A photovoltaic panel cleaning robot failure detection method and system

Abstract

The present application relates to photovoltaic power operation and maintenance technical field, especially in kind of photovoltaic panel cleaning robot fault detection method and system, including: based on the health state image of robot cleaning operation Multimodal time series data construction, then compare the ideal operation behavior model and generate the performance wear coefficient;Combined with the current cleaning task photovoltaic panel pollution characteristic parameters and environmental disturbance parameters, analyze the mapping relationship between performance wear coefficient and corresponding task scene, obtain scene driven component vulnerability atlas;Based on component vulnerability atlas and preset operation and maintenance knowledge base, deduce the expected load impact caused by executing subsequent planning task on each subsystem of robot, and generate decision instruction set accordingly, including maintenance priority and operation suggestion.Through the present application, the fault risk of photovoltaic panel cleaning robot executing future cleaning task can be evaluated in advance, preventive maintenance can be planned in advance, task interruption and unplanned downtime can be avoided, and photovoltaic power station operation and maintenance efficiency can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power operation and maintenance, and particularly relates to a photovoltaic panel cleaning robot fault detection method and system. BACKGROUND

[0002] With the wide application of photovoltaic power generation technology, the photovoltaic panel cleaning robot plays an increasingly important role in power station operation and maintenance. In existing operation and maintenance work, the state monitoring and fault management of the cleaning robot mainly rely on real-time data acquisition and threshold alarm mechanism. For example, by monitoring the motor current, sensor signal and other operating parameters, an alarm is triggered when the data exceeds the preset safety threshold, so that the operation and maintenance personnel can intervene and handle it.

[0003] The existing photovoltaic panel cleaning robot fault detection method can only issue an alarm when an abnormality or fault has occurred or is about to occur, and cannot make a forward-looking assessment of the performance degradation or fault risk that may be caused by the robot performing a specific cleaning task in the future before the fault occurs. It lacks the ability to associate the real-time health status of the robot, the historical performance degradation pattern and the specific task scenario to be performed, and to predict the cumulative impact of the task load on each subsystem of the robot accordingly. Therefore, the operation and maintenance decision often lags behind the actual loss, making it difficult to reasonably arrange the preventive maintenance window, which may cause unplanned downtime and interruption of cleaning tasks, affecting the overall operation and maintenance efficiency of the photovoltaic power station. SUMMARY

[0004] The present application provides a photovoltaic panel cleaning robot fault detection method and system, which can effectively solve the problems in the background art.

[0005] In order to achieve the above purpose, the technical solution adopted by the present application is:

[0006] A photovoltaic panel cleaning robot fault detection method, comprising:

[0007] Based on the multi-modal time series data generated by the cleaning robot during the execution of the cleaning operation, a health status portrait of the cleaning robot is constructed;

[0008] According to the deviation between the health status portrait and the pre-established ideal operation behavior model, an efficiency wear coefficient is generated;

[0009] Combined with the photovoltaic panel pollution characteristic parameters and environmental disturbance parameters of the current cleaning task, the mapping relationship between the efficiency wear coefficient and the corresponding task scenario is analyzed to obtain a scenario-driven component vulnerability map;

[0010] Based on the component vulnerability map and the pre-set operation and maintenance knowledge base, the expected load impact of executing subsequent planned tasks on each subsystem of the robot is deduced, and a decision instruction set is generated accordingly, including maintenance priority and operation suggestions.

[0011] Further, the multi-modal time-series data at least includes dynamics data, water circulation and brush body load data, and sensor raw signal quality data;

[0012] The dynamics data represents a mechanical execution state, the water circulation and brush body load data represents a cleaning efficiency, and the sensor raw signal quality data represents an environmental perception quality.

[0013] Further, the method for constructing the health state image comprises:

[0014] The dynamics data is subjected to joint frequency domain and amplitude domain analysis to extract frequency spectrum features and statistical features representing track walking smoothness, motor driving stability, and joint transmission efficiency;

[0015] The water circulation and brush body load data is subjected to process consistency analysis to determine whether the dynamic processes of cleaning liquid supply, sewage recovery, and rolling brush contact pressure conform to a preset cleaning operation mode;

[0016] The sensor raw signal quality data is subjected to signal-to-noise ratio and feature point stability evaluation to quantify the consistency of the laser radar, the confidence of the camera in identifying stains, and the fluctuation range of the ultrasonic ranging. Figure One

[0017] Further, the dynamics data includes three-phase current ripple of the driving motor, vibration acceleration frequency spectrum of the walking wheel system, and torque following error of each joint servo motor;

[0018] The water circulation and brush body load data includes water pump outlet pressure pulsation, water tank liquid level change rate, and power spectral density of the cleaning rolling brush motor;

[0019] The sensor raw signal quality data includes plane fitting residual error of the laser radar point cloud, local contrast gradient of the camera image, and energy attenuation curve of the ultrasonic echo signal.

[0020] Further, the photovoltaic panel pollution feature parameters are obtained through pre-task scanning, including pollution type identification results, pollution area area proportion, and pollution adhesion intensity grade;

[0021] The environmental disturbance parameters include real-time wind speed, environmental dust concentration, and photovoltaic panel surface temperature.

[0022] Further, the scene-driven component vulnerability map is obtained, comprising:

[0023] The efficiency wear coefficient is decomposed into a driving module, a cleaning module, a perception module, and an energy module;

[0024] ​For different combinations of photovoltaic panel pollution characteristic parameters and environmental disturbance parameters in historical tasks, the change patterns and sensitivity of the performance wear coefficients of each module are counted;

[0025] A prediction relationship network is established, taking pollution type, pollution degree, wind speed level, and temperature interval as input, and expected wear increment and main failure mode of each module as output, as a component vulnerability atlas.

[0026] Further, the expected load impact of subsequent planning tasks on each subsystem of the robot is deduced, including:

[0027] For subsequent planning tasks, the estimated photovoltaic panel pollution characteristic parameters and environmental disturbance parameters are extracted according to the task description;

[0028] The estimated photovoltaic panel pollution characteristic parameters and environmental disturbance parameters are input into the component vulnerability atlas to obtain the expected wear increment of the driving module, cleaning module, sensing module, and energy module;

[0029] Combined with the current health state portrait of each module of the robot, the expected wear increment is accumulated to predict the state trajectory of each module after executing the planning task;

[0030] When the predicted state trajectory reaches the preset performance boundary, it is determined that the module is subjected to load impact.

[0031] Further, the ideal operation behavior model is constructed based on the multi-modal time series data of the robot in the initial state without loss at the factory, and is iteratively corrected based on historical statistical data of the same type of robot without failure.

[0032] Further, the operation and maintenance knowledge base is constructed based on historical operation and maintenance data of the same type of robot, component failure handling specifications, and power station operation and maintenance processes, including module failure impact weight, maintenance process standards, and spare parts adaptation information.

[0033] On the other hand, the present application also provides a photovoltaic panel cleaning robot fault detection system, comprising:

[0034] A health portrait construction module is used to construct a health state portrait of the cleaning robot based on multi-modal time series data generated by the cleaning robot during cleaning operation;

[0035] A performance wear coefficient module is used to generate a performance wear coefficient according to the deviation between the health state portrait and the pre-established ideal operation behavior model;

[0036] A vulnerability atlas module is used to analyze the mapping relationship between the performance wear coefficient and the corresponding task scenario by combining the photovoltaic panel pollution characteristic parameters and environmental disturbance parameters of the current cleaning task, and obtain a scenario-driven component vulnerability atlas;

[0037] An operation and maintenance decision generation module is configured to deduce expected load impact on each subsystem of the robot caused by subsequent planning tasks based on the component vulnerability map and the preset operation and maintenance knowledge base, and generate a decision instruction set including maintenance priority and operation suggestions accordingly.

[0038] The technical scheme of the present application can achieve the following technical effects:

[0039] By multi-dimensional analysis of the robot multi-modal time series data, a comprehensive health status portrait can be constructed to accurately quantify the actual running state and performance degradation of each subsystem of the robot. By deeply correlating the efficiency wear coefficient with the photovoltaic panel pollution characteristics, environmental disturbance and other task scene parameters, a scene-driven component vulnerability map can be established to realize the organic combination of the real-time health status, historical performance degradation mode and specific task scene of the robot. Based on the map, the expected load impact of subsequent planning tasks on the driving, cleaning, sensing and energy modules can be accurately deduced, the state trajectory of each module after executing the task can be predicted, the fault risk points can be identified in advance, and a decision instruction set including maintenance priority and operation suggestions can be generated to help operation and maintenance personnel to reasonably plan preventive maintenance windows in advance, avoid unplanned downtime and cleaning task interruption, and improve the scientific nature and pertinence of photovoltaic panel cleaning robot operation and maintenance decisions, thereby ensuring the stable operation of the photovoltaic power station from the operation and maintenance end, and ultimately improving the overall operation and maintenance efficiency and economic benefits of the photovoltaic power station.

[0040] The above description is only a summary of the technical scheme of the present application. In order to more clearly understand the technical means of the present application, the specific embodiments of the present application can be implemented in accordance with the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following will specifically describe the embodiments of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical scheme in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments described in the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0042] Figure 1 The flowchart of the photovoltaic panel cleaning robot fault detection method of the present application is shown in the figure;

[0043] Figure 2 The structure diagram of the photovoltaic panel cleaning robot fault detection system of the present application is shown in the figure; DETAILED DESCRIPTION

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0046] like Figure 1 As shown, a fault detection method for a photovoltaic panel cleaning robot according to the present invention specifically includes the following steps:

[0047] Step S1: Based on the multimodal time-series data generated by the cleaning robot during cleaning operations, construct a health status profile of the cleaning robot;

[0048] Step S2: Generate the performance wear coefficient based on the degree of deviation between the health status profile and the pre-established ideal work behavior model;

[0049] Step S3: Combining the photovoltaic panel contamination characteristic parameters and environmental disturbance parameters of the current cleaning task, analyze the mapping relationship between the efficiency wear coefficient and the corresponding task scenario to obtain the scenario-driven component vulnerability map.

[0050] Step S4: Based on the component vulnerability map and the preset operation and maintenance knowledge base, deduce the expected load impact on each subsystem of the robot caused by the execution of subsequent planned tasks, and generate a set of decision instructions, including maintenance priorities and operation suggestions.

[0051] In this embodiment, multi-stage collaborative logic can predict the cumulative impact of specific tasks on various subsystems in advance, accurately locating components prone to performance degradation in different scenarios. Based on this prediction, maintenance priorities and operational suggestions can ensure that maintenance decisions align with the actual wear and tear of the robot, rationally plan preventative maintenance windows, avoid unplanned downtime and interruptions to cleaning tasks, ensure the continuity of photovoltaic power plant maintenance processes, and create a positive linkage between robot operation efficiency and power plant maintenance efficiency. Through precise mapping of scenario characteristics and component vulnerabilities, maintenance actions can be more targeted, reducing ineffective maintenance investment, while also reducing irreversible wear and tear on robot subsystems caused by task load, extending the stable operation cycle of equipment, and solving the problem of maintenance decisions lagging behind actual wear and tear.

[0052] In a specific implementation, as an embodiment, since the performance degradation of the photovoltaic panel cleaning robot is the result of the joint action of multiple factors, the state of each component of the robot and the operating environment data need to be obtained, while ensuring the time sequence continuity of the data and the relevance to the work scene, and the health state portrait of the cleaning robot is constructed; the embodiment realizes the multi-modal time sequence data recording of the target robot and the construction of the health state portrait through multi-source data collaborative collection and deep processing, as follows:

[0053] In step S11, multi-modal time sequence data of the cleaning robot during work is obtained, which corresponds to the actual running state of each functional module of the robot; the obtained data is selected according to the principles of covering key working conditions, quantifiable state representation and reflecting early wear, and the collection frequency needs to match the dynamic change characteristics of various data to ensure time sequence continuity; at the same time, a timestamp and a work position marker are added to each type of data to ensure the correspondence between the data and the actual work scene; the multi-modal time sequence data at least includes:

[0054] a. Kinetic data: used to represent the mechanical execution state, which can select the three-phase current ripple of the drive motor, the vibration acceleration spectrum of the walking wheel system and the torque following error of each joint servo motor, which is a direct mapping index of the running state of the robot mechanical system; among them, the three-phase current ripple of the drive motor can reflect the motor winding state, the inverter working stability and the load fluctuation, compared with a single current value, the ripple signal can capture the early insulation aging of the motor, rotor eccentricity and other hidden problems, which is collected by the current sensor built-in the motor controller, covering the whole work stage of motor starting, uniform running and steering; the vibration acceleration spectrum of the walking wheel system reflects the track walking smoothness, and the wheel system wear and track attachment foreign matter will cause characteristic frequency components in the vibration spectrum, which is collected by the three-axis acceleration sensor installed at the wheel system bearing, which can capture wear characteristics in low-frequency vibration and impact signals in high-frequency vibration; the torque following error of each joint servo motor reflects the joint transmission efficiency, and the transmission gear wear and insufficient lubrication will increase the following error, which is obtained by real-time reading the difference between the command torque and the actual output torque of the servo driver, and the corresponding joint action time sequence is recorded synchronously;

[0055] b. Water Circulation and Brush Load Data: Used to characterize cleaning efficiency, this includes selecting water pump outlet pressure pulsation, water tank level change rate, and cleaning roller brush motor power spectral density, directly correlated with the robot's cleaning module operation. Specifically, water pump outlet pressure pulsation reflects the stability of the cleaning fluid supply; pipe blockage and valve wear can cause abnormal pressure pulsation amplitude and frequency, collected by a high-frequency pressure sensor installed at the water pump outlet. Water tank level change rate reflects the balance between cleaning fluid consumption and recovery; abnormal level change rates may originate from pipe leaks or recovery system malfunctions, continuously collected by a built-in water tank level sensor to calculate the level change per unit time. Cleaning roller brush motor power spectral density reflects the brush load status; roller brush wear and changes in contaminant adhesion intensity can cause power spectral characteristic shifts, collected by the roller brush motor controller in real-time, converted into power spectral density data through spectrum analysis, capturing the frequency characteristics of load changes.

[0056] c. Raw sensor signal quality data: Used to characterize the quality of environmental perception. This can include the residuals of the lidar point cloud plane fitting, the local contrast gradient of the camera image, and the energy attenuation curve of the ultrasonic echo signal. This avoids relying solely on sensor output results while ignoring the reliability of the signal itself. Among these, the lidar point cloud plane fitting residuals reflect the overall system performance. Figure One To improve the accuracy of point cloud data fitting, dust and sensor lens contamination can increase the fitting residual. After acquiring point cloud data from the photovoltaic panel surface using LiDAR, planar fitting is performed on local areas, and the residual is calculated to quantify the mapping accuracy. The local contrast gradient of the camera image reflects the confidence level of stain recognition. Lens contamination and changes in lighting can reduce image contrast, affecting the accuracy of stain recognition. The contrast gradient value of the region of interest is extracted from the original image acquired by the camera to reflect the image signal quality. The ultrasonic echo signal energy attenuation curve reflects the fluctuation range of ultrasonic ranging. Air humidity and surface contamination can affect the attenuation pattern of echo energy. The attenuation process of energy over propagation time is recorded by acquiring echo signals using an ultrasonic sensor to quantify signal stability.

[0057] Step S12, frequency domain and amplitude domain joint analysis is performed on the kinetic data to make up for the limitations of single domain analysis; frequency domain analysis can use fast Fourier transform to convert time domain signals of three-phase current ripple, vibration acceleration and torque following error into frequency domain signals, extract characteristic frequency components and amplitude values, for example, harmonic amplitude values near 50Hz reflect the degree of electromagnetic interference of the motor, and vibration peak values in a specific frequency range correspond to wheel train wear characteristics; amplitude domain analysis extracts signal peak value, mean value, variance, kurtosis and other statistical characteristics, among which kurtosis can effectively identify the impact component in the vibration signal and capture the instantaneous abnormal contact between the running wheel train and the track; through the fusion of frequency domain and amplitude domain characteristics, the vibration frequency spectrum characteristics and the amplitude domain kurtosis are used to quantify the track running smoothness, the current ripple characteristic frequency amplitude and the statistical variance are used to quantify the motor drive stability, and the torque following error mean value and the frequency domain harmonic content are used to quantify the joint transmission efficiency, forming a kinetic dimension feature set;

[0058] Step S13, process consistency analysis is performed on the water circulation and brush body load data, and a dynamic process evaluation standard is constructed based on the preset cleaning operation mode; the preset mode is constructed based on historical operation data of the robot under ideal working conditions, which is cleaning photovoltaic panels without environmental interference, and includes the normal fluctuation range of water pump outlet pressure pulsation, the matching relationship between liquid level change rate and operation progress, and the baseline curve of roller brush motor power spectral density; during the analysis, the deviation of the current water pump pressure pulsation from the baseline range is compared in real time to determine whether the cleaning liquid supply is continuous and stable; the operation area and time are combined to calibrate the water tank liquid level change rate, and if the calibrated value deviates, it is determined that there is an abnormality in the supply or recovery system; the roller brush motor power spectral density is compared with the baseline curve to extract the spectral peak shift and amplitude difference, and it is determined whether the roller brush load meets the cleaning operation requirements, forming a cleaning efficiency dimension feature set;

[0059] Step S14, signal-to-noise ratio and feature point stability evaluation are performed on the sensor original signal quality data to quantify the environmental perception quality; the signal-to-noise ratio calculation uses the ratio of signal effective amplitude to noise amplitude, among which the laser radar point cloud signal noise is extracted by background point cloud filtering, the camera image noise is calibrated by dark field acquisition, and the ultrasonic wave signal noise is calibrated in a non-reflective environment; the feature point stability evaluation is performed on the core feature parameters of each sensor, the laser radar consistency is quantified by the variation coefficient of plane fitting residual, the smaller the variation coefficient, the more stable the mapping; the camera stain recognition confidence is represented by the uniformity of local contrast gradient distribution, the more uniform the gradient distribution, the better the image quality, and the more conducive to stain recognition; the ultrasonic ranging fluctuation range is quantified by the fitting accuracy of the echo signal energy decay curve, the smaller the fitting error, the more stable the ranging, forming an environmental perception dimension feature set; Figure One

[0060] ​Step S15, the characteristics set of the three dimensions of kinetics, cleaning efficiency and environmental perception are fused to construct a complete health status portrait; when fusing, cross-dimension feature association rules are first established, and the association weight is determined based on historical operation data statistics, and the influence degree is allocated according to the influence of each feature on the performance degradation of the corresponding subsystem, and the influence degree is marked by the correspondence between the feature and the historical failure case; redundant features are removed by calculating the correlation between features, when the contribution rate of two features to the same working condition overlaps more than a certain threshold, the feature more sensitive to early performance degradation is retained, and the core feature capable of capturing implicit wear is preferentially retained; the portrait is presented in the form of a feature matrix, the rows of the matrix correspond to the time sequence collection nodes, and the time interval of each node is consistent with the data collection frequency to ensure the time sequence continuity; the columns correspond to the core features of each dimension, and each feature value is standardized to eliminate the influence of dimension difference on the representation result; through the feature matrix, the real-time state and time sequence change trend of each subsystem of the robot can be simultaneously reflected, and the quantitative representation of the comprehensive working condition of the robot is realized.

[0061] In the embodiment, by synchronously collecting multi-modal time sequence data related to the operation of each core module of the robot and the environment, the health status portrait is constructed through dimension-by-dimension deep processing and feature fusion, and the working conditions such as mechanical execution, cleaning efficiency and environmental perception are accurately quantified; the embodiment does not rely on a single parameter threshold to determine, but links the operation state of each functional module of the robot and the environmental data, mines the implicit wear characteristics of the data through targeted processing, realizes the dynamic and comprehensive quantitative representation of the comprehensive working condition of the robot, and does not only trigger an alarm when a fault occurs or is about to occur.

[0062] In some embodiments of the present application, the internal wear of the robot in the prior art cannot be directly measured, there is a lack of means for indirect quantification relying on the deviation relationship between the health status portrait and the reference behavior model, and there is no ideal model supporting the actual operation, making it difficult to generate a unified index accurately representing the wear; based on the above problems, the embodiment pre-establishes an ideal operation behavior model that fits the actual operation of the robot, and then generates an efficiency wear coefficient through multi-dimensional deviation analysis to quantify the deviation degree of the health status portrait from the model, maps the multi-dimensional state characteristics to a single efficiency wear coefficient, and accurately represents the internal wear; the following operations are performed:

[0063] Step S21, the pre-established ideal operation behavior model is called, the model is based on the operation data of the robot in the initial state without wear at the factory, and is optimized by combining the historical statistical data of the fault-free operation of the robot of the same type, to ensure that the model fits the original operation ability of the robot; the model corresponds to the feature dimensions of the health status portrait one by one, and each dimension includes the reference range and time sequence change law of each feature; the construction process of the ideal operation behavior model is as follows:

[0064] The multi-modal time series data of the robot under no load and rated load within a preset time period after the robot leaves the factory is selected as the basic data source. The time period is set to ensure that the components are not worn and the performance is at the peak value, and the data can fully reflect the original working state of the robot; for the dynamics dimension, the reference spectrum of the three-phase current ripple of the drive motor, the normal amplitude range of the walking wheel system vibration acceleration, and the torque following error of the joint servo motor are recorded, and the characteristic change law of different working stages is labeled; for the water circulation and brush body load dimension, the reference fluctuation amplitude of the water pump outlet pressure fluctuation, the matching ratio of the water tank liquid level change rate and the working area, and the reference curve of the cleaning brush motor power spectral density are determined; for the sensor original signal quality dimension, the reference value of the laser radar point cloud plane fitting residual, the normal distribution range of the local contrast gradient of the camera image, and the standard curve of the ultrasonic wave echo signal energy attenuation are calibrated; the above reference parameters are iteratively corrected by using the historical data of the same type of robot running without failure, and the abnormal data in extreme environment is eliminated to ensure the stability and universality of the model;

[0065] In step S22, based on the health state portrait feature matrix, corresponding dimensions are compared and analyzed with the ideal model. Dimension deviation calculation and cross-dimension fusion can be used to avoid misjudgment caused by single feature deviation. In each dimension, the deviation of each feature value in the health state portrait from the ideal model reference range is calculated, and then combined with the influence weight of each feature on the corresponding subsystem loss to obtain the comprehensive deviation degree of this dimension. The deviation is represented by the difference between the feature value and the reference range. The larger the difference ratio is, the higher the feature deviation degree is. The influence weight is calibrated based on historical failure cases, and the probability and severity of failure caused by different feature deviations are statistically distributed, for example, the probability of mechanical failure caused by the deviation of the three-phase current ripple of the drive motor is higher than that of the deviation of the camera image contrast gradient, and the corresponding influence weight is higher. After the comprehensive deviation degrees of the dynamics dimension, the water circulation and brush body load dimension, and the sensor original signal quality dimension are calculated, the cross-dimension fusion is performed according to the influence proportion of each dimension on the overall performance of the robot to obtain the overall deviation degree value of the robot. The influence proportion of each dimension is determined by the statistical duration of the failure shutdown of each subsystem of the robot.

[0066] Step S23, generating an efficiency wear coefficient based on the overall deviation degree value of the robot, the value range of the efficiency wear coefficient corresponding to the internal wear level of the robot, the coefficient value gradually increasing with the wear level from no wear to severe wear, and the coefficient value being in a positive correlation mapping relationship with the overall deviation degree value; the mapping rule is determined by training historical data, specifically, collecting actual wear data of internal components of the robot under different deviation degrees, establishing a corresponding relationship between the deviation degree value and the wear amount, and converting the wear amount into a standardized efficiency wear coefficient; during the coefficient generation process, the cumulative working time of the robot needs to be corrected, under the same deviation degree, the longer the cumulative working time, the higher the efficiency wear coefficient, which fits the cumulative characteristics of component wear; at the same time, dynamic calibration is carried out for different working scenes, for example, the wear amount corresponding to the same deviation degree in a high temperature environment is higher than that in a normal temperature environment, and the scene coefficient is adjusted to ensure that the efficiency wear coefficient accurately reflects the actual internal wear; the efficiency wear coefficient is finally output in the form of a single numerical value, and the larger the numerical value, the more serious the internal wear; at the same time, the contribution proportion of each dimension deviation to the coefficient is attached, and the main wear source is clear.

[0067] In the embodiment, the ideal model is constructed based on the original state of the robot and historical fault-free data, and the deviation quantifiable baseline reliability can be obtained; the deviation degree can be calculated by dimension fusion, the single feature misjudgment can be avoided, and the wear evaluation accuracy can be improved; the efficiency wear coefficient realizes the conversion of multi-dimensional state features to internal wear, provides a unified index for analyzing the mapping relationship between wear and task scene; the coefficient is attached with the contribution proportion of the wear source, which can directly locate the main wear component and provide a clear direction for maintenance; combined with the dynamic calibration of time and scene, the wear representation is more suitable for actual operation conditions.

[0068] In specific implementation, as an embodiment, the existing technology cannot associate the overall wear of the robot with the specific working scene, and it is more difficult to locate the vulnerability of each module under different scenes, resulting in lack of pertinence of preventive maintenance; based on this, the embodiment combines the current task scene parameters, disassembles the efficiency wear coefficient to each functional module, establishes the mapping relationship between the scene and the module wear through historical data statistics, constructs a scene-driven component vulnerability map, and specifically implements the following steps:

[0069] Step S31, obtaining the photovoltaic panel pollution characteristic parameters and environmental disturbance parameters of the current cleaning task, both types of parameters directly determine the working load intensity and the wear impact on each module; wherein, the photovoltaic panel pollution characteristic parameters are obtained by scanning before the task, the scanning range should cover all the photovoltaic panel areas of the current task to avoid local pollution omission leading to scene representation deviation; the photovoltaic panel pollution characteristic parameters include:

[0070] a. Pollutant type identification result: Differentiate dust, oil, bird droppings, fallen leaves and other types through visual image gray scale and texture features; Different pollutants have different effects on the identification accuracy of the cleaning module load and the sensing module, for example, oil requires greater roller brush pressure, and dust easily interferes with sensor signals;

[0071] b. Polluted area area ratio: Reflect the total workload through laser radar scanning data segmentation and statistics, the higher the ratio, the longer the continuous running time of the cleaning module and the driving module, and the faster the cumulative loss;

[0072] c. Pollutant adhesion strength grade: Determine through motor load feedback during roller brush pre-contact, the higher the adhesion strength, the greater the roller brush motor load and water circulation system pressure, and the more significant the cleaning module loss;

[0073] After scanning, the parameters are standardized to unify the data format and magnitude, which is convenient for correlation analysis with the efficiency wear coefficient;

[0074] Environmental disturbance parameters are collected in real time by external sensors carried by the robot, and the collection time and work position are recorded synchronously to ensure the spatiotemporal correspondence with pollution characteristic parameters and efficiency wear coefficients; Environmental disturbance parameters include:

[0075] a. Real-time wind speed: Collected by wind speed sensor, wind speed affects the stability of driving module load; Under high wind speed, the robot needs to increase driving force to keep walking smoothly, which can increase the wear of driving motor and wheel train;

[0076] b. Ambient dust concentration: Collected by dust sensor, dust particles can increase the load of cleaning module and adhere to the sensor lens to affect the sensing accuracy and increase the wear of sensing module;

[0077] c. Photovoltaic panel surface temperature: Collected by infrared sensor, high or low temperature can affect the power supply efficiency of energy module and the stability of motor operation, indirectly increasing the wear of each module;

[0078] Data filtering is required during collection to eliminate transient extreme values and retain continuous stable parameter sequences to ensure the reliability of scene disturbance representation;

[0079] Step S32, decompose the performance wear coefficient to the driving module, the cleaning module, the sensing module and the energy module; the decomposition is based on the correspondence between the three-dimensional features based on the health state image and each module, combined with the contribution proportion of each module to the overall loss to determine, to ensure that the decomposition result fits the actual loss distribution; wherein the driving module corresponds to the kinetic dimension feature, the cleaning module corresponds to the water circulation and brush body load dimension feature, the sensing module corresponds to the sensor original signal quality dimension feature, the energy module loss is indirectly represented by the energy consumption change of each module, which is calculated based on the operating current and voltage data of each module, reflecting the power supply system load and efficiency attenuation;

[0080] In the decomposition process, first, the historical fault data is used to calibrate the correlation weight of each dimension feature to the loss of the corresponding module, for example, the contribution proportion of the current ripple deviation of the driving motor in the kinetic dimension to the loss of the driving module is higher than that of the joint torque following error, and the corresponding weight is higher; then, according to the contribution degree of each dimension feature to the performance wear coefficient, the coefficient is decomposed to the corresponding module in proportion to the weight, to obtain the sub-module performance wear coefficient of each module; at the same time, the decomposition result is preliminarily corrected combined with the current scene parameters, for example, in high wind speed scene, the driving module sub-coefficient is appropriately adjusted upwards, which fits the additional loss impact of wind speed on the driving module, to ensure that the decomposition result is adapted to the scene;

[0081] Step S33, for different combinations of photovoltaic panel pollution feature parameters and environmental disturbance parameters in historical tasks, the change mode and sensitivity of the performance wear coefficient of each module are counted; the historical data selects complete task data of the same type of robot in the same type of work scene, covering different pollution types, pollution levels, wind speed levels, temperature intervals, and eliminates data under equipment failure or extreme abnormal environment, to ensure the effectiveness of the statistical sample;

[0082] In the statistical process, the data is classified and arranged according to the scene parameter combination, for example, dust pollution + high proportion + low adhesion intensity + high wind speed + normal temperature; oil pollution + medium proportion + high adhesion intensity + low wind speed + high temperature, etc.; for each combination, the change trend of each module sub-coefficient with the task progress is tracked, and the change mode is summarized, for example, in the high adhesion intensity pollution combination, the cleaning module sub-coefficient increases linearly with the work time, and the driving module sub-coefficient increases at a flat rate; in the high wind speed combination, the driving module sub-coefficient increases at a higher rate than other modules; the sensitivity is determined by the influence amplitude of the same parameter change on the sub-coefficients of different modules, for example, when the dust concentration increases, the change amplitude of the sensing module sub-coefficient is greater than that of the energy module, indicating that the sensing module is more sensitive to dust; when the temperature exceeds the normal temperature interval, the change amplitude of the energy module sub-coefficient is the largest, reflecting the sensitivity of the energy module to temperature; after the statistics is completed, the corresponding data set of scene combination and module loss rule is formed;

[0083] Step S34, a prediction relationship network is established, taking the pollution type, pollution degree, wind speed level, and temperature interval as inputs, and taking the expected wear increment and main failure mode of each module as outputs, as a scene-driven component vulnerability map;

[0084] The network is constructed based on the aforementioned historical statistical data set, and the mapping relationship between the scene parameters and the module loss is fitted through sample training; the standardized pollution type, pollution degree, wind speed level, and temperature interval are taken as the core features at the input end, wherein the pollution degree is represented by the comprehensive representation of the pollution area ratio and the adhesion strength level, and the wind speed level and the temperature interval are divided into intervals according to the industry general standard; the output end corresponds to each module, and outputs the expected wear increment and the main failure mode of each module, the expected wear increment represents the increase of the module loss after completing a single task in this scene, and the main failure mode is determined based on historical data statistics, for example, the main failure mode of the cleaning module in the high adhesion strength pollution scene is the overloading of the brush motor, and the main failure mode of the driving module in the high wind speed scene is the aggravation of the wheel train wear; after the network training is completed, the prediction deviation is corrected through new task data iteration optimization, to ensure the adaptability of the map to different scenes;

[0085] The vulnerability map is finally presented in a structured form, labeling the wear risk and failure tendency of each module under different scene combinations, to realize the mapping from scene to module vulnerability.

[0086] In this embodiment, the scene parameter acquisition adopts a combination of collaborative scanning and real-time sensing to ensure accurate scene representation; the efficiency wear coefficient module is decomposed to match the actual loss of each module, to realize the extension of loss positioning from the whole to the local; the historical data statistics cover multiple scene combinations, so that the prediction relationship network has scene adaptation capability; the vulnerability map provides the expected wear increment and failure mode of the module level, so that the operation and maintenance personnel can accurately master the high-risk components in different scenes, facilitate targeted preventive maintenance, and reduce unplanned downtime.

[0087] In a specific implementation, as an embodiment, since the load impact needs to be deduced in combination with the subsequent task scene and the current state of the robot, and the decision instruction needs to take into account the risk level and the operation and maintenance feasibility, the task estimation parameters need to be extracted first, and then the map and the knowledge base are linked to complete the deduction and instruction generation; this embodiment realizes the deduction of the load impact of the subsequent task through the scene-driven component vulnerability map, combines the preset operation and maintenance knowledge base, judges the impact risk through state trajectory prediction, and finally generates a decision instruction set containing the maintenance priority and operation suggestions, to build a complete closed loop from risk prediction to operation and maintenance landing, and the specific implementation steps are as follows:

[0088] Step S41, extract the estimated pollution characteristic parameters of the photovoltaic panel and the environmental disturbance parameters of the subsequent planning task; the parameter extraction is based on the task description and combined with the historical similar task scene data correction to ensure that the estimation accuracy is consistent with the actual operation scene, providing reliable input for load impact deduction; wherein the task description needs to clearly cover the operation range, photovoltaic panel array layout, estimated operation time and other information, and the scene boundary is preliminarily defined based on the information;

[0089] The estimation of the pollution characteristic parameters of the photovoltaic panel adopts a combination of historical scene matching and on-site investigation, first retrieves the historical task data similar to the operation range, season and climate condition of the current task from the operation knowledge base, extracts the statistical values of the corresponding pollution type, pollution area area ratio and adhesion strength grade as the initial estimation basis, and then uses the scanning components carried by the robot to conduct on-site rapid investigation, covering the key points of the operation area, correcting the initial estimation parameters and avoiding the deviation of historical data from the current actual scene; for example, in the dry season outdoor power station task, the historical scene parameters dominated by dust pollution in the same season are preferentially matched, and then it is confirmed whether there is local oil pollution or bird droppings pollution through investigation, and then the parameter ratio and strength grade are adjusted;

[0090] The environmental disturbance parameter estimation relies on the real-time meteorological data of the power station operation platform and the historical same period statistical rule, the real-time meteorological data is used to obtain the wind speed and temperature forecast value of the power station deployment area in the future operation period, and the historical same period data is used to correct the forecast deviation, for example, the gust often occurs in the afternoon in a certain area, which can be combined with the historical data to increase the estimated wind speed fluctuation amplitude; the environmental dust concentration is derived based on the wind speed estimation result and the distribution of surrounding pollution sources, the higher the wind speed and the closer the surrounding pollution sources, the higher the estimated dust concentration;

[0091] All estimated parameters are standardized to ensure compatibility with the input end of the component vulnerability map;

[0092] Step S42, input the standardized estimated parameters into the component vulnerability map, the map outputs the expected wear increment of the driving module, cleaning module, sensing module and energy module based on the preset prediction relationship network, and simultaneously outputs the main failure mode prompt of each module; the expected wear increment directly corresponds to the newly added amount of wear of each module in the current task scene, and the failure mode prompt provides a basis for subsequent risk judgment and operation suggestion, for example, if the map outputs a high expected wear increment of the cleaning module and prompts an overload risk of the brush motor, the state change of the module needs to be tracked;

[0093] Step S43, in combination with the current health state image of each module of the robot, the expected wear increment is accumulated to predict the state trajectory of each module after performing the planned task; the state trajectory presents the change trend of the module-by-module performance wear coefficient and the key node value synchronously with time sequence as the axis; the trajectory drawing needs to be consistent with the cumulative characteristics of the loss of each module, for example, the energy module loss grows nonlinearly with the working time, and the trajectory needs to reflect this law; the three-dimensional features of the health state image need to be associated during the accumulation process to correct abnormal fluctuations, for example, there is a slight signal deviation in the current perception module, and the influence of the deviation on the trajectory needs to be appropriately amplified when accumulating the expected wear increment to ensure that the trajectory is consistent with the actual state of the module;

[0094] Step S44, preset the performance boundary as the load impact judgment reference, the boundary value is calibrated based on the robot factory performance standard and historical failure data, and is set for each module respectively, and the performance boundaries of different modules are adapted to their functional positioning and loss tolerance ability; for example, the perception module sets the boundary based on the signal recognition accuracy threshold; when the predicted module state trajectory reaches or exceeds the preset performance boundary, it is determined that the module is subjected to load impact, and the estimated task node and associated failure mode of the impact occurrence are recorded to provide accurate risk information for decision instruction generation;

[0095] Step S45, based on the load impact deduction result and the preset operation and maintenance knowledge base, a set of decision instructions is generated, which includes maintenance priority and operation suggestions, taking into account the risk urgency and operation and maintenance operability to ensure that the instructions can directly guide the on-site operation; the preset operation and maintenance knowledge base is constructed based on historical operation and maintenance data of the same type of robot, component failure handling specifications, and power station operation and maintenance processes, which can include module failure impact weight, maintenance process standard, spare part adaptation information, etc.; among them, the module failure impact weight is determined by the downtime length and maintenance cost statistics caused by historical failures, for example, the drive module failure causes the highest downtime proportion, so the impact weight is higher than that of other modules; the maintenance process standard clearly specifies the standard processing flow of different failure modes, for example, the brush motor overload needs to perform load debugging and lubrication maintenance, and the wheel train wear needs to be detected and replaced; the spare part adaptation information records the model, adaptation condition and replacement period of the core components of each module;

[0096] The maintenance priority is ranked according to the impact risk level and the module impact weight, the priority of the module subjected to the load impact is higher than that of the module not subjected to the impact, and under the same impact level, the priority of the module with a large impact weight is higher; for example, the driving module and the sensing module are both subjected to the impact, and the driving module has a high priority; the operation suggestion is formulated according to the estimated failure mode and the impact node of each module, and combined with the maintenance process standard in the knowledge base, the specific operation content, execution time and key requirements are clear, for example, for the risk of overload of the roller brush motor of the cleaning module, it is suggested to adjust the roller brush pressure before the task is executed, to monitor the load every interval of a preset time length during the operation, and to perform motor inspection after the task; for the temperature-sensitive impact of the energy module, it is suggested to optimize the operation period to avoid the high-temperature period, and to carry the corresponding spare parts for emergency; the decision instruction set is output in a structured form.

[0097] In the embodiment, by extracting the subsequent task estimation parameters, the component vulnerability map and the operation and maintenance knowledge base are linked to complete the module-level load impact deduction and state trajectory prediction, and finally generate a standardized decision instruction set, realizing the transition from forward-looking risk prediction to operation and maintenance landing; the estimation parameter extraction combined with historical data and field investigation ensures the accuracy of scene simulation; the state trajectory prediction and performance boundary determination realize the accurate positioning and node prediction of the load impact; the operation and maintenance knowledge base provides standardized support for instruction generation, ensuring that the operation suggestion is compliant and feasible; the decision instruction set takes into account the priority and operability, making the operation and maintenance work more targeted, avoiding unplanned downtime in advance, ensuring the operation and maintenance efficiency of the power station, and reducing the cost loss caused by blind maintenance.

[0098] Based on the same inventive concept as the fault detection method of the photovoltaic panel cleaning robot in the foregoing embodiment, the present application also provides a photovoltaic panel cleaning robot fault detection system, as shown in Figure 2 The system comprises:

[0099] A health portrait construction module is configured to construct a health state portrait of the cleaning robot based on multi-modal time series data generated by the cleaning robot during execution of a cleaning task;

[0100] An efficiency wear coefficient module is configured to generate an efficiency wear coefficient according to a deviation between the health state portrait and a pre-established ideal operation behavior model;

[0101] A vulnerability map module is configured to analyze a mapping relationship between the efficiency wear coefficient and a corresponding task scene by combining a photovoltaic panel pollution characteristic parameter and an environmental disturbance parameter of a current cleaning task, and obtain a scene-driven component vulnerability map;

[0102] An operation and maintenance decision generation module is configured to deduce an expected load impact caused by execution of a subsequent planned task on each subsystem of the robot based on the component vulnerability map and a pre-established operation and maintenance knowledge base, and generate a decision instruction set including a maintenance priority and an operation suggestion accordingly.

[0103] The above system in the present application can effectively realize a photovoltaic panel cleaning robot fault detection method, which can achieve the technical effects as described in the above embodiments, which will not be repeated here.

[0104] Although the present application has been described in connection with specific features and embodiments thereof, it will be evident to those of ordinary skill in the art that various modifications and combinations can be made without departing from the spirit and scope of the application. Accordingly, it is intended that the description and drawings be regarded as illustrative rather than restrictive. It is intended that the application cover all modifications, changes, combinations, or equivalents thereof falling within the spirit and scope of the application. Obviously, various modifications and changes can be made to the present application without departing from the scope of the present application. Thus, it is intended that the present application include all such modifications and changes as fall within the scope of the present application and its equivalents.

Claims

1. A photovoltaic panel cleaning robot failure detection method, characterized in that, The application relates to a method for generating a decision instruction set for a cleaning robot based on a health state portrait of the cleaning robot and a preset operation and maintenance knowledge base. The method comprises the following steps: based on multi-modal time series data generated by the cleaning robot during cleaning operation, a health state portrait of the cleaning robot is constructed; the multi-modal time series data at least comprises kinetic data, water circulation and brush body load data and sensor original signal quality data; wherein the kinetic data represents a mechanical execution state, the water circulation and brush body load data represents cleaning efficiency, and the sensor original signal quality data represents environment sensing quality; the health state portrait is constructed by the following method: frequency domain and amplitude domain joint analysis is performed on the kinetic data to extract frequency spectrum features and statistical features representing track walking smoothness, motor driving stability and joint transmission efficiency; process consistency analysis is performed on the water circulation and brush body load data to determine whether the dynamic processes of cleaning liquid supply, sewage recovery and rolling brush contact pressure conform to a preset cleaning operation mode; signal-to-noise ratio and feature point stability evaluation is performed on the sensor original signal quality data to quantify laser radar mapping consistency, camera stain identification confidence and ultrasonic ranging fluctuation range; According to the deviation degree between the health state portrait and a pre-established ideal operation behavior model, an efficiency wear coefficient is generated; Combined with photovoltaic panel pollution characteristic parameters and environmental disturbance parameters of a current cleaning task, a mapping relationship between the efficiency wear coefficient and a corresponding task scene is analyzed to obtain a scene-driven component vulnerability graph, which comprises the following steps: the efficiency wear coefficient is decomposed into a driving module, a cleaning module, a sensing module and an energy module; for different combinations of photovoltaic panel pollution characteristic parameters and environmental disturbance parameters in historical tasks, the change mode and sensitivity degree of the efficiency wear coefficient of each module are counted; a prediction relationship network taking pollution type, pollution degree, wind speed level and temperature interval as input and taking expected wear increment and main failure mode of each module as output is established as the component vulnerability graph; Based on the component vulnerability graph and the preset operation and maintenance knowledge base, expected load impact of executing a subsequent planned task on each subsystem of the robot is deduced, and a decision instruction set is generated according to the expected load impact, which comprises a maintenance priority and an operation suggestion; the expected load impact of executing the subsequent planned task on each subsystem of the robot comprises the following steps: for the subsequent planned task, estimated photovoltaic panel pollution characteristic parameters and environmental disturbance parameters are extracted according to a task description; the estimated photovoltaic panel pollution characteristic parameters and environmental disturbance parameters are input into the component vulnerability graph to obtain expected wear increments of the driving module, the cleaning module, the sensing module and the energy module; the expected wear increments of the driving module, the cleaning module, the sensing module and the energy module are respectively accumulated in combination with the current health state portrait of each module of the robot to predict state trajectories of each module after executing the planned task; when the state trajectory of any predicted module reaches a preset efficiency boundary, it is determined that the module is subjected to load impact; the preset efficiency boundary is calibrated based on robot factory efficiency standards and historical failure data and is set for different modules, and the preset efficiency boundaries of different modules are adapted to the functional positioning and loss tolerance capacity of the modules. ​ 2. The photovoltaic panel cleaning robot failure detection method according to claim 1, characterized in that, The kinetic data includes three-phase current ripple of the driving motor, vibration acceleration spectrum of the walking wheel system, and torque following error of each joint servo motor; The water circulation and brush body load data includes water pump outlet pressure pulsation, water tank liquid level change rate, and power spectral density of the cleaning roller brush motor; The sensor raw signal quality data includes plane fitting residual of laser radar point cloud, local contrast gradient of camera image, and energy attenuation curve of ultrasonic echo signal.

3. The photovoltaic panel cleaning robot failure detection method of claim 1, wherein, The photovoltaic panel pollution feature parameters are obtained through pre-task scanning, including pollution type judgment result, pollution area area ratio, and pollution adhesion intensity grade; The environmental disturbance parameters include real-time wind speed, environmental dust concentration, and photovoltaic panel surface temperature.

4. The photovoltaic panel cleaning robot failure detection method of claim 1, wherein, The ideal operation behavior model is constructed based on the multi-modal time series data of the robot in the initial state without loss, and is iteratively corrected based on the historical statistical data of the same type of robot without fault.

5. The photovoltaic panel cleaning robot failure detection method of claim 1, wherein, The operation and maintenance knowledge base is constructed based on the historical operation and maintenance data of the same type of robot, component fault handling specification, and power station operation and maintenance process, including module fault influence weight, maintenance process standard, and spare parts adaptation information. 6.A photovoltaic panel cleaning robot failure detection system, the system being applied to the photovoltaic panel cleaning robot failure detection method according to claim 1, characterized in that, Comprise: A health portrait construction module for constructing a health state portrait of the cleaning robot based on multi-modal time series data generated by the cleaning robot during cleaning operation; An efficiency wear coefficient module for generating an efficiency wear coefficient according to the deviation between the health state portrait and the pre-established ideal operation behavior model; A vulnerability map module for analyzing the mapping relationship between the efficiency wear coefficient and the corresponding task scene by combining the photovoltaic panel pollution feature parameters and the environmental disturbance parameters of the current cleaning task, and obtaining a scene-driven component vulnerability map; An operation and maintenance decision generation module for deducing the expected load impact of executing subsequent planned tasks on each subsystem of the robot based on the component vulnerability map and the preset operation and maintenance knowledge base, and generating a decision instruction set including maintenance priority and operation suggestion accordingly.

Images