layer in the spiking neural network used to analyze the spatiotemporal characteristics of pollution maps, containing a cluster of feature extraction neurons. These feature extraction neurons are computational units specifically designed to receive and process the membrane potential pulses input from the pollution treatment layer, extracting pollution features through spatiotemporal correlation analysis. The pollution diffusion trend feature value is a vector parameter that measures the rate of change of the pollution area boundary, characterizing the spatial spread direction and speed of pollutants.

[0080] In some implementations, the spatiotemporal feature coding layer establishes a full connection with the contamination processing layer through feature extraction neurons. This allows each feature extraction neuron to receive membrane potential pulses from neurons in its 3x3 neighborhood of the contamination processing layer. Thus, when a contamination processing layer neuron generates a pulse due to an event, the feature extraction neuron synchronously initiates spatiotemporal feature computation. By statistically analyzing the pulse frequencies of neurons in the neighborhood within a 100ms time window, a Gaussian weighted sum is applied. Calculate the characteristic value of cumulative pollution intensity ,in For distance decay weights, The pulse frequency is used. Simultaneously, the spatial gradient vector of the pulse event within the neighborhood is calculated. Obtain the characteristic value of pollution diffusion trend ,in L represents the peak environmental pollution level inside and outside the neighborhood, and L is the side length of the neighborhood.

[0081] Therefore, the characteristic value of cumulative pollution intensity can be directly obtained. and pollution diffusion trend characteristic values After normalization, they are mapped to floating-point numbers between 0 and 1, and then converted into an excitation pulse sequence for the action neuron array through pulse frequency modulation. At this time, the firing frequency of the excitation pulse is positively correlated with the magnitude of the feature value.

[0082] It should be noted that the spatiotemporal feature encoding layer employs a local perception mechanism, where each feature extraction neuron processes pollution information only within its neighborhood, avoiding global computation through spatiotemporal correlation constraints. Simultaneously, the pollution accumulation intensity feature value is dynamically updated using a sliding time window mechanism, the size of which can adaptively adjust according to the pollution diffusion rate. Furthermore, a pollution level gradient threshold determination mechanism is introduced in the calculation of the pollution diffusion trend feature value. When the gradient value is less than a set threshold, it is determined to be a stable pollution region, and the diffusion trend calculation is stopped to reduce energy consumption.

[0083] In a 10x10 photovoltaic module cleaning scenario, when bird droppings suddenly accumulate at coordinates (5,5), the event camera triggers a high-frequency negative polarity pulse event. At this time, the neuron corresponding to the contamination processing layer accumulates a pulse frequency of 200Hz within 100ms, while the feature extraction neuron of the spatiotemporal feature coding layer calculates... =0.85, =(0.3,0.4).

[0084] At this point, after pulse frequency modulation, this feature value can generate an excitation pulse sequence (firing frequency of 150Hz) for the "forward" action neuron and an inhibitory pulse of 50Hz for the "cleaning" action neuron. Under the competitive mechanism, the "forward" action neuron wins because its membrane potential reaches the threshold first, driving the robot towards the pollution source. When the robot approaches the polluted area, the pollution diffusion trend feature value can be updated. =(0.1,0.2), and automatically reduce the priority of the cleaning command, prioritizing the path correction action.

[0085] Step 402: Convert the pollution accumulation intensity feature value and pollution diffusion trend feature value into an excitation pulse sequence for different action neurons in the action neuron array according to the preset action mapping rule.

[0086] The action mapping rule is a pre-defined mathematical transformation model used to map contamination feature values ​​to excitation pulse parameters of action neurons. The action neuron array is a decision layer composed of multiple spiking neurons encoding different cleaning actions, where each neuron corresponds to a robot action command.

[0087] In some implementations, the cumulative intensity eigenvalue is first... With diffusion trend characteristic value Perform vector superposition Generate a comprehensive pollution feature vector At this point, the comprehensive pollution feature vector... The eigenvalues ​​will be obtained by normalizing the hyperbolic tangent function. Then, based on the preset threshold segmentation function, the normalized feature values ​​are... Mapped to the activation weights of the four action channels: when When ∈ [0, 0.3), only the "forward" action neurons are activated, with weights of . ×5; when When the input range is [0.3, 0.7), both "forward" and "left turn" are activated simultaneously, with weights of 0.6 for each channel. and 0.4 ;when ≥0.7 activates the "cleaning" action neurons, with a weight of 1- .

[0088] At this time, the excitation pulse frequency of each action neuron Pulse width =5ms, and frequency encoding is achieved through pulse interval modulation. Finally, the pulse sequences of all action neurons compete in the time dimension, and the neuron with the fastest membrane potential rise is given the right to perform the action.

[0089] It should be noted that the mapping rule adopts a segmented activation strategy, using dynamic weight allocation to control action priority: when the pollution spread trend exceeds a set threshold, the pollution boundary tracking mode is automatically activated, increasing the weight of the spread trend feature value by 30%. Simultaneously, to prevent the simultaneous activation of multiple action neurons, a lateral inhibition mechanism is introduced, ensuring that the intensity of the inhibitory impulse of each action neuron is proportional to its excitation weight. Furthermore, this mapping rule can also support online learning and optimization, dynamically adjusting the threshold cutoff point and weight allocation coefficients through an error backpropagation algorithm, with an optimization cycle of every 10 cleaning cycles.

[0090] In a photovoltaic array cleaning scenario, when the cumulative intensity of contamination... =0.85, diffusion trend When = (0.3, 0.4), the calculation yields... =0.56, while after normalization =0.53. Based on the mapping rule, the "forward + left turn" compound action can be activated directly, with the activation weights of the two channels being 0.318 and 0.212, respectively. In the corresponding pulse sequence, the "forward" neuron fires pulses at a frequency of 31.8 Hz, and the "left turn" neuron fires pulses at a frequency of 21.2 Hz, with pulse intervals of 31.4 ms and 47.2 ms, respectively. During the competition, because the "forward" action has a higher activation weight and more frequent pulse firing, its membrane potential reaches the firing threshold of 0.6 V first at 23 ms, triggering the robot's compound action command to shift 15° to the left and continue forward.

[0091] Step 403: All action neurons in the action neuron array compete to fire asynchronous pulses in parallel based on their current membrane potential, forming a cleaning path instruction.

[0092] In some implementations, the action neuron array employs a circular topology, where each neuron corresponds to a standardized cleaning action (forward / left turn / right turn / clean). Furthermore, each neuron maintains an independent membrane potential variable. The initial value is the resting potential. .

[0093] When the neuron receives the excitation pulse sequence from the spatiotemporal feature coding layer, it will modulate according to the pulse frequency rules. Update membrane potential ,in For the input pulse frequency, =0.05 is the incentive weight. =0.002 is the leakage coefficient. =1ms is the simulation time step. When ≥threshold When the voltage reaches 0.6V, the neuron immediately fires a pulse, and a competitive arbitration mechanism determines the winning action. The arbitration logic follows the fastest-first-served principle: when multiple neurons reach the threshold simultaneously, the neuron that first breaks the threshold gains the right to execute, and the pulses of the remaining neurons are suppressed. The action code of the ultimately winning neuron is converted into a PWM signal by a pulse encoder, driving the chassis motor and the cleaning nozzle to perform the corresponding action.

[0094] It should be noted that the competition mechanism employs a fully hardware parallel architecture, where the membrane potential calculation and comparison of each action neuron are performed independently on a dedicated neuromorphic chip, requiring no software intervention. Furthermore, to prevent action oscillations, a dynamic threshold adjustment mechanism can be introduced, adjusting the threshold after each action execution. The algorithm will increase by 10% in the next cycle, continuing for 3 time steps before returning to the baseline value. For multi-target conflict scenarios (such as simultaneously detecting highly contaminated areas and obstacles), action weights can also be set through a priority encoder. The weight for cleaning actions can be set to 1.5, obstacle avoidance actions to 1.2, and regular movement actions to 1.0. Online learning optimization is also supported, dynamically adjusting action priority weight parameters through a Q-learning algorithm.

[0095] For example, when the cumulative intensity of pollution =0.85, diffusion trend When the value is (0.3, 0.4), the sequence of excitation pulses received by the action neuron array is: "Forward" neuron frequency 150Hz, "Left Turn" neuron frequency 100Hz, and "Clean" neuron frequency 50Hz. During parallel computation, the "Forward" neuron, due to its highest excitation weight and earliest pulse emission (t=23ms), is allowed to break through the threshold first, and its emitted pulse signal locks the execution right through arbitration logic. The chassis motor immediately moves forward at a speed of 0.5m / s, while the cleaning nozzle remains in standby mode. When the robot moves to the edge of the contaminated area, the event camera detects a change in the pollutant boundary, triggering a new event stream to update the dynamic contamination map. A new round of action competition then begins, and the "Clean" neuron gains the execution right because the current contamination level meets the standard, and the nozzle activates high-pressure water flow to perform targeted cleaning.

[0096] Based on the above technical solution, by establishing a local sensing connection between the feature extraction neurons and the neuron cluster of the pollution treatment layer, the membrane potential pulse flow in a 3x3 neighborhood can be focused. By using Gaussian weighted statistics and spatial gradient calculation, the characteristic values ​​of pollution accumulation intensity and pollution diffusion trend can be accurately obtained. This avoids the limitation of traditional planning that only focuses on the static pollution location. It achieves dual accurate capture of the severity of pollution and the spread trend from the sensing level, and solves the problem of the ambiguity of the cleaning target caused by the inability to quantify the dynamic characteristics of pollution in traditional solutions. It enables the robot to clearly identify high-pollution hotspots and pollution diffusion directions, avoiding the waste of water resources and energy caused by blind cleaning. Simultaneously, by employing segmented activation rules and lateral inhibition mechanisms, abstract features are mapped to excitation pulse sequences of action neuron arrays, establishing a direct mapping relationship from pollution features to action commands. This addresses the problems of disconnect between action decisions and actual pollution conditions, as well as prioritization issues, in traditional planning. It allows for dynamic allocation of action weights based on pollution accumulation intensity and diffusion trends; for example, prioritizing the activation of movement-type action neurons during mild diffusion and strengthening cleaning-type action excitation during severe pollution, ensuring a high degree of adaptation between action decisions and the pollution scenario. Finally, a ring topology structure is adopted, enabling each action neuron to compete asynchronously and in parallel based on independent membrane potential evolution. The optimal action is determined through a fastest-priority arbitration mechanism, generating cleaning path commands and resolving [the remaining issues]. Addressing the core issues of traditional solutions, such as high decision-making delays and inability to respond in real time to dynamic environmental changes, this technology can adapt to sudden situations such as the emergence of new pollution and changes in light and shadow within milliseconds. By combining precise feature perception with reasonable action mapping, it forms a progressive technical structure characterized by accurate perception, action adaptation, and rapid decision-making. This not only reduces the delay between perception and execution from the second level to the hundredth millisecond level, enabling the robot to clean sudden pollution such as bird droppings instantly, but also enhances its adaptability to complex dynamic scenarios. Furthermore, it reduces redundant energy consumption through an event-driven computing model and, in conjunction with a multi-machine pulse signal coordination mechanism, further achieves decentralized and efficient collaborative operation, comprehensively improving the cleaning efficiency and power generation benefits of photovoltaic modules and reducing power plant operation and maintenance costs.

[0097] In one possible implementation of the embodiments of this application, combined with Figure 1-5 As shown, the process of collecting the event stream of reflectivity changes on the surface of photovoltaic modules and generating asynchronous data packets containing event coordinates, timestamps, and polarity can be achieved through the following steps 501 to 505, which are explained in detail below: Step 501: Deploy a dynamic vision sensor, which operates in an asynchronous mode, with each pixel unit responding independently to changes in light intensity.

[0098] In asynchronous operation mode, the sensor does not rely on an external clock signal; it only triggers data acquisition and transmission when a change in a physical quantity is detected. A pixel unit is the smallest photosensitive element constituting the sensor's imaging surface, possessing independent photoelectric signal conversion capabilities. Changes in light intensity refer to the differences in the intensity of reflected or transmitted light caused by changes in the physical properties of the measured object's surface (such as the adhesion of contaminants).

[0099] In some implementations, dynamic vision sensors with event-aware capabilities are deployed on the surface of photovoltaic modules. The pixel arrays within these dynamic vision sensors employ an independent asynchronous operating mechanism. Each pixel unit continuously monitors changes in the light intensity of its corresponding area.

[0100] For example, when localized oil contamination appears on the surface of a photovoltaic module, a dynamic visual sensor deployed directly above the module monitors changes in light reflection in real time.

[0101] Step 502: When the logarithmic change in the light intensity of a local area exceeds a preset intensity threshold, an event acquisition is triggered to obtain a reflectivity change event stream.

[0102] In some implementations, when the logarithmic change in illumination intensity at a pixel location exceeds a preset threshold (e.g., ±5%), an event generation mechanism is immediately triggered. This is output as an asynchronous data packet, containing the pixel coordinates (x, y) where the event occurred, a microsecond-level timestamp (t) indicating the triggering of the event, and a polarity parameter representing the direction of the brightness change (positive polarity indicates increased brightness, negative polarity indicates decreased brightness). For example, when bird droppings land on the surface of a photovoltaic panel, the corresponding pixel will detect a sudden drop in reflectivity, triggering an event data packet containing coordinates (12, 34), a timestamp (15:23:45.678901), and a negative polarity.

[0103] For example, when a pixel unit of a dynamic vision sensor detects a 12% decrease in the reflectivity of a certain area due to oil contamination, it immediately generates an event data packet containing coordinates (56, 78), a timestamp (09:45:12.345678), and a negative polarity. After filtering out homopolar noise caused by ambient light fluctuations, the data packet is marked as a valid contamination event and assigned a high priority.

[0104] Step 503: Perform real-time filtering on the reflectivity change event stream to remove global homopolar events caused by ambient light fluctuations, and then generate a preprocessed event stream.

[0105] Among them, global same-polarity events are events of the same polarity collectively triggered by sensor pixel units due to ambient light fluctuations (such as cloud cover or shadow movement). The preprocessed event stream is a sequence of local pollution-related events retained after filtering, which is used for subsequent pollution level calculation and path planning.

[0106] In some implementations, a dynamic filtering algorithm based on a sliding time window is deployed at the event camera output, establishing a 50ms time window and statistically analyzing the polarity distribution ratio of all events in the original event stream within the window. When the proportion of positive / negative polarity events exceeds 80%, it is determined to be a globally homogeneous event (such as a sudden change in ambient light), and all events within that window are marked as noise and removed. A polarity weighting strategy is applied to the remaining events. For example, contaminated events (negative polarity) can be assigned a weighting coefficient of 0.8, and clean events (positive polarity) can be assigned a weighting coefficient of 0.3. This allows the Kalman filter to smooth the weighted event stream, eliminating high-frequency jitter noise.

[0107] For example, when a photovoltaic panel is suddenly contaminated by bird droppings, the event camera detects 200 negative polarity events within 50ms. After the polarity consistency check is passed, the event stream is retained. However, when rapid cloud movement causes a decrease in the overall brightness of the module surface, the system detects that the polarity consistency exceeds the threshold and filters out all 1500 negative polarity events within that time period. After filtering, the proportion of contamination events in the event stream increases, improving the signal-to-noise ratio of contamination detection.

[0108] It should be noted that the filtering technology adopts a hardware-level parallel processing architecture, enabling real-time processing with a latency of 5μs on an FPGA chip. The time window size (adjustable from 20-100ms) and polarity threshold (adjustable from 50% to 90%) can be dynamically adjusted to adapt to various complex lighting environments, from strong desert light to weak rainy light.

[0109] For example, under cloudy weather conditions, the surface of photovoltaic modules is alternately exposed to direct sunlight and cloud cover. When the clouds move rapidly, the event camera detects 5000 negative polarity events (sudden drop in brightness) and 3000 positive polarity events (brightness recovery) within 200ms. Since the polarity consistency exceeds 95% through a sliding time window, it is identified as a global ambient light event, and all events within that time period are filtered out. Subsequently, when oil pollution occurs in a certain area, the event camera detects 20 negative polarity events within 50ms. After polarity weighting (total weight 16) and Kalman filtering, the event stream is retained. With pollution events accounting for 85% of the preprocessed event stream, the corresponding neuron cluster is quickly activated, generating a high-priority preprocessed event package containing coordinates (45, 67), a timestamp (14:32:01.234), and negative polarity, providing accurate input for subsequent pollution hotspot tracking.

[0110] Step 504: Establish the mapping relationship between the physical coordinate system of the photovoltaic module and the pixel coordinate system of the event camera, and assign dynamic priority and weight to the pollution events according to the event polarity parameters.

[0111] The photovoltaic module physical coordinate system is a two-dimensional planar coordinate system established based on the actual physical dimensions of the photovoltaic panel, with the horizontal and vertical axes corresponding to the length and width of the module, respectively. The event camera pixel coordinate system is a virtual coordinate system defined by the geometric arrangement of the sensor chip pixel array, with each pixel unit corresponding to a fixed spatial sampling point.

[0112] In some implementations, during system initialization, a uniform grid partitioning method is used to establish the mapping relationship between the physical coordinate system of the photovoltaic module and the pixel coordinate system of the event camera. Assuming the photovoltaic module size is 2m × 1.5m and the sensor resolution is 1920 × 1080 pixels, then each physical coordinate point (x, y) corresponds to the pixel coordinates... Where W and H are the width and height of the component, and (u0, v0) are the coordinates of the optical center of the sensor.

[0113] When the event camera captures a reflectance change event, the pixel coordinates (u=987, v=543) can first be converted to physical coordinates (x=1.234m, y=0.678m) using this mapping relationship. Then, dynamic priority is calculated based on the event polarity parameter: if the event polarity is negative (contamination event), the base priority is set to 100, multiplied by the time decay factor. (λ=0.1 / s, Δt is the time elapsed since the event occurred); if it is a positive polarity (clean event), the basic priority is set to 30. The final weighted value P = basic priority × (1 + pollution diffusion coefficient D), where D is dynamically adjusted according to the historical event density (D∈[0.5,2]). For example, if a pollution event occurred 50ms ago and the diffusion coefficient D=1.5, then its weighted priority... This is higher than that of ordinary events.

[0114] For example, when oil contamination appears in the lower right corner of the photovoltaic module, the event camera detects a negative polarity event at pixel coordinates (1890, 945). The preset coordinate mapping rules can then be immediately invoked to convert it to physical coordinates (x=1.83m, y=0.91m). Since the event polarity is negative, the basic priority is set to 100, and the time decay factor is calculated based on the current timestamp (32ms from now). =0.968, multiplied by the diffusion coefficient D=1.5 (the area has historically experienced rapid pollution diffusion), the final weighted priority P=100×0.968×1.5=145.2. This high-priority event is then injected in real-time into the pollution processing layer of the spiking neural network, triggering a surge in the membrane potential of the corresponding neuron cluster, driving the robot to plan a compound action sequence of "right movement → downward movement → cleaning". By comparing this with the case where the event only receives a basic priority of 100 without weighted processing, the weighting mechanism improves the cleaning response speed, verifying the effectiveness of the dynamic prioritization strategy.

[0115] Step 505: Sort the weighted preprocessed event stream by priority and output it to form an asynchronous data packet sequence for tracking pollution hotspots.

[0116] Among them, pollution hotspot tracking is an intelligent monitoring strategy that uses spatiotemporal correlation analysis to locate areas with high pollution density.

[0117] In some implementations, a four-dimensional data structure containing event coordinates, timestamps, polarity weights, and dynamic priorities is established during initialization. When a contaminated event stream undergoes weighted calculation (e.g., a negative polarity event with a weighted priority of 148.5), it is automatically inserted at the head of the priority queue; when a clean event (weighted priority 30) arrives, it is inserted at the tail. Thus, a streaming sorting engine implemented in FPGA hardware can dynamically reorganize millions of events within 5μs.

[0118] Once sorted, the system can directly generate an asynchronous data packet sequence according to the strategy of "prioritizing contaminated hotspots, followed by adjacent areas, and lastly clean areas." Each data packet encapsulates 100 high-priority events and uses compression coding technology to map spatiotemporal coordinates to differential index values, compressing the size of a single data packet to less than 2KB. For example, when a series of oil spill events occur in a certain area, the system generates an asynchronous data packet sequence containing 200 high-priority events within 50ms, transmitting it to the spiking neural network processing unit via gigabit Ethernet to ensure real-time transmission of contaminated hotspot information.

[0119] It should be noted that when a sudden contamination event is detected, the system automatically activates the emergency prioritization mode, raising the priority of the contamination event to the highest level (P=255) to ensure an immediate response to cleaning commands. For multi-robot collaborative operation scenarios, each robot maintains an independent priority queue, and a conflict detection algorithm prevents overlapping work areas.

[0120] For example, in a distributed photovoltaic power station, when a series of oil spill events occur in a certain area, the system generates an asynchronous data packet sequence containing 200 high-priority events within 50ms. These data packets are transmitted via gigabit Ethernet to the spiking neural network processing unit, immediately triggering the pollution hotspot tracking program. At this time, the neural network quickly locates the pollution spread front area (coordinate range x=2.3-2.7m, y=1.8-2.1m) based on spatiotemporal correlation analysis. Simultaneously, adjacent robots automatically adjust their cleaning paths to avoid the pollution hotspot area by receiving the pulse suppression signal from that area. Throughout the tracking process, the system maintains the capacity to process 8000 high-priority events per second, ensuring that the pollution spread speed (0.5m / min) remains lower than the robot cleaning speed (0.8m / min), thus achieving closed-loop control of the pollution hotspot.

[0121] Based on the above technical solution, a dynamic vision sensor operates in asynchronous mode, allowing each pixel to independently respond to changes in illumination. This avoids redundant data from traditional synchronous acquisition, significantly reducing energy consumption and laying a solid foundation for accurate pollution detection. Simultaneously, the acquisition of events triggered by thresholds of logarithmic changes in illumination intensity accurately identifies sudden changes in reflectivity caused by pollution, reducing false triggers from invalid signals and improving the accuracy of pollution detection. Real-time filtering removes globally homogeneous events caused by ambient light fluctuations, separating pollution signals from environmental interference and improving the signal-to-noise ratio of the event stream. Combined with a dual-coordinate system mapping relationship, and assigning dynamic weighted priorities to pollution events based on their polarity, the system accurately locates the physical position of pollution, clearly defines the cleaning target, and avoids blind operations. Finally, the pre-processed event stream is output according to priority and forms a pollution hotspot tracking data packet sequence, ensuring that high-priority pollution events are processed first, reducing transmission latency, and effectively driving the robot to focus on highly polluted areas for efficient operation. This comprehensively optimizes the real-time performance, accuracy, and economy of photovoltaic cleaning.

[0122] In one possible implementation of the embodiments of this application, combined with Figure 1-6 As shown, the process of converting asynchronous data packets into a dynamic pollution level map can be achieved through the following steps 601 to 603, which are explained in detail below: Step 601: Parse the asynchronous data packets and calculate the pollution level to establish pollution level mapping rules.

[0123] Among them, the pollution level mapping rule is a quantitative standard that converts the frequency of events occurring per unit time into pollution level values.

[0124] In some implementations, the asynchronous data packets output by the event camera are first parsed to extract the event coordinates (x, y), microsecond-level timestamp t, and polarity parameter p. Then, by establishing a mapping relationship between the photovoltaic module's physical coordinate system and the sensor's pixel coordinate system, the pixel coordinates are converted into physical coordinates. Furthermore, dynamic priorities are assigned to pollution events based on the polarity parameter (negative polarity weighting for pollution events and positive polarity weighting for clean events). For example, if a pollution event, after coordinate mapping, is located at the lower right corner of the module (x=1.83m, y=0.91m), its negative polarity event weighting priority is increased to 148.5, ensuring that subsequent processing prioritizes responses to high-pollution signals.

[0125] Obtaining pollution detection thresholds and saturation threshold This allows for the direct establishment of the pollution level L and event frequency. Nonlinear piecewise mapping relationship between , where a and b are piecewise fitting coefficients. For example, when When the frequency is 20Hz, L = 65 is calculated; when When the frequency is 100Hz, L=100.

[0126] It should be noted that the pollution level mapping rule adopts a piecewise linear calibration strategy, setting differentiated parameters for different pollution types such as oil pollution (slope 1.2) and dust (slope 0.8) to improve the differentiation of pollution types.

[0127] Step 602: Use a sliding time window to perform frame processing on asynchronous data packets and extract the event frequency of each spatial coordinate.

[0128] In some implementations, a sliding time window (default 100ms, dynamically adjustable from 20-200ms) is used to segment the preprocessed event stream into frames. Within each time window, the number of event triggers at spatial coordinates (x, y) is counted. It can be passed by spatial coordinates. Calculate event trigger frequency Where T is the window duration. For example, if a region triggers 200 negative polarity events within a 50ms window, the frequency is calculated. =4Hz, indicating that there are high-frequency pollution events in this area.

[0129] It should be noted that the sliding time window size is adaptively adjusted according to the pollution diffusion rate. In dust storm scenarios, the window is shortened to 20ms, while in clear weather, the window is extended to 200ms, balancing response speed and noise suppression.

[0130] Step 603: Integrate the FPGA parallel computing platform to fuse historical pollution accumulation values ​​with real-time event frequencies to generate a dynamic pollution level map.

[0131] In some implementations, an FPGA parallel computing platform is integrated to combine the historical pollution accumulation value H(x,y) with the real-time event frequency. By weight To integrate. Among them... This is the attenuation coefficient (default 0.3). For example, if a region's historical pollution level H=70 and real-time frequency calculation yields L=85, after merging... =75.5. A dynamic pollution level map is ultimately generated, visually displaying the pollution distribution through a color gradient (red-yellow-green). Red areas (L≥80) identify high-pollution hotspots requiring priority cleaning. Based on the above technical solution, by parsing asynchronous data packets and establishing pollution level mapping rules, different pollution types can be accurately distinguished using coordinate mapping and polarity / frequency quantization, solving the problem of ambiguity in traditional pollution level classification and providing a quantitative basis for dynamic pollution tracking. Simultaneously, dynamic sliding time window frame processing is employed to extract the frequency of events at each spatial coordinate in real time, balancing the response speed to sudden pollution and noise suppression, thus solving the problem of lagging pollution change capture in traditional acquisition modes and achieving real-time dynamic pollution perception. Furthermore, an integrated FPGA parallel computing platform rapidly merges historical pollution accumulation values ​​with real-time frequencies to generate a dynamic pollution level map, reducing processing latency and addressing the poor dynamic adaptability of traditional serial computing. This provides real-time and comprehensive pollution distribution data for robots to accurately focus on highly polluted areas, reducing resource waste.

[0132] In one possible implementation of the embodiments of this application, combined with Figure 1-7 As shown, the process of dynamically updating the dynamic pollution map and the action selection competition mechanism can be implemented through the following steps 701 to 703, which are explained in detail below: Step 701: Real-time monitoring of changes in light intensity in the cleaning area triggers the generation of a reset event. Based on the polarity parameters of the reset event, the membrane potential of the dynamic pollution simulation neuron corresponding to the coordinates of the pollution treatment layer is reset.

[0133] Among them, the reset event refers to a special event stream triggered by the event camera after the cleaning action is completed. It includes the coordinates of the cleaning area, a timestamp, and polarity parameters, indicating that the contaminated area has been restored to a clean state. The dynamic contamination simulation neuron refers to the neuron node in the spiking neural network responsible for simulating contamination diffusion and cleanliness restoration. Its dynamic membrane potential change represents the degree of contamination. The membrane potential reset operation is the process of adjusting the neuron membrane potential according to the contamination state after cleaning, and it is divided into two modes: hard reset for complete cleaning and soft reset for residual contamination.

[0134] In some implementations, dynamic visual sensors are deployed to monitor changes in light intensity in the cleaning area in real time. When the detected change in reflectivity ΔR after cleaning exceeds a preset threshold (e.g., ΔR ≥ 20%), a reset event is triggered. This reset event includes the coordinates of the cleaning area (x, y), a timestamp t, and a polarity parameter p.

[0135] After the contamination treatment layer receives a reset event, it can locate the corresponding neuron N(x,y) using a lookup table of physical coordinates. If p is positive, the neuron's membrane potential V is directly reset to the initial baseline value V0; if p is negative, an exponential decay model is initiated. For example, when cleaning an oily area, if monitoring continues for 10 seconds after cleaning, the membrane potential will gradually decay from 45mV before cleaning to 45×e. (-0.03×10)=33.4mV, ultimately forming a dynamic pollution map update that matches the degree of residual pollution.

[0136] Step 702: Based on the change in the frequency of contamination events before and after cleaning, adjust the connection weights between neurons using the pulse time-dependent plasticity rule.

[0137] Among them, pulse time-dependent plasticity (STDP) refers to the synaptic weight adjustment rule in a spiking neural network based on the time difference between consecutive pulses, achieving dynamic weight optimization through ΔW=η·f(Δt). (The text also mentions changes in the frequency of pollution events.) The difference in the number of triggered events in the same spatial area per unit time before and after the cleaning action reflects the effectiveness of pollution removal. Interneuronal connection weights: a quantitative indicator of synaptic transmission efficiency in a spiking neural network, determining the spatiotemporal correlation in pollution diffusion simulation.

[0138] In some implementations, the system uses an event camera to monitor changes in light intensity in the cleaning area in real time and calculates the change in event frequency within a 100ms time window before and after cleaning. When the cleaning action effectively removes the contaminants ( >0), if the pulse time before cleaning Later than the pulse time after cleaning (Right now > Then, the weight adjustment is calculated according to the STDP rule. The learning rate η = 0.01. For example, if the event frequency before cleaning is 40Hz and drops to 15Hz after cleaning, then Δf = 25Hz, ΔW = 0.01(100-25) = 0.75mV, corresponding to a 0.75mV increase in the connection weight of adjacent neurons on the pollution propagation path, strengthening the spatiotemporal correlation of pollution diffusion simulation. If the pollution residue after cleaning ( <0), then The weights are adjusted in reverse to suppress the simulation of error propagation.

[0139] Step 703: Based on the cleaning action's performance, dynamically adjust the inhibition weight coefficient in the competition mechanism.

[0140] In some implementations, the polarity parameter of the reset event determines the effectiveness of the cleaning action; a positive polarity indicates complete removal of contaminants, while a negative polarity indicates residual contaminants. If three consecutive cleaning actions achieve a positive polarity (i.e., complete removal of contaminants), the suppression weight coefficient in the competition mechanism is dynamically adjusted. Updated to If negative polarity (i.e., cleaning residue) is present, then Updated to During the overall adjustment process, it is necessary to ensure that the γ value is within the effective range [0.5, 2.0], and take the corresponding extreme value when it exceeds the boundary.

[0141] It should be noted that the determination of complete removal of contamination three times in a row requires real-time updating of the count, and a single instance of residual contamination will trigger gamma downregulation. It can be set to 1.0 by default, or pre-configured according to the initial pollution level of the scene.

[0142] Based on the above technical solution, by capturing the abrupt change in reflectivity after cleaning to generate a reset event, and resetting the hard or soft graded membrane potential of neurons according to polarity, the problem of lagging updates in traditional pollution status is solved. This ensures that the dynamic pollution map is synchronized with the actual clean status in real time, providing accurate data support for path planning. Simultaneously, by comparing the frequency changes of pollution events before and after cleaning, the connection weights of neurons are dynamically adjusted using the pulse time-dependent plasticity rule, improving the accuracy of pollution diffusion and clean recovery simulations and avoiding invalid operations caused by model bias. Finally, the competitive mechanism suppresses the weight coefficients based on the cleaning effect, overcoming the rigidity of traditional decision-making mechanisms. This allows action decisions to adapt to the cleaning effect in real time, forming a feedback, optimization, and adaptation closed loop that progressively improves the accuracy and efficiency of photovoltaic cleaning while reducing water and energy waste and lowering power plant operation and maintenance costs.

[0143] The above primarily describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as a path planning system for a photovoltaic cleaning robot based on the degree of contamination, includes at least one of the hardware structures and software modules corresponding to the execution of each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in a hardware or computer software-driven hardware manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0144] When using integrated units, Figure 8The diagram illustrates a possible structural schematic of the path planning system for a photovoltaic cleaning robot based on contamination level, as described in the above embodiments. This system includes an event perception module, a pulse processing module, and an execution module. The event perception module collects and preprocesses the independent responses of each pixel unit to changes in light intensity, generating asynchronous data packet sequences and establishing a mapping relationship between the physical coordinate system of the photovoltaic module and the sensor pixel coordinate system, assigning dynamic priorities to contamination events. The pulse processing module constructs a three-level spiking neural network, outputting a distributed dynamic contamination map and an excitation pulse sequence for an action selection competition mechanism through the asynchronous data packet sequences. The execution module performs cleaning actions and executes state feedback and environmental change event streams based on the excitation pulse sequence of the action selection competition mechanism.

[0145] In one possible implementation, the three-level spiking neural network comprises an event input layer, a contamination processing layer, and an action output layer: The event input layer locates the neuron corresponding to the event through a fixed mapping relationship from physical coordinates to virtual neurons, calculates the membrane potential increment, and drives the contamination processing layer to update the membrane potential. The contamination processing layer calculates the potential increments of positive and negative events, generating a dynamic contamination level map. The action output layer selects the optimal action through a spiking competition mechanism; when the membrane potential of an action neuron reaches the threshold first, a cleaning instruction is triggered.

[0146] In one possible implementation, the construction of the event-aware module includes: deploying a dynamic visual sensor array, with each pixel unit independently responding to changes in light intensity; establishing a dynamic mapping relationship between the physical coordinate system of the photovoltaic module and the sensor pixel coordinate system, accelerating coordinate transformation via an FPGA; and integrating an event filter, employing a sliding time window and polarity weighting strategy.

[0147] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A path planning method for photovoltaic cleaning robots based on contamination levels, characterized in that, include: Collect the event stream of reflectivity changes on the surface of photovoltaic modules and generate asynchronous data packets containing event coordinates, timestamps, and polarity; The asynchronous data packets are converted into a dynamic pollution level map, and the pollution level value is positively correlated with the frequency of events occurring per unit time. A distributed dynamic pollution map is constructed using a spiking neural network, and the pollution diffusion and cleanup recovery process is simulated through the evolution of neuronal membrane potential. The construction process of the spiking neural network specifically includes: A spiking neural network structure is constructed, comprising an event input layer, a contamination processing layer, and an action output layer. The event input layer consists of multiple event response neurons, each corresponding to a specific coordinate position of the photovoltaic module, used to receive asynchronous data packets from the event camera and generate pulse signals. The contamination processing layer is composed of a dynamic contamination simulation neural network, where neurons simulate the accumulation and diffusion process of contamination through membrane potential evolution. The action output layer includes competitive action neurons, each encoding a basic cleaning action, and selecting the optimal action through a pulse competition mechanism. This competition mechanism is based on the dynamic evolution of membrane potential; when the membrane potential of a certain action neuron first reaches the firing threshold, the corresponding cleaning instruction is triggered. It integrates a dedicated neuromorphic computing chip, processes pulse streams in parallel, and embeds a dynamic reconfiguration interface to adapt to the spatial mapping of photovoltaic arrays of different sizes; The accuracy of contamination simulation is optimized by adjusting the connection weights between neurons through pulse timing. The process of constructing a distributed dynamic pollution map using a spiking neural network and simulating pollution diffusion and cleanup recovery through the evolution of neuronal membrane potential specifically includes: Based on the physical dimensions of the photovoltaic module, each event response neuron in the event input layer is mapped to the spatial coordinates of the photovoltaic module surface, establishing a fixed mapping relationship from physical coordinates to virtual neurons, and forming a lookup table from physical coordinates to neurons; The asynchronous data packet is input into the spiking neural network, and the corresponding event response neuron is found according to the event coordinates. The membrane potential increment is calculated and the contamination treatment layer is driven to update the membrane potential. Monitor the reflectivity of the photovoltaic module after cleaning, generate a reset event and send it to the contamination treatment layer, and update the dynamic contamination simulation neuron potential at the corresponding coordinate in the contamination treatment layer; Extract the neural network from the pollution treatment layer to form a pollution state distribution map; An action selection competition mechanism is established in the spiking neural network to generate cleaning path instructions based on the spatiotemporal distribution characteristics of the dynamic pollution map. The process of establishing an action selection competition mechanism in a spiking neural network and generating cleaning path instructions based on the spatiotemporal distribution characteristics of the dynamic pollution map specifically includes: A spatiotemporal feature coding layer is constructed, which includes feature extraction neurons connected to the neuron cluster in the pollution treatment layer, for receiving membrane potential pulse flow and calculating pollution accumulation intensity feature value and pollution diffusion trend feature value; The pollution accumulation intensity characteristic value and pollution diffusion trend characteristic value are converted into excitation pulse sequences for different action neurons in the action neuron array according to a preset action mapping rule; All action neurons in the action neuron array compete to fire parallel asynchronous pulses based on their current membrane potential, forming a cleaning path instruction. The system performs cleaning actions and provides feedback on environmental change events, dynamically updating the dynamic pollution map and the action selection competition mechanism.

2. The path planning method for photovoltaic cleaning robots based on contamination level according to claim 1, characterized in that, The process of collecting the reflectivity change event stream on the surface of the photovoltaic module and generating asynchronous data packets containing event coordinates, timestamps, and polarity specifically includes: A dynamic vision sensor is deployed, and the dynamic vision sensor operates in an asynchronous mode, with each pixel unit responding independently to changes in light intensity; When the logarithmic change in light intensity in a local area is detected to exceed a preset intensity threshold, an event is triggered to collect a reflectivity change event stream. The reflectivity change event stream is filtered in real time to remove global homopolar events caused by ambient light fluctuations, and then a preprocessed event stream is generated. Establish a mapping relationship between the physical coordinate system of photovoltaic modules and the pixel coordinate system of event cameras, and assign dynamic priorities and weights to pollution events based on event polarity parameters; The weighted preprocessed event stream is sorted by priority and then output to form an asynchronous data packet sequence for tracking pollution hotspots.

3. The path planning method for photovoltaic cleaning robots based on contamination level according to claim 2, characterized in that, The process of converting the asynchronous data packets into a dynamic pollution level graph specifically includes: The asynchronous data packets are parsed, and the pollution level is calculated to establish a pollution level mapping rule; A sliding time window is used to perform frame processing on asynchronous data packets, and the event frequency of each spatial coordinate is extracted; An integrated FPGA parallel computing platform is used to fuse historical pollution accumulation values ​​with real-time event frequencies to generate a dynamic pollution level map.

4. The path planning method for photovoltaic cleaning robots based on contamination level according to claim 3, characterized in that, The process of dynamically updating the dynamic pollution map and the action selection competition mechanism specifically includes: Real-time monitoring of changes in light intensity in the cleaning area triggers the generation of a reset event, and based on the polarity parameter of the reset event, the membrane potential of the dynamic pollution simulation neuron corresponding to the coordinate of the pollution treatment layer is reset. Based on the change in the frequency of contamination events before and after cleaning, the connection weights between neurons are adjusted by the pulse time-dependent plasticity rule. The inhibition weight coefficient in the competition mechanism is dynamically adjusted based on the effectiveness of the cleaning action.

5. A photovoltaic cleaning robot path planning system based on contamination level, characterized in that... The path planning method for photovoltaic cleaning robots based on contamination levels, as described in any one of claims 1-4, includes an event sensing module, a pulse processing module, and an execution module. The event perception module is used to collect the independent response of each pixel unit to changes in light intensity and perform preprocessing, generate an asynchronous data packet sequence, establish a mapping relationship between the physical coordinate system of the photovoltaic module and the sensor pixel coordinate system, and assign dynamic priority to pollution events. The pulse processing module constructs a three-level spiking neural network and outputs a distributed dynamic pollution map and an incentive pulse sequence for the action selection competition mechanism through the asynchronous data packet sequence; The execution module performs cleaning actions and executes state feedback and environmental change event streams based on the excitation pulse sequence of the action selection competition mechanism.

6. The photovoltaic cleaning robot path planning system based on contamination level according to claim 5, characterized in that, The three-level spiking neural network includes an event input layer, a contamination processing layer, and an action output layer: The event input layer locates the neuron corresponding to the event through a fixed mapping relationship from physical coordinates to virtual neurons, calculates the membrane potential increment, and drives the contamination treatment layer to update the membrane potential. The pollution treatment layer calculates the potential increments of positive and negative polarity events to generate a dynamic pollution level map. The action output layer selects the optimal action through a pulse competition mechanism. When the membrane potential of a certain action neuron reaches the threshold first, a cleaning instruction is triggered.

7. The photovoltaic cleaning robot path planning system based on contamination level according to claim 6, characterized in that, The construction of the event-aware module includes: Deploy a dynamic visual sensor array, with each pixel unit responding independently to changes in light intensity; Establish a dynamic mapping relationship between the physical coordinate system of photovoltaic modules and the coordinate system of sensor pixels, and accelerate coordinate transformation through FPGA; An integrated event filter is used, employing a sliding time window and polarity weighting strategy.

Images