Intelligent pressure adjusting and protecting method and system for photovoltaic cleaning crawler robot

Abstract

The invention provides an intelligent pressure adjustment and protection method and system for a photovoltaic cleaning crawler robot, and relates to the technical field of photovoltaic cleaning robots, and the method comprises the steps: processing a pressure signal through a dual-time scale analysis method, extracting features through a chaos twin neural network, building a baseline compensation model, and generating a self-adaptive pressure threshold value. And the optimal relative displacement and speed value are calculated based on the real-time motion state parameters of the crawler, and a multi-target optimization controller is constructed by adopting sorting learning to adjust the output characteristics. The photovoltaic panel cleaning device can effectively prevent the crawler from damaging the photovoltaic panel, and the cleaning efficiency and safety are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic cleaning robots, and in particular to a method and system for intelligently regulating and protecting the pressure of a photovoltaic cleaning crawler robot. Background Art

[0002] With the widespread adoption of photovoltaic power generation technology, cleaning and maintaining the surface of photovoltaic panels has become a critical step in ensuring power generation efficiency. PV cleaning tracked vehicles, as automated cleaning equipment, can replace manual cleaning of photovoltaic panels, significantly improving cleaning efficiency and reducing labor costs. During the photovoltaic cleaning process, controlling the contact pressure between the tracked vehicle and the photovoltaic panel is particularly important, directly impacting the cleaning effect and the service life of the photovoltaic panel.

[0003] Currently, pressure control systems for photovoltaic cleaning crawler robots primarily rely on fixed parameter settings or simple feedback adjustment mechanisms, failing to intelligently adjust to the dynamic changes in the photovoltaic panel surface condition and cleaning environment. Traditional pressure control methods suffer from numerous technical drawbacks.

[0004] Existing technologies lack the ability to perform refined analysis of pressure signals and are unable to effectively distinguish between continuous pressure baselines and transient impact pressures, resulting in insufficient pressure control accuracy under complex working conditions, which can easily cause damage to the photovoltaic panel surface or incomplete cleaning. Traditional pressure control systems use a fixed threshold design and cannot dynamically adjust pressure parameters based on factors such as photovoltaic panel material properties, surface contamination levels, and ambient temperature. They lack adaptive capabilities and find it difficult to achieve a balance between optimal cleaning and protection effects under different working conditions. Existing tracked vehicle control systems typically separate pressure control from motion control, lack comprehensive consideration of the correlation between the tracked vehicle's motion state and pressure output, and are unable to achieve multi-objective collaborative optimization. This leads to problems such as large pressure fluctuations and delayed response during the actual cleaning process, affecting cleaning quality and equipment service life. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for intelligent pressure regulation and protection of a photovoltaic cleaning crawler robot, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for intelligently regulating and protecting pressure of a photovoltaic cleaning crawler vehicle robot, comprising:

[0007] A dual-time-scale analysis method is used to process the tracked vehicle pressure signal. The collected pressure data is decomposed into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network, an impact feature matrix is ​​generated, and a pressure baseline trend vector is constructed.

[0008] The impact characteristic matrix and pressure baseline trend vector are subjected to variational mode decomposition and joint analysis to establish a baseline compensation model. The energy loss is calculated based on the surface load response characteristics of the photovoltaic panel to generate an adaptive pressure threshold.

[0009] Collect the real-time motion state parameters of the crawler vehicle and calculate the optimal relative displacement and optimal relative speed between the track wheel and the photovoltaic panel based on the adaptive pressure threshold;

[0010] Based on the optimal relative displacement and relative speed, a multi-objective optimization controller is constructed using ranking learning, and the output characteristics of the tracked vehicle are adjusted through a competition-cooperation mechanism.

[0011] In an optional embodiment, a dual-time-scale analysis method is used to process the pressure signal of the tracked vehicle, decomposing the collected pressure data into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network to generate an impact feature matrix, and a pressure baseline trend vector is constructed, including:

[0012] The pressure signal of the photovoltaic cleaning crawler robot is subjected to multi-scale entropy analysis to obtain a time scale demarcation threshold, and the optimal time scale is determined according to the inflection point of the entropy curve of the time scale demarcation threshold. The pressure signal is integrated using the optimal time scale to obtain a slow time scale pressure baseline component, and the slow time scale pressure baseline component is subtracted from the pressure signal to obtain a fast time scale transient impact component.

[0013] The fast time scale transient impact component is reconstructed by cross-correlation analysis and false neighbor analysis to determine the reconstruction parameters, determine the phase space matrix, extract the topological eigenvector from the phase space matrix, and input it into the chaotic twin neural network, and obtain the chaotic feature representation through contrast loss function training;

[0014] The chaotic feature representation is constructed into a transient impact feature matrix, and the slow time scale pressure baseline component is averaged within a sliding time window to obtain a pressure baseline trend vector. The transient impact feature matrix and the pressure baseline trend vector constitute the characteristic representation of the pressure signal of the photovoltaic cleaning crawler vehicle robot.

[0015] In an optional embodiment, the fast time scale transient impact component is subjected to cross-correlation analysis and false neighbor analysis to determine reconstruction parameters, determine a phase space matrix, extract a topological eigenvector from the phase space matrix, and input it into a chaotic twin neural network. The chaotic feature representation is obtained by contrast loss function training, including:

[0016] The fast time scale transient impact component is calculated by cross-correlation analysis to obtain a candidate reconstruction delay sequence, the optimal reconstruction delay is determined based on the cross-correlation function of the candidate reconstruction delay sequence, and the minimum embedding dimension is obtained by using the optimal reconstruction delay through false neighbor analysis;

[0017] Based on the optimal reconstruction delay and minimum embedding dimension, the fast time-scale transient impact component is reconstructed into a phase space matrix. The phase space dispersion is obtained by calculating the Euclidean distance between adjacent reconstructed vectors in the phase space matrix, and the nearest neighbor search radius is determined. The local density distribution of the reconstructed vector is obtained by calculation within the nearest neighbor search radius.

[0018] Performing density clustering on the local density distribution to obtain a set of feature points, calculating the topological relationship between the feature points, and constructing a topological feature vector;

[0019] The topological feature vector is input into the chaotic twin neural network, a nonlinear transformation is performed through the feature encoder to obtain an initial feature map, and the Mahalanobis distance between the features is calculated through the metric learner;

[0020] A contrast loss function is constructed based on the Mahalanobis distance, and the chaotic twin neural network is trained to obtain an optimized feature mapping model, which is then used to extract the final chaotic feature representation.

[0021] In an optional embodiment, variational modal decomposition is performed on the impact characteristic matrix and the pressure baseline trend vector, and a joint analysis is performed to establish a baseline compensation model. Energy loss is calculated in combination with the surface load response characteristics of the photovoltaic panel. Generating an adaptive pressure threshold includes:

[0022] Performing spectral analysis on the impact characteristic matrix and the pressure baseline trend vector to obtain a spectral energy distribution, calculating the energy coverage based on the spectral energy distribution, determining the number of decomposition levels of variational modal decomposition, and decomposing the impact characteristic matrix and the pressure baseline trend vector according to the number of decomposition levels to obtain a first modal function sequence and a second modal function sequence;

[0023] Based on the first modal function sequence and the second modal function sequence, multi-scale decomposition is performed, the mutual information matrix, distance correlation coefficient and maximum information coefficient are calculated, a comprehensive correlation matrix is ​​constructed, mode pairs are selected according to the corresponding eigenvalues, and a compensation function is constructed;

[0024] A stress-strain relationship is established based on the elastic coefficient tensor and viscosity coefficient tensor of the photovoltaic panel surface, and an integration operation is performed in the time domain to obtain the energy loss per unit area. The energy loss per unit area is integrated over the photovoltaic panel surface to obtain the total energy loss.

[0025] A comprehensive evaluation function including a compensation error term, an energy loss term, and a threshold gradient constraint term is constructed, the output value of the compensation function and the total energy loss are input into the comprehensive evaluation function, and an adaptive pressure threshold is generated through optimization using a dynamic programming algorithm.

[0026] In an optional embodiment, multi-scale decomposition is performed based on the first modal function sequence and the second modal function sequence, the mutual information matrix, the distance correlation coefficient and the maximum information coefficient are calculated, a comprehensive correlation matrix is ​​constructed, a modal pair is selected according to the corresponding eigenvalues, and a compensation function is constructed, including:

[0027] Performing multi-scale decomposition on the first modal function sequence and the second modal function sequence to obtain first modal components and second modal components of different scales, calculating a joint probability distribution based on the first modal components and the second modal components, and constructing a multi-scale mutual information matrix using the joint probability distribution;

[0028] Calculating a dual-centered distance matrix for the first modal function sequence and the second modal function sequence, calculating a distance correlation coefficient based on the dual-centered distance matrix, and gridding the first modal function sequence and the second modal function sequence to obtain a maximum information coefficient;

[0029] The multi-scale mutual information matrix, the distance correlation coefficient, and the maximum information coefficient are weighted and combined using a preset fusion coefficient to construct a comprehensive correlation matrix, the comprehensive correlation matrix is ​​subjected to eigenvalue decomposition to obtain eigenvectors, and a modal pair is determined based on the eigenvectors;

[0030] The deviation between the comprehensive correlation matrix and a preset reference correlation degree is calculated, an adaptive weight coefficient is constructed based on the deviation, and the adaptive weight coefficient is combined with the output of the kernel function mapping to generate a compensation function.

[0031] In an optional embodiment, based on the optimal relative displacement and the optimal relative speed, a multi-objective optimization controller is constructed using ranking learning, and the output characteristics of the crawler vehicle are adjusted through a competition-cooperation mechanism, including:

[0032] generating a control parameter vector including a proportional parameter, an integral parameter, a differential parameter, and a compensation parameter, and constructing a multidimensional candidate solution space based on the control parameter vector;

[0033] Performing non-dominated sorting on candidate solutions in the multidimensional candidate solution space, calculating the optimal relative displacement deviation and the optimal relative velocity value deviation corresponding to each candidate solution, determining the dominance number of each candidate solution, and constructing a Pareto frontier surface based on the dominance number;

[0034] The optimal control parameter vector is selected based on the Pareto frontier, and a displacement competition function including an adaptive weight coefficient and a speed coordination function including a coordination coefficient are constructed.

[0035] The displacement competition function and the speed cooperation function are combined through a competition-cooperation balance factor to generate a mechanism fusion function, and the mechanism fusion function is combined with the optimal control parameter vector to construct an optimized control law;

[0036] An output characteristic compensation amount is calculated based on the optimized control law, the output characteristic compensation amount is combined with the dominant relationship in the Pareto front surface, an adaptive adjustment function is constructed, and the output characteristic of the actuator is adjusted according to the adaptive adjustment function.

[0037] In an optional embodiment, selecting an optimal control parameter vector based on the Pareto frontier, and constructing a displacement competition function including an adaptive weight coefficient and a speed coordination function including a coordination coefficient include:

[0038] A comprehensive evaluation function is constructed based on candidate solutions in the Pareto front surface, wherein the comprehensive evaluation function includes a dominance relationship determination term and a displacement-velocity deviation term, wherein the dominance relationship determination term is the number of times the candidate solution is dominated by other candidate solutions, and the displacement-velocity deviation term is the weighted sum of the squared displacement deviation term and the squared velocity deviation term corresponding to the candidate solution; the dominance relationship determination term and the displacement-velocity deviation term are weightedly combined using a preset weight factor to obtain an evaluation value, and the candidate control parameter vector with the smallest evaluation value is selected from the Pareto front surface to determine the optimal control parameter vector;

[0039] A displacement competition function is constructed based on the optimal control parameter vector, and an adaptive weight coefficient is calculated according to the cumulative effect of the displacement deviation, wherein the adaptive weight coefficient includes an exponential decay term and a deviation integral term, and a competition action radius is determined based on the ratio of the displacement deviation to the speed deviation;

[0040] A speed coordination function is constructed based on the optimal control parameter vector, a dynamic coordination coefficient is calculated according to the coupling relationship between the displacement deviation and the speed deviation, and a coordination radius is determined based on the speed deviation change rate.

[0041] A second aspect of the embodiments of the present invention provides a photovoltaic cleaning crawler robot pressure intelligent regulation and protection system, comprising:

[0042] The first unit is used to process the tracked vehicle pressure signal using a dual-time-scale analysis method. The collected pressure data is decomposed into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network, the impact feature matrix is ​​generated, and a pressure baseline trend vector is constructed.

[0043] The second unit is used to perform variational mode decomposition and joint analysis on the impact characteristic matrix and the pressure baseline trend vector, establish a baseline compensation model, calculate the energy loss based on the surface load response characteristics of the photovoltaic panel, and generate an adaptive pressure threshold;

[0044] The third unit is used to collect the real-time motion state parameters of the crawler vehicle and calculate the optimal relative displacement and optimal relative speed between the track wheel and the photovoltaic panel based on the adaptive pressure threshold;

[0045] The fourth unit is used to construct a multi-objective optimization controller based on the optimal relative displacement and the optimal relative speed value by using ranking learning, and to adjust the output characteristics of the tracked vehicle through a competition-cooperation mechanism.

[0046] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0051] In an embodiment of the present invention, the pressure signal of the tracked vehicle is processed by a dual-time-scale analysis method and a chaotic twin neural network, thereby achieving precise separation and feature extraction of the pressure baseline and transient impact, effectively improving the accuracy and stability of the pressure signal processing, and providing a reliable data basis for subsequent intelligent adjustment; combining variational mode decomposition and the surface load response characteristics of the photovoltaic panel, a baseline compensation model is established and an adaptive pressure threshold is generated, so that the system can automatically adjust the pressure control parameters according to different working conditions, effectively preventing the tracked vehicle from causing excessive pressure damage to the photovoltaic panel and extending the service life of the photovoltaic panel; based on the optimal relative displacement and speed value, a multi-objective optimization controller is constructed using sorting learning, and the intelligent adjustment of the output characteristics of the tracked vehicle is achieved through a competition-cooperation mechanism, thereby improving the cleaning efficiency while ensuring the safety of the photovoltaic panel, and solving the technical problems of low pressure control accuracy and poor adaptability of traditional photovoltaic cleaning equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a method for intelligent pressure regulation and protection of a photovoltaic cleaning crawler robot according to an embodiment of the present invention;

[0053] Figure 2 It is a schematic diagram of the multi-scale entropy analysis curve;

[0054] Figure 3 This is a schematic diagram of pressure signal decomposition;

[0055] Figure 4 Schematic diagram comparing the feature extraction performance of chaotic twin neural networks and traditional methods. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0057] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0058] Figure 1 FIG. 1 is a flow chart of a method for intelligently regulating and protecting pressure of a photovoltaic cleaning crawler robot according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] A dual-time-scale analysis method is used to process the tracked vehicle pressure signal. The collected pressure data is decomposed into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network, an impact feature matrix is ​​generated, and a pressure baseline trend vector is constructed.

[0060] The impact characteristic matrix and pressure baseline trend vector are subjected to variational mode decomposition and joint analysis to establish a baseline compensation model. The energy loss is calculated based on the surface load response characteristics of the photovoltaic panel to generate an adaptive pressure threshold.

[0061] Collect the real-time motion state parameters of the crawler vehicle and calculate the optimal relative displacement and optimal relative speed between the track wheel and the photovoltaic panel based on the adaptive pressure threshold;

[0062] Based on the optimal relative displacement and relative speed, a multi-objective optimization controller is constructed using ranking learning, and the output characteristics of the tracked vehicle are adjusted through a competition-cooperation mechanism.

[0063] In an optional embodiment, a dual-time-scale analysis method is used to process the tracked vehicle pressure signal, decomposing the collected pressure data into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network to generate an impact feature matrix, and a pressure baseline trend vector is constructed, including:

[0064] The pressure signal of the photovoltaic cleaning crawler robot is subjected to multi-scale entropy analysis to obtain a time scale demarcation threshold, and the optimal time scale is determined according to the inflection point of the entropy curve of the time scale demarcation threshold. The pressure signal is integrated using the optimal time scale to obtain a slow time scale pressure baseline component, and the slow time scale pressure baseline component is subtracted from the pressure signal to obtain a fast time scale transient impact component.

[0065] The fast time scale transient impact component is reconstructed by cross-correlation analysis and false neighbor analysis to determine the reconstruction parameters, determine the phase space matrix, extract the topological eigenvector from the phase space matrix, and input it into the chaotic twin neural network, and obtain the chaotic feature representation through contrast loss function training;

[0066] The chaotic feature representation is constructed into a transient impact feature matrix, and the slow time scale pressure baseline component is averaged within a sliding time window to obtain a pressure baseline trend vector. The transient impact feature matrix and the pressure baseline trend vector constitute the characteristic representation of the pressure signal of the photovoltaic cleaning crawler vehicle robot.

[0067] In a specific embodiment, pressure signal data of a photovoltaic cleaning crawler robot is collected. The collected pressure signal usually contains components of different frequencies, and a multi-scale entropy analysis is required to determine the time scale demarcation threshold. Specifically, the sample entropy values ​​at different time scales are calculated for the collected pressure signal to obtain a curve showing the change of entropy value with time scale. An inflection point is found on the curve, and the time scale corresponding to the inflection point is the optimal time scale. For example, in a group of collected pressure signals, the sample entropy values ​​at time scales from 1 to 20 are calculated, and it is found that when the time scale is 8, the entropy curve has an obvious inflection point, and the entropy value drops from 0.85 to 0.72. At this time, the optimal time scale is determined to be 8.

[0068] After determining the optimal time scale, the original pressure signal is integrated to obtain the slow-time-scale pressure baseline component. This integration is performed using a sliding average method with a window length equal to the optimal time scale. For example, if the optimal time scale is 8, the average of eight adjacent sampling points is taken as the slow-time-scale component at the current moment. For a pressure signal sampled at 1000 Hz, this is equivalent to averaging the signal over an 8ms time window.

[0069] The slow-time-scale pressure baseline component is subtracted from the original pressure signal to obtain the fast-time-scale transient impact component. This signal contains information about the transient impact generated during the operation of the tracked vehicle and is of great significance for condition monitoring.

[0070] For the fast-time-scale transient impact component, phase space reconstruction is required to extract its dynamic characteristics. The time delay parameter is determined through cross-correlation analysis. The specific method is to calculate the cross-correlation function between the signal and its time-delayed sequence. The delay value corresponding to the first time the cross-correlation function drops to 1 / e is used as the time delay parameter. In practical applications, for a transient impact signal with a sampling frequency of 1000 Hz, the calculated time delay parameter is typically between 5 and 15, for example, 9.

[0071] The embedding dimension is determined through false neighbor analysis. The false neighbor ratio is calculated for different embedding dimensions. The dimension where the false neighbor ratio falls below a threshold is considered the appropriate embedding dimension. In practice, a threshold of 0.05 is set, and the calculated embedding dimension is typically between 3 and 7, for example, 5.

[0072] Based on the determined time delay parameters and embedding dimension, a phase space matrix is ​​constructed. For a transient impulse signal of length N, a time delay of τ, and an embedding dimension of m, the constructed phase space matrix contains N-(m-1)×τ rows, with each row containing m elements, representing the values ​​of the original signal at different time delays.

[0073] Topological eigenvectors are extracted from the phase space matrix, including the density distribution of phase space points, the curvature of phase space trajectories, and the estimated Lyapunov exponents of phase space points. These eigenvectors can characterize the dynamic characteristics of transient impact signals.

[0074] The extracted topological feature vectors are fed into a chaotic twin neural network for training. The chaotic twin neural network consists of two weight-sharing subnetworks, each with three fully connected layers, 64 and 32 hidden layer neurons, respectively, using ReLU as the activation function. The network input is the phase space topological feature vector, and the output is a low-dimensional feature representation.

[0075] During training, a contrastive loss function was used to keep samples from the same operating condition close in feature space, while samples from different operating conditions were farther apart. The contrastive loss function's boundary value was set to 0.5, the initial learning rate was 0.001, and the Adam optimizer was used for optimization. The number of training epochs was 200.

[0076] After training is complete, the transient impact signal is passed through the trained chaotic twin neural network to obtain a chaotic feature representation. The chaotic feature representations within the continuous time window are combined into a transient impact feature matrix, where each row of the matrix corresponds to the feature representation of a time window.

[0077] For the slow-time-scale pressure baseline component, the mean is calculated within a sliding time window to generate a pressure baseline trend vector. The time window length is set to 10% of the original signal length, and the window sliding step is 50% of the window length. For example, for a pressure signal with a length of 10,000 points, a window length of 1,000 points and a sliding step of 500 points results in a pressure baseline trend vector containing 19 elements.

[0078] The transient impact feature matrix and the pressure baseline trend vector together constitute the characteristic representation of the pressure signal of a photovoltaic cleaning crawler robot. This characteristic representation method can effectively capture the slowly varying baseline trend and rapidly varying transient impact characteristics in the pressure signal, providing effective feature input for crawler vehicle condition monitoring and fault diagnosis.

[0079] Figure 2 This is a schematic diagram of the multi-scale entropy analysis curve. A key area is clearly marked in the figure (marked with a dotted rectangle), showing the inflection point of the entropy curve at a time scale of 8. This inflection point is where the entropy curve changes significantly, indicating that it is the optimal time scale demarcation threshold. At this location, the entropy value of normal operating conditions drops from approximately 0.85 to 0.72, and other operating conditions also have similar sharp drops. This multi-scale entropy analysis method can be used to determine the optimal time scale for signal decomposition, providing a theoretical basis for subsequent dual-time scale analysis. By observing the entropy value change characteristics under different operating conditions, changes in the system's dynamic characteristics can be identified, which can then be used for condition monitoring and fault diagnosis.

[0080] Figure 3 This is a schematic diagram of pressure signal decomposition. The dual-time-scale decomposition method effectively separates the original pressure signal into a slowly varying baseline component and a fast-varying impact component, making the fault characteristics more obvious in the fast-time-scale component, which facilitates subsequent feature extraction and condition monitoring.

[0081] In this embodiment, the optimal time scale is determined through multi-scale entropy analysis, and the pressure signal is integrated, which can clearly separate the baseline trend of the slow time scale from the transient impact component of the fast time scale, avoiding the loss of useful impact information by traditional filtering methods; the transient impact component of the fast time scale retains the sudden impact signal during the cleaning process, and the phase space reconstruction parameters are optimized with the help of cross-correlation and false neighbor analysis, so that the phase space matrix can truly reflect the system dynamics, which is conducive to extracting highly sensitive topological features; the topological feature vector is input into the chaotic twin neural network and trained using the contrast loss function, which can learn the feature representation under the chaotic dynamics of the system and better capture nonlinear and chaotic behaviors than traditional neural networks; the transient impact feature matrix and the pressure baseline trend vector under the sliding window are combined to form a complete feature representation, which can simultaneously reflect short-term impacts and long-term trends, providing comprehensive information for subsequent fault warning or performance evaluation.

[0082] In an optional embodiment, the fast time scale transient impact component is reconstructed by cross-correlation analysis and false neighbor analysis to determine the reconstruction parameters, determine the phase space matrix, extract the topological eigenvector from the phase space matrix, and input it into the chaotic twin neural network. The chaotic feature representation is obtained by contrast loss function training, including:

[0083] The fast time scale transient impact component is calculated by cross-correlation analysis to obtain a candidate reconstruction delay sequence, the optimal reconstruction delay is determined based on the cross-correlation function of the candidate reconstruction delay sequence, and the minimum embedding dimension is obtained by using the optimal reconstruction delay through false neighbor analysis;

[0084] Based on the optimal reconstruction delay and minimum embedding dimension, the fast time-scale transient impact component is reconstructed into a phase space matrix. The phase space dispersion is obtained by calculating the Euclidean distance between adjacent reconstructed vectors in the phase space matrix, and the nearest neighbor search radius is determined. The local density distribution of the reconstructed vector is obtained by calculation within the nearest neighbor search radius.

[0085] Performing density clustering on the local density distribution to obtain a set of feature points, calculating the topological relationship between the feature points, and constructing a topological feature vector;

[0086] Inputting the topological feature vector into the chaotic twin neural network, performing nonlinear transformation through the feature encoder to obtain an initial feature map, and calculating the Mahalanobis distance between features through the metric learner;

[0087] A contrast loss function is constructed based on the Mahalanobis distance, and the chaotic twin neural network is trained to obtain an optimized feature mapping model, which is then used to extract the final chaotic feature representation.

[0088] In a specific embodiment, starting with the processing of the fast time-scale transient impact component, the candidate reconstruction delay sequence is calculated by cross-correlation analysis. Specifically, for the collected time series signal x(t), the fast time-scale transient impact component s(t) is extracted by wavelet decomposition or empirical mode decomposition. The autocorrelation function C(τ) of s(t) is calculated, where τ is a delay parameter with a value range of 1 to 100 sampling points. For each τ value, the amplitude of C(τ) is calculated. When C(τ) first drops to 1 / e of its initial value, the τ value is recorded as the candidate reconstruction delay τ0. In practical applications, τ0=15 can be taken as the initial estimate, and the optimal value is searched within a range of ±5 around it.

[0089] The optimal reconstruction delay is determined based on the cross-correlation function of the candidate reconstruction delay sequence. The average mutual information (AMI(τ)) of the cross-correlation function R(τ) corresponding to different values ​​of τ is calculated. The value of τ at which AMI(τ) reaches its first local minimum is the optimal reconstruction delay (τopt). In practice, if AMI(τ) reaches its first local minimum at τ=18, then τopt=18. Using the optimal reconstruction delay, the minimum embedding dimension is determined through false nearest neighbor analysis. Specifically, as the embedding dimension d increases, the false nearest neighbor ratio (FNN(d)) is calculated. When FNN(d) falls below a preset threshold (e.g., 1%), the corresponding value of d is the minimum embedding dimension (dmin). In practice, d can be started from 2 and gradually increased to 10. If FNN drops to 0.8% at d=5, then dmin=5.

[0090] Based on the optimal reconstruction delay τopt = 18 and the minimum embedding dimension dmin = 5, the fast-time-scale transient impulse component is reconstructed into a phase space matrix. For time series s(t), t = 1, 2, ..., N, reconstruction vectors Xi = [s(i), s(i + τopt), s(i + 2τopt), ..., s(i + (dmin - 1)τopt)] are constructed, i = 1, 2, ..., N - (dmin - 1)τopt. These reconstruction vectors form the phase space matrix X = [X1, X2, ..., XM], where M = N - (dmin - 1)τopt is the total number of reconstruction vectors. In practice, for a time series s(t) of length 2000, M = 2000 - (5 - 1) × 18 = 1928 reconstruction vectors are obtained, each with a dimension of 5.

[0091] The phase space dispersion is obtained by calculating the Euclidean distance between adjacent reconstruction vectors in the phase space matrix. For any two reconstruction vectors Xi and Xj, their Euclidean distance dist(Xi, Xj) = [(Xi-Xj)^T(Xi-Xj)] is calculated. 1 / 2 . The phase space dispersion Disp is defined as the ratio of the standard deviation of all distances to the mean. In the actual data, Disp=0.685 is calculated. Determine the nearest neighbor search radius ε, generally take ε=0.1×max(dist). For actual data, if the maximum distance is 3.5, then ε=0.35. The local density distribution of the reconstructed vector is calculated within the nearest neighbor search radius. For each reconstructed vector Xi, the number of reconstructed vectors ni with a distance less than ε is counted, and the local density ρi=ni / M. For actual cases, the possible local density distribution range is 0.001 to 0.15, with an average value of about 0.042.

[0092] Density clustering is performed on the local density distribution to obtain a set of feature points. A density-based spatial clustering algorithm is used, with a density threshold of ρth = 0.08 and a distance threshold of δth = 0.5. High-density points are identified where ρi > ρth, and the minimum distance δi from each high-density point to a higher-density point is calculated. Points that simultaneously satisfy ρi > ρth and δi > δth are selected as cluster centers, forming a set of feature points P = {P1, P2, ..., Pk}. In the actual clustering results, k = 8 feature points may be identified from 1928 reconstructed vectors. The topological relationships between the feature points are calculated to construct a topological feature vector. Based on the relative positions and connectivity between the feature points, a topological distance matrix TD is calculated, where TD(i, j) represents the geodesic distance between feature points Pi and Pj. The distance matrix is ​​flattened and combined with the density values ​​of the feature points to form a topological feature vector TV = [TD(1, 2), TD(1, 3), ..., TD(k-1, k), ρ1, ρ2, ..., ρk]. For a feature point set of k=8, the dimension of the topological feature vector is 8×7 / 2+8=36.

[0093] The topological feature vector is input into the chaotic twin neural network, and the initial feature map is obtained by nonlinear transformation through the feature encoder. The feature encoder consists of a three-layer fully connected network with 36 input layer nodes, 64 hidden layer nodes, and 32 output layer nodes. The ReLU activation function and batch normalization technique are used. TV is encoded to obtain the initial feature map F0 = Encoder(TV), with a dimension of 32. The Mahalanobis distance between features is calculated using a metric learner. The metric learner learns a parameter matrix W, which is used to calculate the Mahalanobis distance D(F1, F2) = (F1-F2)^T·W·(F1-F2) between two feature vectors F1 and F2. W is a 32×32 symmetric positive definite matrix, initially set to the identity matrix, and optimized through network training.

[0094] A contrastive loss function based on the Mahalanobis distance was constructed, and a chaotic twin neural network was trained to obtain an optimized feature mapping model. For positive pairs (features from the same category) and negative pairs (features from different categories), a contrastive loss function L = Σ[y·D(F1, F2)+(1-y)·max(0,mD(F1, F2))] was defined, where y is the label of the pair (1 for the same category and 0 for different categories), and m is a margin parameter set to 2.0. The network was trained using stochastic gradient descent with a batch size of 64. The learning rate was initially set to 0.001 and halved every 30 epochs. After 100 epochs of training, the loss on the validation set dropped from an initial 0.85 to 0.21. The trained feature mapping model was used to extract the final chaotic feature representation. For new time series data, the aforementioned processing steps were repeated to obtain the topological feature vector TV'. The trained encoder and metric learner were then used to extract a 32-dimensional chaotic feature representation CF = Encoder(TV'). This feature representation can be used for subsequent tasks such as classification and anomaly detection.

[0095] Figure 4 This is a schematic diagram comparing the feature extraction performance of chaotic twin neural networks and traditional methods. Figure 4 As shown in the figure, it can be clearly seen from the data that the comprehensive correlation matrix method proposed in the present invention performs best in all test scenarios. Specifically, under standard working conditions, the recognition accuracy of the three methods are 46.5%, 56.1% and 59.4% respectively; in minor fault scenarios, they reach 56.4%, 64.4% and 72.4% respectively; in severe fault scenarios, they are 60.4%, 68.4% and 75.4% respectively; in complex environments, they are 54.0%, 61.5% and 75.0% respectively. The advantages of the method of the present invention are particularly obvious in severe fault and complex environment scenarios, which are about 15-21 percentage points higher than the traditional phase space reconstruction method and about 7-14 percentage points higher than the statistical learning method.

[0096] In the existing technology, time series feature extraction mainly relies on statistical features, spectral analysis and traditional phase space reconstruction methods. These methods often ignore the topological structure information in chaotic systems and have difficulty capturing the nonlinear dynamic characteristics in transient impact signals. For example, traditional phase space reconstruction methods use fixed uniform time delays and empirically selected embedding dimensions, resulting in poor reconstruction quality; existing density clustering algorithms are sensitive to parameters and difficult to adapt to different working conditions; the feature extraction process lacks a metric learning mechanism and cannot adaptively adjust the feature space. Starting from the perspective of chaotic dynamics and topological data analysis, this application achieves effective capture of the topological structure of fast time-scale transient impact components by optimizing reconstruction parameter selection, improving density clustering and introducing chaotic twin neural networks.

[0097] In an optional embodiment, variational mode decomposition is performed on the impact characteristic matrix and the pressure baseline trend vector and a joint analysis is performed to establish a baseline compensation model. Energy loss is calculated in combination with the surface load response characteristics of the photovoltaic panel to generate an adaptive pressure threshold, including:

[0098] Performing spectral analysis on the impact characteristic matrix and the pressure baseline trend vector to obtain a spectral energy distribution, calculating the energy coverage based on the spectral energy distribution, determining the number of decomposition levels of variational modal decomposition, and decomposing the impact characteristic matrix and the pressure baseline trend vector according to the number of decomposition levels to obtain a first modal function sequence and a second modal function sequence;

[0099] Based on the first modal function sequence and the second modal function sequence, multi-scale decomposition is performed, the mutual information matrix, distance correlation coefficient and maximum information coefficient are calculated, a comprehensive correlation matrix is ​​constructed, mode pairs are selected according to the corresponding eigenvalues, and a compensation function is constructed;

[0100] A stress-strain relationship is established based on the elastic coefficient tensor and viscosity coefficient tensor of the photovoltaic panel surface, and an integration operation is performed in the time domain to obtain the energy loss per unit area. The energy loss per unit area is integrated over the photovoltaic panel surface to obtain the total energy loss.

[0101] A comprehensive evaluation function including a compensation error term, an energy loss term, and a threshold gradient constraint term is constructed, the output value of the compensation function and the total energy loss are input into the comprehensive evaluation function, and an adaptive pressure threshold is generated through optimization using a dynamic programming algorithm.

[0102] In one specific embodiment, an impact signature matrix X and a pressure baseline trend vector Y are obtained from the photovoltaic panel surface. The impact signature matrix X has dimensions m×n, where m represents the number of sensors and n represents the number of sampling points. The pressure baseline trend vector Y has dimensions 1×n, representing the pressure baseline values ​​at n time points. A fast Fourier transform is performed on the impact signature matrix X to obtain a frequency domain representation X_f, and the energy distribution E_x of each frequency component is calculated. Similarly, a fast Fourier transform is performed on the pressure baseline trend vector Y to obtain a frequency domain representation Y_f, and the energy distribution E_y of each frequency component is calculated.

[0103] Based on the energy distributions E_x and E_y, the energy coverage is calculated. Specifically, the frequencies are sorted in descending order of energy, and the energy is accumulated until 95% of the total energy is reached. The corresponding number of frequencies is the required decomposition level K. For example, if the energy of the impact feature matrix is ​​mainly concentrated in the low-frequency band, it is calculated that 6 frequency components are required to cover 95% of the energy; while the energy distribution of the pressure baseline trend vector is more dispersed, 8 frequency components are required to cover 95% of the energy, so K=8 is used as the decomposition level.

[0104] Using variational modal decomposition, the impact signature matrix X is decomposed to obtain K modal functions u_1, u_2, ..., u_K, each corresponding to a central frequency ω_1, ω_2, ..., ω_K. Similarly, variational modal decomposition is performed on the pressure baseline trend vector Y to obtain K modal functions v_1, v_2, ..., v_K, corresponding to central frequencies η_1, η_2, ..., η_K. These modal functions constitute the first and second modal function sequences.

[0105] Perform multi-scale analysis on the first and second modal function sequences and calculate the mutual information matrix (MI). The elements (MI(i, j)) of the mutual information matrix (MI) represent the mutual information between modal functions u_i and v_j, reflecting the nonlinear correlation between the two modal functions. For example, a mutual information value of 0.85 between u_3 and v_2 indicates a strong correlation between the two modal functions.

[0106] Calculate the distance correlation coefficient matrix DC, where the element DC(i, j) represents the distance correlation coefficient between the modal functions u_i and v_j. This coefficient can capture nonlinear and non-monotonic correlations. For example, when the distance correlation coefficient between u_1 and v_1 is 0.92, it means that the distribution of these two modal functions in the geometric space is highly correlated.

[0107] Calculate the maximum information coefficient matrix MIC, where the element MIC(i, j) represents the maximum information coefficient between modal functions u_i and v_j. This coefficient can identify various functional relationships. For example, when the maximum information coefficient between u_5 and v_4 is 0.78, it indicates that there is a strong functional relationship between these two modal functions.

[0108] After normalizing the mutual information matrix MI, the distance correlation coefficient matrix DC, and the maximum information coefficient matrix MIC, a comprehensive correlation matrix CR is constructed by weighted averaging, with weights of 0.4, 0.3, and 0.3, respectively. Based on the comprehensive correlation matrix CR, the modal pairs with the highest correlation are selected, such as (u_1, v_1), (u_3, v_2), (u_5, v_4), etc.

[0109] A compensation function C(t) is constructed based on the selected modal pairs. Specifically, for each pair of modal functions (u_i, v_j), a mapping relationship f_ij is established using a local linear embedding algorithm, such that v_j ≈ f_ij(u_i). The compensation function C(t) is the weighted sum of the mapping functions, with the weights proportional to the correlation between the corresponding modal pairs. For example, when three pairs of modal functions are selected with correlations of 0.92, 0.85, and 0.78, the normalized weights are 0.36, 0.33, and 0.31, respectively.

[0110] Based on the material properties of the photovoltaic panel surface, a stress-strain relationship is established. The elastic modulus tensor E and the viscosity coefficient tensor V of the photovoltaic panel surface are fourth-order tensors that describe the elastic and viscous properties of the material in different directions. In practical applications, for silicon-based photovoltaic panels, the elastic modulus is 130 GPa, the Poisson's ratio is 0.28, and the viscosity coefficient is 2.5×10 -5 Pa·s.

[0111] When a pressure p(t) is applied to the surface of a photovoltaic panel, a strain ε(t) is generated. Based on the stress-strain relationship, the stress σ(t) is calculated. In the time domain, the energy loss per unit area, W, is the time integral of the product of the stress and strain rate of change. For a sampled time series, the energy loss can be approximated using a discrete summation. For example, when the pressure increases from 0 to 5000 Pa over a duration of 0.5 seconds, the calculated energy loss per unit area is 0.023 J / m².

[0112] Integrate the energy loss per unit area, W, over the panel surface to obtain the total energy loss, W_total. For a 2m² panel, if the average energy loss per unit area is 0.023J / m², the total energy loss is 0.046J.

[0113] A comprehensive evaluation function J is constructed, consisting of a compensation error term J_1, an energy loss term J_2, and a threshold gradient constraint term J_3. The compensation error term J_1 represents the mean squared error between the compensated pressure and the actual pressure; the energy loss term J_2 represents the ratio of energy loss to a preset threshold; and the threshold gradient constraint term J_3 is used to limit the smoothness of threshold changes. The weights of these three terms are 0.5, 0.3, and 0.2, respectively.

[0114] A dynamic programming algorithm is used to optimize the comprehensive evaluation function J and generate an adaptive pressure threshold p_th(t). Specifically, the time axis is discretized into n points, and m candidate thresholds are set for each time point. A state transition equation is constructed, and the optimal threshold sequence is obtained through backtracking. For example, when the photovoltaic panel is operating normally, the generated adaptive pressure threshold is 3500Pa. When the ambient temperature rises and the material softens, the threshold is automatically adjusted to 3200Pa. When dust accumulates on the photovoltaic panel surface, the threshold is adjusted to 3800Pa.

[0115] In this embodiment, the number of variational modal decomposition layers is adaptively determined by spectral energy distribution and energy coverage, which can effectively decompose the impact feature matrix and the pressure baseline trend vector into several intrinsic mode functions, retaining key impact information while suppressing noise interference; based on the first and second modal function sequences, the mutual information matrix, distance correlation coefficient and maximum information coefficient are calculated to construct a comprehensive correlation matrix, which can accurately identify the coupling relationship between each mode, and based on this, the most representative modal pair is selected to construct the compensation function; the elastic coefficient tensor and the viscosity coefficient tensor are used to establish the stress-strain relationship, and the energy loss per unit area is obtained by integration in the time domain, and then the total energy loss on the surface of the photovoltaic panel is obtained by integration, providing a quantitative indicator for performance degradation or structural fatigue; the compensation function output, total energy loss and threshold gradient constraint term are jointly incorporated into the comprehensive evaluation function, and solved by a dynamic programming algorithm to achieve global optimal adaptive adjustment of the pressure threshold, thereby improving the sensitivity and robustness of anomaly detection.

[0116] In an optional embodiment, multi-scale decomposition is performed based on the first modal function sequence and the second modal function sequence, the mutual information matrix, the distance correlation coefficient and the maximum information coefficient are calculated, a comprehensive correlation matrix is ​​constructed, a modal pair is selected according to the corresponding eigenvalue, and a compensation function is constructed, including:

[0117] Performing multi-scale decomposition on the first modal function sequence and the second modal function sequence to obtain first modal components and second modal components of different scales, calculating a joint probability distribution based on the first modal components and the second modal components, and constructing a multi-scale mutual information matrix using the joint probability distribution;

[0118] Calculating a dual-centered distance matrix for the first modal function sequence and the second modal function sequence, calculating a distance correlation coefficient based on the dual-centered distance matrix, and gridding the first modal function sequence and the second modal function sequence to obtain a maximum information coefficient;

[0119] The multi-scale mutual information matrix, the distance correlation coefficient, and the maximum information coefficient are weighted and combined using a preset fusion coefficient to construct a comprehensive correlation matrix, the comprehensive correlation matrix is ​​subjected to eigenvalue decomposition to obtain eigenvectors, and a modal pair is determined based on the eigenvectors;

[0120] The deviation between the comprehensive correlation matrix and a preset reference correlation degree is calculated, an adaptive weight coefficient is constructed based on the deviation, and the adaptive weight coefficient is combined with the output of the kernel function mapping to generate a compensation function.

[0121] In a specific embodiment, when performing multi-scale decomposition on the first mode function sequence and the second mode function sequence, the empirical mode decomposition technique is used to decompose the complex signal into a series of intrinsic mode functions. Specifically, for the first mode function sequence A(t)={a1(t), a2(t), ..., a n (t)} and the second modal function sequence B(t)={b1(t), b2(t), ..., b m (t)} are decomposed separately, where t represents the time index. The multi-scale decomposition process determines each intrinsic mode through iterative screening. For the first modal function sequence, by calculating all local extreme points, the upper envelope e_max(t) and the lower envelope e_min(t) are constructed, and the mean m(t)=(e_max(t)+e_min(t)) / 2 is calculated. Then, the mean is subtracted from the original data to obtain h(t)=A(t)-m(t). Repeat this process until h(t) meets two conditions: the number of extreme points and the number of zero-crossing points do not differ by more than 1; at any point in the entire data set, the mean of the upper and lower envelopes is zero. After the conditions are met, h(t) is determined to be the first intrinsic mode function, denoted as c1(t). Subtract c1(t) from the original data, and perform the same processing on the residual to obtain c2(t). Repeat this process until the residual becomes a monotonic function. The second modal function sequence is processed in the same way, and finally the first modal components of multiple scales {c1 A (t), c2 A (t), ..., c p A (t)} and the second modal component {c1 B (t), c2 B (t), ..., c q B (t)}.

[0122] Based on the obtained modal components of different scales, the joint probability distribution is calculated. For the i-th component c_i of the first modality A and the jth component c_j of the second mode B , construct a two-dimensional histogram, and obtain the joint probability distribution P(c_i A , c_j B ). In actual operation, c_i A and c_j B The range of is divided into N_A and N_B equal-width intervals, and the number of data points in each pair of interval combinations is counted and divided by the total number of samples to obtain the joint probability. Kernel density estimation can also be used to obtain a smoother joint probability distribution. Mutual information I(c_i A ; c_j B), which represents the degree of information sharing between the two modal components. The mutual information values ​​of different scale combinations are combined to form a multi-scale mutual information matrix MI, where the matrix element MI_ij represents the mutual information value between the i-th component of the first modal and the j-th component of the second modal.

[0123] At the same time, the dual-centered distance matrix is ​​calculated for the original first modal function sequence and the second modal function sequence. First, the Euclidean distance matrix D_A within the first modality is calculated, where D_A(i, j) represents the distance between the i-th and j-th sample points in the first modality; similarly, the distance matrix D_B of the second modality is calculated. The two distance matrices are dual-centered, that is, the mean of the corresponding row and column is subtracted from each element, and then the mean of the entire matrix is ​​added to obtain the dual-centered distance matrices H_A and H_B. Based on the dual-centered distance matrix, the distance correlation coefficient dCor is calculated, specifically by calculating the square root of the sum of the product of the elements of the two dual-centered distance matrices, and then dividing it by the square root of the sum of the squares of the elements of each matrix. The range of the distance correlation coefficient is [0, 1]. The larger the value, the stronger the correlation between the two modal data.

[0124] In the grid partitioning step, the first and second modal function sequences are divided into grids of different sizes. Specifically, for different grid sizes D, the first modal data A and the second modal data B are divided into D×D subregions. The mutual information for each subregion is calculated, and the grid partition that maximizes the mutual information is selected to obtain the maximum information coefficient (MIC). MIC has excellent capabilities for detecting various functional relationships, especially nonlinear ones.

[0125] The multi-scale mutual information matrix MI, distance correlation coefficient dCor, and maximum information coefficient MIC obtained above are weighted and combined using the preset fusion coefficients α, β, and γ to construct the comprehensive correlation matrix CR. The calculation formula is CR = α·MI + β·dCor + γ·MIC, where α+β+γ=1. In practical applications, the weights of each coefficient can be adjusted according to the specific scenario. For example, when the data exhibits obvious nonlinear relationships, the weight of MIC can be appropriately increased; when the data scale is large, the weight of dCor can be increased. A set of feasible parameter settings is α=0.4, β=0.3, and γ=0.3.

[0126] Perform eigenvalue decomposition on the comprehensive correlation matrix CR and obtain the eigenvalues ​​λ1≥λ2≥...≥λ n and the corresponding eigenvectors v1, v2, ..., v n The main eigenvectors are selected according to the size of the eigenvalues. Usually, the first k eigenvectors with a cumulative explained variance of 90% are retained to form a matrix V = [v1, v2, ..., v k]. The modal pairs are determined based on the eigenvector matrix V. The specific method is to calculate the projection of the original modal data in the eigenvector space and find the dimension combination with the largest projection value. These dimension combinations constitute the modal pairs.

[0127] To further improve the accuracy of correlation analysis, the deviation between the comprehensive correlation matrix CR and the preset reference correlation R is calculated. The reference correlation can be set based on domain knowledge or prior information, or it can be obtained through multiple experimental statistics. The deviation D = |CR-R| is calculated, and the adaptive weight coefficient w is constructed based on the deviation. The weight coefficient is calculated using a decreasing function, so that the area with larger deviation has smaller weight, for example, w = exp(-D 2 / σ 2 ), where σ is an adjustable parameter that controls the speed of weight decay. Practice shows that setting σ to 0.2 has a good effect.

[0128] Next, the adaptive weight coefficient w is combined with the output of the kernel function mapping to generate the compensation function. The kernel function mapping uses the radial basis function (RBF), which is expressed as K(x, y) = exp(-||xy|| 2 / 2θ 2 ), where x and y are data points, and θ is the bandwidth parameter. The original data is mapped into a high-dimensional feature space using a kernel function, yielding the mapping output Φ. The compensation function C, defined as C = w·Φ, is used to adjust the association analysis results, particularly in areas where the original association analysis method underperforms.

[0129] Traditional cross-modal association analysis mainly relies on a single correlation measure, such as the Pearson correlation coefficient, the Spearman rank correlation coefficient, etc., but these methods have limited ability to capture nonlinear relationships or complex dependencies. Some improved methods such as canonical correlation analysis (CCA) can process multiple variables at the same time, but are still limited by the linear assumption. In response to these limitations, the method of this embodiment integrates three highly complementary association measures: multi-scale mutual information, distance correlation coefficient, and maximum information coefficient, constructs a comprehensive association matrix, and introduces adaptive weights and kernel function mapping for compensation, which significantly improves the accuracy and robustness of cross-modal data association analysis, and is particularly suitable for analysis scenarios with complex nonlinear relationships.

[0130] In an optional embodiment, based on the optimal relative displacement and the optimal relative speed, a multi-objective optimization controller is constructed using ranking learning, and the output characteristics of the crawler vehicle are adjusted through a competition-cooperation mechanism, including:

[0131] generating a control parameter vector including a proportional parameter, an integral parameter, a differential parameter, and a compensation parameter, and constructing a multidimensional candidate solution space based on the control parameter vector;

[0132] Performing non-dominated sorting on candidate solutions in the multidimensional candidate solution space, calculating the optimal relative displacement deviation and the optimal relative velocity value deviation corresponding to each candidate solution, determining the dominance number of each candidate solution, and constructing a Pareto frontier surface based on the dominance number;

[0133] The optimal control parameter vector is selected based on the Pareto frontier, and a displacement competition function including an adaptive weight coefficient and a speed coordination function including a coordination coefficient are constructed.

[0134] The displacement competition function and the speed cooperation function are combined through a competition-cooperation balance factor to generate a mechanism fusion function, and the mechanism fusion function is combined with the optimal control parameter vector to construct an optimized control law;

[0135] An output characteristic compensation amount is calculated based on the optimized control law, the output characteristic compensation amount is combined with the dominant relationship in the Pareto front surface, an adaptive adjustment function is constructed, and the output characteristic of the actuator is adjusted according to the adaptive adjustment function.

[0136] In one specific embodiment, a control parameter vector is generated, comprising a proportional parameter, an integral parameter, a differential parameter, and a compensation parameter. This control parameter vector can be expressed as [Kp, Ki, Kd, ​​Kc], where Kp is the proportional parameter, ranging from [1.5 to 3.0]; Ki is the integral parameter, ranging from [0.1 to 0.5]; Kd is the differential parameter, ranging from [0.5 to 1.5]; and Kc is the compensation parameter, ranging from [0.2 to 0.8]. Based on these parameter ranges, a multidimensional candidate solution space is constructed. For example, 50 points can be uniformly sampled within each parameter range, forming a 50×50×50×50 candidate solution space, for a total of 6,250,000 candidate solutions.

[0137] Perform a non-dominated sort on the candidate solutions in the multidimensional candidate solution space. To improve computational efficiency, Latin hypercube sampling can be used to select 500 representative samples from the candidate solution space. For each candidate solution, calculate its corresponding optimal relative displacement deviation and optimal relative velocity deviation. The optimal relative displacement deviation is defined as the absolute difference between the actual displacement and the target displacement, and the optimal relative velocity deviation is defined as the absolute difference between the actual velocity and the target velocity. For example, for the candidate solution [2.1, 0.3, 0.8, 0.5], the optimal relative displacement deviation is 0.15 meters, and the optimal relative velocity deviation is 0.08 meters per second, calculated through simulation of the tracked vehicle dynamics model.

[0138] Determine the domination number of each candidate solution—the number of times it is dominated by other solutions. If all objective function values ​​of solution A are not inferior to those of solution B, and at least one objective function value is superior to that of solution B, then solution A is said to dominate B. For example, if candidate solution A has a displacement deviation of 0.12 meters and a velocity deviation of 0.07 meters per second, while candidate solution B has a displacement deviation of 0.15 meters and a velocity deviation of 0.08 meters per second, then solution A dominates B. By calculating the domination number of each solution, solutions with a domination number of 0 are assigned to the first frontier. From the remaining solutions, find solutions with a domination number of 0 and assign them to the second frontier. This process continues, constructing a complete Pareto frontier hierarchy.

[0139] The optimal control parameter vector is selected based on the Pareto frontier. On the first frontier, the distribution of solutions is assessed using a crowding calculation method. The crowding degree indicates the sparseness of solutions in the target space; a higher crowding degree indicates a smaller number of solutions surrounding it. Solutions with higher crowding degrees are selected as the optimal control parameter vector to ensure solution diversity. For example, calculations determined the optimal control parameter vector to be [2.3, 0.35, 0.9, 0.45].

[0140] A displacement competition function with an adaptive weight coefficient is constructed. The displacement competition function is expressed as the product of the displacement deviation and the adaptive weight coefficient. The adaptive weight coefficient is dynamically adjusted based on the current displacement deviation. When the displacement deviation is large, the weight coefficient increases, and vice versa. For example, when the displacement deviation is 0.2 meters, the adaptive weight coefficient is 0.8; when the displacement deviation decreases to 0.1 meters, the adaptive weight coefficient decreases to 0.6.

[0141] Construct a speed coordination function that includes a coordination coefficient. The speed coordination function is expressed as the product of the speed deviation and the coordination coefficient. The coordination coefficient is determined based on the relationship between the displacement deviation and the speed deviation. When the displacement deviation and the speed deviation change in the same direction, the coordination coefficient increases; when they change in opposite directions, the coordination coefficient decreases. For example, when the displacement deviation and the speed deviation decrease simultaneously, the coordination coefficient is 0.7; when the displacement deviation decreases and the speed deviation increases, the coordination coefficient is 0.4.

[0142] The displacement competition function and the speed coordination function are combined using a competition-cooperation balance factor to generate a mechanism fusion function. The competition-cooperation balance factor is dynamically adjusted based on the current control stage. In the early stages of control, displacement competition dominates, and the balance factor favors displacement competition. As the control process progresses, speed coordination gradually strengthens, and the balance factor tilts toward speed coordination. For example, in the first 10 seconds of control, the balance factor is 0.7, favoring displacement competition. Between 10 and 20 seconds, the balance factor gradually drops to 0.5, achieving equilibrium between displacement competition and speed coordination. After 20 seconds, the balance factor drops further to 0.3, favoring speed coordination.

[0143] The mechanism fusion function is combined with the optimal control parameter vector to construct an optimized control law. The optimized control law combines the parameters in the optimal control parameter vector with the mechanism fusion function to generate a control signal. For example, when the mechanism fusion function value is 0.6, the control signal is 2.3 × 0.6 + 0.35 × integral term + 0.9 × differential term + 0.45 × compensation term.

[0144] The output characteristic compensation is calculated based on the optimized control law. The output characteristic compensation is calculated based on the difference between the control signal and the ideal response. For example, when the control signal is 2.5 and the ideal response is 2.7, the output characteristic compensation is 0.2. The output characteristic compensation is combined with the dominant relationship in the Pareto front to construct an adaptive adjustment function. The adaptive adjustment function adjusts the application of the compensation based on the position of the current solution in the Pareto front. For example, for solutions located at the edge of the first front, the adaptive adjustment function enhances the effect of the compensation; for solutions located at the center of the first front, the adaptive adjustment function weakens the effect of the compensation.

[0145] The output characteristics of the actuator are adjusted according to the adaptive adjustment function. The actuator adjusts the driving force and steering force of the crawler vehicle according to the adjusted control signal to achieve precise control of the crawler vehicle's motion state.

[0146] In this embodiment, by constructing a multidimensional candidate solution space and based on non-dominated sorting (Pareto front), proportional, integral, differential, and compensation parameters can be simultaneously weighed to ensure that the selected control parameters achieve a global optimal compromise between displacement and speed performance, rather than the local optimality of traditional single-objective regulation. A displacement competition function and a speed cooperation function are introduced, and through dynamic fusion of the competition-cooperation balance factor, the controller can automatically adjust the priority of each indicator according to the real-time operating conditions, thereby always maintaining the optimal response speed and minimum overshoot under different operating conditions. By utilizing the diversified optimal solutions on the Pareto front surface and combining it with an adaptive adjustment function, the output characteristics of the actuator can be quickly adjusted to different disturbances or load changes, significantly enhancing the system's robustness to model uncertainties and external interferences. By combining the output characteristic compensation amount with the Pareto dominance relationship, real-time compensation for nonlinear characteristics and dynamic errors is achieved, effectively suppressing residual oscillations and steady-state errors, and improving tracking accuracy and stability.

[0147] In an optional embodiment, selecting an optimal control parameter vector based on the Pareto frontier, and constructing a displacement competition function including an adaptive weight coefficient and a speed coordination function including a coordination coefficient include:

[0148] A comprehensive evaluation function is constructed based on candidate solutions in the Pareto front surface, wherein the comprehensive evaluation function includes a dominance relationship determination term and a displacement-velocity deviation term, wherein the dominance relationship determination term is the number of times the candidate solution is dominated by other candidate solutions, and the displacement-velocity deviation term is the weighted sum of the squared displacement deviation term and the squared velocity deviation term corresponding to the candidate solution; the dominance relationship determination term and the displacement-velocity deviation term are weightedly combined using a preset weight factor to obtain an evaluation value, and the candidate control parameter vector with the smallest evaluation value is selected from the Pareto front surface to determine the optimal control parameter vector;

[0149] A displacement competition function is constructed based on the optimal control parameter vector, and an adaptive weight coefficient is calculated according to the cumulative effect of the displacement deviation, wherein the adaptive weight coefficient includes an exponential decay term and a deviation integral term, and a competition action radius is determined based on the ratio of the displacement deviation to the speed deviation;

[0150] A speed coordination function is constructed based on the optimal control parameter vector, a dynamic coordination coefficient is calculated according to the coupling relationship between the displacement deviation and the speed deviation, and a coordination radius is determined based on the speed deviation change rate.

[0151] In a specific embodiment, when constructing a comprehensive evaluation function, the function includes two parts: a dominance relationship judgment term and a displacement-velocity deviation term. The dominance relationship judgment term represents the number of times a candidate solution is dominated by other candidate solutions, and the displacement-velocity deviation term represents the weighted sum of the squared displacement deviation term and the squared velocity deviation term corresponding to the candidate solution. In the specific implementation process, for each candidate solution on the Pareto front, the number of times it is dominated by other candidate solutions is calculated. For example, if there are five candidate solutions, namely solutions 1 to 5, and if solution 1 is dominated by solutions 2, 3, and 5, then the dominance relationship judgment term value of solution 1 is 3.

[0152] The displacement-velocity deviation term is calculated by combining the squared displacement deviation and the squared velocity deviation of the candidate solution using a weighting coefficient. In one embodiment, if the candidate solution has a displacement deviation of 3.5 cm and a velocity deviation of 0.8 cm / s, the displacement deviation weight is 0.7, and the velocity deviation weight is 0.3, then the displacement-velocity deviation term is: 0.7×(3.5) 2 +0.3×(0.8) 2 =8.745.

[0153] The comprehensive evaluation function is obtained by weighted combination of the dominance relationship judgment item and the displacement velocity deviation item, and the weight factor can be adjusted according to the actual application scenario. In a specific example, if the weight of the dominance relationship judgment item is 0.6, the weight of the displacement velocity deviation item is 0.4, the dominance relationship judgment item value of the candidate solution is 3, and the displacement velocity deviation item value is 8.745, then the comprehensive evaluation value is: 0.6×3+0.4×8.745=5.298. After calculating the evaluation values ​​of all candidate solutions on the Pareto front surface, the candidate solution with the smallest evaluation value is selected as the optimal control parameter vector. In actual applications, there may be dozens of candidate solutions on the Pareto front surface. By screening out the solution with the smallest evaluation value through the above method, the optimal control parameter vector can be determined.

[0154] A displacement competition function is constructed based on the optimal control parameter vector, which introduces an adaptive weight coefficient to balance the system response. The adaptive weight coefficient includes an exponential decay term and a deviation integral term, which are calculated based on the cumulative effect of the displacement deviation. In one embodiment, the exponential decay term adopts a negative exponential form. When the displacement deviation continues to decrease, the value of the exponential decay term will gradually increase, thereby enhancing the control effect; when the displacement deviation is large, the value of the exponential decay term is small, weakening the control effect. For example, if the displacement deviation is 5 cm and the attenuation coefficient is 0.2, the value of the exponential decay term is e (-0.2×5) =0.368.

[0155] The deviation integral term accounts for the cumulative effect of historical displacement deviations by integrating the displacement deviations over time. In one embodiment, if the system sampling interval is 0.1 seconds and the displacement deviations at the five most recent sampling points are 4.2 cm, 3.8 cm, 3.5 cm, 3.2 cm, and 2.9 cm, respectively, the value of the deviation integral term is: (4.2 + 3.8 + 3.5 + 3.2 + 2.9) × 0.1 = 1.76.

[0156] The displacement competition function also requires determining the competition radius, which is determined by the ratio of displacement deviation to velocity deviation. In practice, a large ratio indicates that the system is in the displacement adjustment phase, in which case the competition radius is increased; a small ratio indicates that the system is in the velocity adjustment phase, in which case the competition radius is decreased. For example, if the displacement deviation is 6 cm, the velocity deviation is 1.5 cm / s, the base competition radius is 4, and the adjustment coefficient is 0.5, then the actual competition radius is: 4 + (6 / 1.5) × 0.5 = 6.

[0157] A velocity coordination function is constructed based on the optimal control parameter vector. This function introduces a dynamic coordination coefficient, calculated based on the coupling relationship between the displacement deviation and the velocity deviation. In one embodiment, when the displacement deviation and the velocity deviation are in the same direction, the coordination effect is enhanced; when the displacement deviation and the velocity deviation are in opposite directions, the coordination effect is weakened. For example, if the displacement deviation is a positive value of 4 cm, the velocity deviation is a positive value of 1 cm / s, the basic coordination coefficient is 0.8, and the coupling adjustment factor is 0.1, then the dynamic coordination coefficient is: 0.8 + 0.1 × 4 × 1 = 1.2.

[0158] The synergy radius of the speed synergy function is determined by the rate of change of the speed deviation. A larger rate of change indicates a more drastic system speed adjustment, and in this case, the synergy radius is increased; a smaller rate of change indicates a more stable system speed adjustment, and in this case, the synergy radius is decreased. In one embodiment, if the current sampling point speed deviation is 2.5 cm / s, the previous sampling point speed deviation is 3.2 cm / s, the sampling interval is 0.1 seconds, the base synergy radius is 5, and the adjustment coefficient is 0.3, then the actual synergy radius is: 5 + 0.3 × |(2.5 - 3.2) / 0.1| = 7.1.

[0159] In the prior art, photovoltaic cleaning crawler robots often use fixed pressure control parameters when working under different terrain conditions. They are unable to dynamically adjust the pressure according to terrain changes and operating conditions, resulting in poor cleaning effects or damage to photovoltaic panels. Traditional methods usually adopt a single-objective optimization strategy, which makes it difficult to meet the requirements of displacement accuracy and speed response at the same time. The method of this embodiment is based on the Pareto multi-objective optimization principle, constructs a comprehensive evaluation function, and takes into account the two aspects of dominance relationship determination and displacement speed deviation, thereby achieving a balance between multiple optimization objectives. By introducing the displacement competition function and the speed coordination function, the control parameters are adaptively adjusted according to the real-time status of the system, which significantly improves the accuracy and stability of pressure control; effectively avoids excessive pressure damage to the photovoltaic panel, while ensuring the stability of the cleaning quality.

[0160] The photovoltaic cleaning crawler robot pressure intelligent regulation and protection system of the embodiment of the present invention includes:

[0161] The first unit is used to process the tracked vehicle pressure signal using a dual-time-scale analysis method. The collected pressure data is decomposed into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network, the impact feature matrix is ​​generated, and a pressure baseline trend vector is constructed.

[0162] The second unit is used to perform variational mode decomposition and joint analysis on the impact characteristic matrix and the pressure baseline trend vector, establish a baseline compensation model, calculate the energy loss based on the surface load response characteristics of the photovoltaic panel, and generate an adaptive pressure threshold;

[0163] The third unit is used to collect the real-time motion state parameters of the crawler vehicle and calculate the optimal relative displacement and optimal relative speed between the track wheel and the photovoltaic panel based on the adaptive pressure threshold;

[0164] The fourth unit is used to construct a multi-objective optimization controller based on the optimal relative displacement and the optimal relative speed value by using ranking learning, and to adjust the output characteristics of the tracked vehicle through a competition-cooperation mechanism.

[0165] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0166] processor;

[0167] a memory for storing processor-executable instructions;

[0168] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0169] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0170] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The intelligent pressure regulation and protection method of the photovoltaic cleaning crawler robot is characterized by: include: A dual-time-scale analysis method is used to process the tracked vehicle pressure signal. The collected pressure data is decomposed into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network, an impact feature matrix is ​​generated, and a pressure baseline trend vector is constructed. The impact characteristic matrix and pressure baseline trend vector are subjected to variational mode decomposition and joint analysis to establish a baseline compensation model. The energy loss is calculated based on the surface load response characteristics of the photovoltaic panel to generate an adaptive pressure threshold. Collect the real-time motion state parameters of the crawler vehicle and calculate the optimal relative displacement and optimal relative speed between the track wheel and the photovoltaic panel based on the adaptive pressure threshold; Based on the optimal relative displacement and relative speed, a multi-objective optimization controller is constructed using ranking learning, and the output characteristics of the tracked vehicle are adjusted through a competition-cooperation mechanism.

2. The method according to claim 1, characterized in that The dual-time-scale analysis method is used to process the pressure signal of the tracked vehicle. The collected pressure data is decomposed into pressure baseline and transient impact. The multi-dimensional feature map of transient impact is extracted through the chaotic twin neural network to generate the impact feature matrix and construct the pressure baseline trend vector, including: The pressure signal of the photovoltaic cleaning crawler robot is subjected to multi-scale entropy analysis to obtain a time scale demarcation threshold, and the optimal time scale is determined according to the inflection point of the entropy curve of the time scale demarcation threshold. The pressure signal is integrated using the optimal time scale to obtain a slow time scale pressure baseline component, and the slow time scale pressure baseline component is subtracted from the pressure signal to obtain a fast time scale transient impact component. The fast time scale transient impact component is reconstructed by cross-correlation analysis and false neighbor analysis to determine the reconstruction parameters, determine the phase space matrix, extract the topological eigenvector from the phase space matrix, and input it into the chaotic twin neural network, and obtain the chaotic feature representation through contrast loss function training; The chaotic feature representation is constructed into a transient impact feature matrix, and the slow time scale pressure baseline component is averaged within a sliding time window to obtain a pressure baseline trend vector. The transient impact feature matrix and the pressure baseline trend vector constitute the characteristic representation of the pressure signal of the photovoltaic cleaning crawler vehicle robot.

3. The method according to claim 2, characterized in that The fast time scale transient impact component is reconstructed by cross-correlation analysis and false neighbor analysis to determine the reconstruction parameters, determine the phase space matrix, extract the topological eigenvector from the phase space matrix, and input it into the chaotic twin neural network. The chaotic feature representation is obtained by contrast loss function training, including: The fast time scale transient impact component is calculated by cross-correlation analysis to obtain a candidate reconstruction delay sequence, the optimal reconstruction delay is determined based on the cross-correlation function of the candidate reconstruction delay sequence, and the minimum embedding dimension is obtained by using the optimal reconstruction delay through false neighbor analysis; Based on the optimal reconstruction delay and minimum embedding dimension, the fast time-scale transient impact component is reconstructed into a phase space matrix. The phase space dispersion is obtained by calculating the Euclidean distance between adjacent reconstructed vectors in the phase space matrix, and the nearest neighbor search radius is determined. The local density distribution of the reconstructed vector is obtained by calculation within the nearest neighbor search radius. Performing density clustering on the local density distribution to obtain a set of feature points, calculating the topological relationship between the feature points, and constructing a topological feature vector; The topological feature vector is input into the chaotic twin neural network, a nonlinear transformation is performed through the feature encoder to obtain an initial feature map, and the Mahalanobis distance between the features is calculated through the metric learner; A contrast loss function is constructed based on the Mahalanobis distance, and the chaotic twin neural network is trained to obtain an optimized feature mapping model, which is then used to extract the final chaotic feature representation.

4. The method according to claim 1, wherein The impact characteristic matrix and pressure baseline trend vector are subjected to variational mode decomposition and joint analysis to establish a baseline compensation model. The energy loss is calculated based on the surface load response characteristics of the photovoltaic panel, and the adaptive pressure threshold is generated, including: Performing spectral analysis on the impact characteristic matrix and the pressure baseline trend vector to obtain a spectral energy distribution, calculating the energy coverage based on the spectral energy distribution, determining the number of decomposition levels of variational modal decomposition, and decomposing the impact characteristic matrix and the pressure baseline trend vector according to the number of decomposition levels to obtain a first modal function sequence and a second modal function sequence; Based on the first modal function sequence and the second modal function sequence, multi-scale decomposition is performed, the mutual information matrix, distance correlation coefficient and maximum information coefficient are calculated, a comprehensive correlation matrix is ​​constructed, mode pairs are selected according to the corresponding eigenvalues, and a compensation function is constructed; A stress-strain relationship is established based on the elastic coefficient tensor and viscosity coefficient tensor of the photovoltaic panel surface, and an integration operation is performed in the time domain to obtain the energy loss per unit area. The energy loss per unit area is integrated over the photovoltaic panel surface to obtain the total energy loss. A comprehensive evaluation function including a compensation error term, an energy loss term, and a threshold gradient constraint term is constructed, the output value of the compensation function and the total energy loss are input into the comprehensive evaluation function, and an adaptive pressure threshold is generated through optimization using a dynamic programming algorithm.

5. The method according to claim 4, characterized in that Based on the first modal function sequence and the second modal function sequence, multi-scale decomposition is performed, the mutual information matrix, distance correlation coefficient and maximum information coefficient are calculated, a comprehensive correlation matrix is ​​constructed, mode pairs are selected according to the corresponding eigenvalues, and a compensation function is constructed including: Performing multi-scale decomposition on the first modal function sequence and the second modal function sequence to obtain first modal components and second modal components of different scales, calculating a joint probability distribution based on the first modal components and the second modal components, and constructing a multi-scale mutual information matrix using the joint probability distribution; Calculating a dual-centered distance matrix for the first modal function sequence and the second modal function sequence, calculating a distance correlation coefficient based on the dual-centered distance matrix, and gridding the first modal function sequence and the second modal function sequence to obtain a maximum information coefficient; The multi-scale mutual information matrix, the distance correlation coefficient, and the maximum information coefficient are weighted and combined using a preset fusion coefficient to construct a comprehensive correlation matrix, the comprehensive correlation matrix is ​​subjected to eigenvalue decomposition to obtain eigenvectors, and a modal pair is determined based on the eigenvectors; The deviation between the comprehensive correlation matrix and a preset reference correlation degree is calculated, an adaptive weight coefficient is constructed based on the deviation, and the adaptive weight coefficient is combined with the output of the kernel function mapping to generate a compensation function.

6. The method according to claim 1, characterized in that Based on the optimal relative displacement and the optimal relative speed, a multi-objective optimization controller is constructed using ranking learning. The output characteristics of the crawler are adjusted through a competition-cooperation mechanism, including: generating a control parameter vector including a proportional parameter, an integral parameter, a differential parameter, and a compensation parameter, and constructing a multidimensional candidate solution space based on the control parameter vector; Performing non-dominated sorting on candidate solutions in the multidimensional candidate solution space, calculating the optimal relative displacement deviation and the optimal relative velocity value deviation corresponding to each candidate solution, determining the dominance number of each candidate solution, and constructing a Pareto frontier surface based on the dominance number; The optimal control parameter vector is selected based on the Pareto frontier, and a displacement competition function including an adaptive weight coefficient and a speed coordination function including a coordination coefficient are constructed. The displacement competition function and the speed cooperation function are combined through a competition-cooperation balance factor to generate a mechanism fusion function, and the mechanism fusion function is combined with the optimal control parameter vector to construct an optimized control law; An output characteristic compensation amount is calculated based on the optimized control law, the output characteristic compensation amount is combined with the dominant relationship in the Pareto front surface, an adaptive adjustment function is constructed, and the output characteristic of the actuator is adjusted according to the adaptive adjustment function.

7. The method according to claim 6, characterized in that The optimal control parameter vector is selected based on the Pareto frontier, and the displacement competition function including the adaptive weight coefficient and the speed coordination function including the coordination coefficient are constructed. A comprehensive evaluation function is constructed based on candidate solutions in the Pareto front surface, wherein the comprehensive evaluation function includes a dominance relationship determination term and a displacement-velocity deviation term, wherein the dominance relationship determination term is the number of times the candidate solution is dominated by other candidate solutions, and the displacement-velocity deviation term is the weighted sum of the squared displacement deviation term and the squared velocity deviation term corresponding to the candidate solution; the dominance relationship determination term and the displacement-velocity deviation term are weightedly combined using a preset weight factor to obtain an evaluation value, and the candidate control parameter vector with the smallest evaluation value is selected from the Pareto front surface to determine the optimal control parameter vector; A displacement competition function is constructed based on the optimal control parameter vector, and an adaptive weight coefficient is calculated according to the cumulative effect of the displacement deviation, wherein the adaptive weight coefficient includes an exponential decay term and a deviation integral term, and a competition action radius is determined based on the ratio of the displacement deviation to the speed deviation; A speed coordination function is constructed based on the optimal control parameter vector, a dynamic coordination coefficient is calculated according to the coupling relationship between the displacement deviation and the speed deviation, and a coordination radius is determined based on the speed deviation change rate.

8. A photovoltaic cleaning crawler robot pressure intelligent regulation and protection system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to process the tracked vehicle pressure signal using a dual-time-scale analysis method. The collected pressure data is decomposed into a pressure baseline and transient impact. The multi-dimensional feature map of the transient impact is extracted through a chaotic twin neural network, the impact feature matrix is ​​generated, and a pressure baseline trend vector is constructed. The second unit is used to perform variational mode decomposition and joint analysis on the impact characteristic matrix and the pressure baseline trend vector, establish a baseline compensation model, calculate the energy loss based on the surface load response characteristics of the photovoltaic panel, and generate an adaptive pressure threshold; The third unit is used to collect the real-time motion state parameters of the crawler vehicle and calculate the optimal relative displacement and optimal relative speed between the track wheel and the photovoltaic panel based on the adaptive pressure threshold; The fourth unit is used to construct a multi-objective optimization controller based on the optimal relative displacement and the optimal relative speed value by using ranking learning, and to adjust the output characteristics of the tracked vehicle through a competition-cooperation mechanism.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images