Photovoltaic panel pollution classification system and method based on unmanned aerial vehicle and semantic communication

Abstract

The application discloses a photovoltaic panel pollution classification system and method based on unmanned aerial vehicle and semantic communication. The unmanned aerial vehicle collects high-resolution images of the photovoltaic panel, and converts the high-resolution images into compact task-related feature representations through an on-board feature encoder, so as to greatly reduce the data volume. Then, the features are converted into complex-valued symbols through weighted mapping, and are subjected to channel-aware precoding processing, so as to realize adaptation to the channel state. Subsequently, the processed signals are transmitted to a remote data center, and after signal recovery, a classifier module is used to realize accurate classification of the pollution type. Through extraction and transmission of low-dimensional task-related features, and by considering the feature importance weighting and pairing embedding enhancement of channel awareness, the importance of different features to the classification result is determined, resource optimization allocation is realized according to the importance, and high-precision and high-robustness pollution state recognition under the limited bandwidth and power constraints is realized.

Description

Technical Field

[0001] This invention relates to photovoltaic panel pollution treatment, specifically to a photovoltaic panel pollution classification system and method based on unmanned aerial vehicles and semantic communication, belonging to the interdisciplinary fields of wireless communication technology, computer vision, and photovoltaic operation and maintenance. Background Technology

[0002] With the expansion of photovoltaic power plants and their widespread application in harsh environments such as arid regions and deserts, the problem of reduced power generation efficiency due to contaminants on photovoltaic panels has become increasingly prominent. Different types of contaminants (such as sand and bird droppings) have varying impacts on photovoltaic panels, thus requiring different cleaning methods and levels of urgency. Therefore, accurate classification of photovoltaic contaminants is crucial for the normal operation and routine maintenance of photovoltaic panels. Traditional inspection methods rely on manual or drone-based high-definition image capture and transmission to a data center for manual or automated analysis, which presents the following problems:

[0003] 1. High communication bandwidth pressure: High-definition images involve large amounts of data, resulting in low transmission efficiency and long transmission times in remote areas or environments with limited bandwidth;

[0004] 2. Limited onboard computing resources: If real-time image recognition is performed on the drone, it is difficult to deploy complex models due to limitations in computing power, battery capacity, and heat dissipation.

[0005] 3. Existing task-oriented communication methods are highly coupled: Although some studies have proposed reducing data transmission through semantic communication, most of them adopt end-to-end joint training, which leads to tight coupling between system modules, poor interpretability, difficulty in maintenance and updating, and large training data requirements and high computational costs.

[0006] 4. Inappropriate allocation of transmission resources: Existing feature transmission methods often fail to consider the differences in importance of different features to the classification task, and do not dynamically allocate resources in conjunction with the wireless channel state, resulting in a sharp decline in classification performance under adverse channel conditions.

[0007] Therefore, there is an urgent need for a smart inspection solution for photovoltaic panel pollution that can significantly reduce communication overhead, maintain high classification accuracy in complex wireless environments, and be easy to deploy and maintain on the drone side. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, the present invention aims to propose a photovoltaic panel contamination classification system and method based on UAVs and semantic communication. This invention extracts and transmits low-dimensional task-related features, combines feature importance weighting considering channel awareness with paired embedding enhancement to determine the importance of different features for the classification results, and optimizes resource allocation based on importance, thereby achieving high-precision and robust contamination state identification under limited bandwidth and power constraints.

[0009] The technical solution of this invention is implemented as follows:

[0010] A photovoltaic panel pollution classification system based on unmanned aerial vehicles (UAVs) and semantic communication includes a UAV and a data center; the UAV includes an image acquisition module, a lightweight feature encoder, a channel-aware weighting module, a complex value encoding module, a pre-encoder, and a transmission module; wherein,

[0011] Image acquisition module: used to acquire images of photovoltaic panels;

[0012] Lightweight feature encoder: Used to extract compact multidimensional features from photovoltaic panel images, including color / reflectivity, texture / edge, shape / coverage, and stability enhancement features;

[0013] Channel-aware weighting module: used to calculate the inter-class discriminative power of each feature dimension, and combined with the estimated channel noise, to generate feature importance weights, so as to give greater weight to feature dimensions with high inter-class discriminative power, while maintaining stability against intra-class noise, channel fluctuations or measurement disturbances.

[0014] Complex value encoding module: Based on the correlation between features and the weight of feature importance, it performs optimal pairing of real value feature pairs and performs coordinate system rotation in each pairing subspace to generate a complex value symbol stream, so as to enhance the separability of features after transmission;

[0015] Precoder: With the goal of minimizing the weighted mean square error, under the constraint of total transmit power, the optimal precoding matrix is ​​solved, and the transmit resources are tilted towards high-weight features;

[0016] Transmitting module: Used to transmit pre-coded signals to the data center via a wireless channel;

[0017] The data center includes a receiving module, an equalization module, and a classifier module; among them,

[0018] Receiver module: Used to receive signals transmitted from the transmitter module;

[0019] Equalization module: used to recover the received signal and input it into the pre-trained classifier module;

[0020] Classifier module: Used to classify the pollution categories of photovoltaic panels.

[0021] This invention also provides a method for classifying photovoltaic panel pollution based on unmanned aerial vehicles (UAVs) and semantic communication, comprising the following steps:

[0022] 1) Use drones to collect images of photovoltaic panels;

[0023] 2) Extract compact multidimensional features from photovoltaic panel images, including color / reflectivity, texture / edge, shape / coverage, and stability enhancement features;

[0024] 3) For the extracted multidimensional features, calculate the inter-class discriminant of each dimension feature, and combine it with the estimated channel noise to generate feature importance weights, so as to give greater weight to feature dimensions with high inter-class discriminant and stable intra-class noise, channel fluctuations or measurement disturbances.

[0025] 4) Based on the correlation and importance weights between features, multidimensional features are paired to form complex arrays, with each pair of features forming the real and imaginary parts of a complex number. During pairing, the pairing is biased towards features with higher correlation and higher weights. Then, coordinate system rotation is performed in each paired subspace to obtain new complex arrays, thereby generating a complex value symbol stream. The purpose of coordinate system rotation is to enhance the separability of features after transmission. And under the global power budget, amplitude and phase resources are preferentially allocated to high-weight coordinates.

[0026] 5) With the goal of minimizing the weighted mean square error, under the constraint of total transmit power, the optimal precoding matrix is ​​solved by precoding, and transmit resources are allocated according to the weights of the new complex numbers obtained after rotation in step 4;

[0027] 6) Transmit the pre-coded signal and allocated transmission resources to the data center via a wireless channel;

[0028] 7) The data center uses a corresponding linear equalizer to recover the received signal and inputs it into a pre-trained classifier. The classifier then outputs the photovoltaic panel pollution category.

[0029] Furthermore, steps 1) through 6) are performed by modules corresponding to each step; the modules corresponding to steps 1) through 6) are image acquisition module, lightweight feature encoder, channel-aware weighting module, complex value encoding module, pre-encoder and transmission module, and all of them are set on the UAV.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. Significantly improved communication efficiency: This invention transmits only optimized low-dimensional features instead of the original image, reducing the amount of data by more than 99.98%, which greatly alleviates the uplink bandwidth pressure and reduces transmission latency and energy consumption.

[0032] 2. High classification accuracy and robustness: By using the channel-aware Fisher weighted (CAFW) mechanism, the semantic discriminative power of features is combined with channel noise statistics to ensure that key information is given priority protection during transmission.

[0033] 3. Lightweight and easy-to-deploy system: Adopting a modular design, each functional module (feature extraction, weighting, encoding, precoding, classification) can be designed and optimized independently, eliminating the need for complex end-to-end joint training. The feature extractor is lightweight, making it suitable for edge devices such as drones.

[0034] 4. Intelligent resource allocation: Through weighted mean square error precoding (WMSE precoding), the limited transmit power and spatial degrees of freedom are dynamically and accurately allocated to the feature substreams most important to the classification task, thereby improving the system's survivability under harsh channel conditions.

[0035] 5. Good interpretability and maintainability: The parameters of each module of the system have clear physical meaning (such as feature weights and rotation angles), which facilitates debugging, updating and cross-scene migration.

[0036] 6. Enhanced data privacy: The transmitted data is abstract features rather than the original image, reducing the risk of sensitive visual information being leaked during transmission. Attached Figure Description

[0037] Figure 1 - A schematic diagram of the system structure and classification process of this invention.

[0038] Figure 2 - Schematic diagram of pollutant types in this invention.

[0039] Figure 3 - A schematic diagram illustrating the importance weights of different features in this embodiment of the invention.

[0040] Figure 4 - A diagram illustrating the relationship between the number of features and their corresponding weights under different methods.

[0041] Figure 5 - Accuracy comparison curve between the present invention and existing methods. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0043] See Figure 1 This invention discloses a photovoltaic panel pollution classification system based on unmanned aerial vehicles (UAVs) and semantic communication, comprising a UAV and a data center. The UAV is equipped with an image acquisition module, a lightweight feature encoder, a channel-aware weighting module, a complex value encoding module, and a precoding and transmission module. The data center includes a signal receiving module, an equalization module, and a classifier module.

[0044] The drone first acquires high-resolution images of the photovoltaic panels using its image acquisition module, and then converts these images into compact, task-relevant feature representations using an onboard lightweight feature encoder to significantly reduce the amount of data. Next, these features are weighted and mapped to complex-valued symbols, and then undergo channel-aware precoding to adapt to the channel conditions. Subsequently, the processed signal is transmitted wirelessly to a remote data center. After signal recovery by an equalization module, a pre-trained classifier module accurately classifies the pollution type. This process operates within a modular architecture, with each stage dynamically coordinated through feature weights, ensuring efficient transmission and high-precision inference under limited communication resources.

[0045] The specific workflow is as follows:

[0046] 1) Images of photovoltaic panels taken by drones.

[0047] 2) Lightweight Feature Encoder: Extracts compact multi-family features from images, including color / reflectivity, texture / edge, shape / coverage, and stability enhancement features. It is composed of well-designed image processing algorithms (such as color moments, Local Binary Pattern (LBP), Histogram of Oriented Gradients (HOG), etc.), eliminating the need for large neural networks. The example provides a total of 48 features, offering high computational efficiency and avoiding the problems of large computational loads and difficulty in lightweight deployment associated with traditional neural networks.

[0048] 3) Channel-Aware Fisher Weighted (CAFW) Module: Based on the training data, it calculates the inter-class discriminative power for each feature dimension and, combined with the estimated channel noise, generates a feature importance weight vector. Its purpose is to assign greater weights to feature dimensions that have high inter-class discriminative power while remaining stable against intra-class noise, channel fluctuations, or measurement perturbations. This weighted vector is estimated only once on the training set and then kept fixed thereafter. Let c be the set of sample indices belonging to category c, where ,and The empirical prior probability is For the k-th normalized feature dimension, its class-conditional sample mean and variance are:

[0049]

[0050]

[0051] The prior weighted global mean is

[0052]

[0053] in Let represent the normalized value of the k-th feature of sample i. These statistics are estimated only once on the training set and remain fixed during the validation / testing phase to prevent data leakage. The intra-class divergence and between-class divergence for each dimension are defined as follows:

[0054]

[0055]

[0056] This prior-weighted representation enhances robustness to class imbalance and reduces the variance of the divergence estimate used by CAFW. Therefore, the channel-aware Fisher ratio for each dimension is...

[0057]

[0058] Here, term C is a smaller, data-scale value used to gently regularize the denominator, defined as follows:

[0059]

[0060] here Controlling the magnitude of the basis relative to typical intra-class divergence, C is a tiny numeric protection term that ensures the denominator is strictly positive and numerically stable. C prevents denominator collapse at high signal-to-noise ratios and low variance tails and suppresses rare outliers. Because C is sufficiently small relative to typical divergence, it stabilizes weights across different data subsets and signal-to-noise ratio levels without altering the weight order, thus improving robustness.

[0061] In formula (6), This represents the dimension-dependent transport / measurement noise variance. This noise variance is estimated only once on the training set and remains constant throughout the validation / testing process. The terms, combined with the inherent intra-class divergence, form a heteroscedastic, channel-aware discriminator.

[0062] When a link estimator (such as an MMSE or ZF equalizer) can provide the mean squared error per dimension or per substream, robust per-dimensional noise estimation can be obtained by taking the median. Alternatively, when the channel matrix H and the transmit covariance matrix... When known, the mean squared error covariance matrix of each substream can be estimated in closed-form using linear MMSE:

[0063]

[0064] And using the diagonal elements of this matrix The mean square error for each substream. During the weighting phase, a uniform baseline independent of pairing can be used, defined as...

[0065]

[0066] This estimation method decouples noise estimation from the downstream mapping process while preserving the accuracy of noise estimation. The dependency. When neither estimator logs nor H are available, a fallback based on signal-to-noise ratio can be adopted. The proportional sensing agent, i.e. In all cases, it is possible to... Slight tail reduction is performed to suppress rare spikes. If the reported error is the mean squared error at the original feature scale, it should be divided by the square of the standard deviation of that feature on the training set to align to the normalized scale before being used in Equation (6).

[0067] To suppress outliers while preserving moderate tail behavior, we first analyze the channel-aware scores computed on the training set. Perform tail reduction processing:

[0068]

[0069] in express The percentile, These are the lower / upper percentile hyperparameters. Subsequently, a power-law mapping is applied. Normalization:

[0070]

[0071] in It is a precision protection term, which can be ignored when the denominator is non-zero; exponent

[0072]

[0073] Used to control the degree of concentration of the weight distribution. The larger the value, the more concentrated the weight distribution (peaking); The smaller the value, the more dispersed the distribution. Therefore, we can select values ​​on the training set using a one-dimensional binary search. To find a target effective engagement level And will eventually get The value will be used consistently in subsequent tests.

[0074] The CAFW method proposed in this invention integrates prior-weighted class divergence and link-level noise statistics, supplemented by lightweight stability control, thereby generating interpretable, contrast-adjustable, and normalized weights with well-controlled participation. By pruning extreme outliers while preserving information-rich tails, and calibrating to the target effective participation through contrast exponents, this scheme prevents a few unstable dimensions from dominating, improves numerical conditions, and can reliably generalize to realistic signal-to-noise ratio ranges and downstream classifiers.

[0075] 4) A complex-valued encoder based on Paired Embedding Enhancement (PEE): Real-valued feature pairs are optimally paired based on feature correlation and CAFW weights, and Fisher-Discriminative Rotation is performed within each paired subspace to generate a complex-valued symbol stream, enhancing post-transmission separability of features. Following the CAFW process, the obtained weights can be used to instantiate an R2C (real-to-complex) encoder, which prioritizes amplitude and phase resources for high-weight coordinates under a global power budget. This coupling improves robustness to channel impairments and maintains post-transmission separability without requiring sample-by-sample reweighting.

[0076] R2C coding pairs real-valued coordinates into a two-dimensional subspace. Traditional pairing methods, such as fixed adjacency indices, random pairing, and PCA block grouping, either ignore tag information structure and channel statistics, or suffer from high variance and poor reproducibility, or become vulnerable across multiple SNR intervals. In fact, pairing highly correlated coordinates produces elliptical class-conditional profiles, and Fisher rotation can significantly improve the ratio of inter-class to intra-class variations. In contrast, pairing nearly independent coordinates offers minimal performance improvement. To address these limitations, this invention proposes a PEE encoder that uses stabilized correlation and Fisher weights to drive global pairing assignment of features, aiming to employ more strongly correlated combinations during the transformation from real to complex numbers, thereby improving robustness.

[0077] Based on the normalized training matrix The empirical correlation can be calculated using the following formula:

[0078]

[0079] Where N is the number of samples; because the data sample is limited, to avoid some related coincidence factors, this invention applies Ledoit-Wolf shrinkage to shrink towards the identity matrix to a certain extent, thereby improving its reliability. The formula is as follows.

[0080]

[0081] in Indicates shrinkage strength. The identity matrix is ​​used; its strength is estimated once on the training set using the Ledoit-Wolf closed-form formula and then kept constant. The diagonal of the identity matrix is ​​all 1s, and the rest are all 0s (i.e., completely uncorrelated). Ledoit-Wolf shrinkage compresses the eigenvalue range while ensuring positive definiteness, which helps to remove some weakly correlated eigenvalues ​​and improves overall reliability, thus improving the condition number under limited sample size and link noise.

[0082] To emphasize the strong and signed dependency, a contrast enhancement operation is then applied:

[0083]

[0084] in It represents the Hadamah accumulation. It is a contrast index used to highlight strong correlations and weaken weak correlations. Formula (15) aims to preserve... The sign of each element is taken, and then the absolute value is raised to the power of γ. This ensures that strong correlations remain strong, while weak correlations become even weaker, which is helpful in identifying truly strong correlations. Therefore, given a correlation... and dataset-level weights The selection can be achieved by solving the one-to-one assignment problem. Disjoint pairings: The following objective function can achieve optimal pairings for all real-valued features.

[0085] This problem is solved using the Hungarian algorithm, where Controlling the influence of the learned weight distribution, The magnitude of the value represents the importance of the weight in this process. A larger value indicates a greater emphasis on weight and a less emphasis on relevance. This construction method tends to select values ​​with larger weights. and larger The pairings reflect the strong dependence and high joint importance under the learned weight distribution.

[0086] For each selected pair This invention summarizes the importance of the learned weights using the geometric mean:

[0087]

[0088] This choice achieves a pairing where both coordinates are important and is less sensitive to outliers than the arithmetic mean.

[0089] To select the rotation angle, this invention calculates the between-class and within-class scatter matrices in paired two-dimensional subspaces. and Under a fixed power budget, the allowed energy transformations in two-dimensional space remain orthogonal while maintaining linearity. Therefore, it is possible to maximize the one-dimensional projection. Fisherby, of which It is a unit direction vector:

[0090]

[0091] in Regularized inner terms. Maximized These are the dominant generalized eigenvectors for the following generalized eigenvalue problems:

[0092]

[0093] remember The optimal rotation matrix can be obtained.

[0094]

[0095] in ,and Thus, a power-preserving mapping is obtained. Complex value symbol Finally, we stack all the complex value symbols as... For subsequent beamforming processing.

[0096] 5) Weighted Mean Square Error (WMSE) Precoder: Aiming to minimize the weighted mean square error of CAFW, the optimal precoding matrix is ​​solved under the constraint of total transmit power, tilting transmit resources towards high-weight features. The weighted mean square error (WMSE) metric for the transmitted symbol vector is defined as:

[0097]

[0098] in It is a diagonal weight matrix; , These refer to the signals received at the receiver and the signals transmitted at the transmitter, respectively. When the extracted feature signals are transmitted through a wireless channel, they are inevitably affected by the communication channel and additive noise. Therefore, the signals received by the satellite cannot accurately replicate the transmitted feature vectors, but rather reflect a perturbed version of the random characteristics of the channel matrix and noise. This difference introduces uncertainty into the inference process. To mitigate the impact of communication channel interference and noise on the feature vectors, a linear equalizer is used. For the received signal Processing is performed to recover the transmitted symbol vector, that is:

[0099]

[0100] In this context, WMSE can be recalculated as

[0101]

[0102] in express The covariance matrix can be estimated using training data samples. Therefore, the WMSE minimization problem can be concisely formulated as follows:

[0103]

[0104] This constraint ensures that the total transmission power consumed by the drone does not exceed its allowed power budget; here This represents the trace operator for the enclosed matrix.

[0105] 6) The pre-coded signal is transmitted to the data center via a wireless channel.

[0106] 7) The data center uses the corresponding linear equalizer to recover the feature symbols and inputs them into a pre-trained classifier (such as a support vector machine, SVM) to obtain the contamination category.

[0107] The following are corresponding embodiments to further illustrate the present invention and its implementation effects.

[0108] This embodiment describes three types of pollutants: no pollution, white pollution (corresponding to bird droppings), and brown pollution (corresponding to dust). See [link / reference]. Figure 2 .

[0109] The present invention extracts 48 features from the images collected from each photovoltaic panel. Each feature represents the type of pollution (a type of pollution can often be reflected by multiple features, i.e., multiple features point to the same type of pollution) and also corresponds to different classification importance weights. The numbers in the boxes are the classification importance weights calculated by the method in step 3 of the present invention, and the color depth represents the importance to the result. The darker the color, the greater the weight and the more important it is to the classification result. Figure 3 This is a schematic diagram of the feature value weights in an embodiment of the present invention. Figure 3 This reflects that different feature values ​​have different degrees of influence on the classification results. When communication resources are limited, priority should be given to allocating features with high weights.

[0110] Figure 4This diagram illustrates the relationship between the number of features and their corresponding weights under different methods, representing the proportion of information that the number of features can express relative to the total information. Figure 4 As can be seen, the method proposed in this invention can prioritize the selection of feature values ​​that are more important to the classification results, retain key semantic information, and represent more information with fewer features.

[0111] Figure 5 This is a graph comparing the accuracy of the present invention with existing methods. Figure 5 This demonstrates that the method proposed in this invention can maintain better classification accuracy compared to other methods under low signal-to-noise ratio conditions (i.e., when communication resources are poor).

[0112] Finally, it should be noted that the above examples of the present invention are merely illustrative and not intended to limit the implementation of the invention. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art can make other variations and modifications based on the above description. It is impossible to exhaustively list all possible implementations here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A photovoltaic panel pollution classification system based on unmanned aerial vehicles (UAVs) and semantic communication, comprising UAVs and a data center; characterized in that: The UAV includes an image acquisition module, a lightweight feature encoder, a channel-aware weighting module, a complex value encoding module, a pre-encoder, and a transmission module; wherein... Image acquisition module: used to acquire images of photovoltaic panels; Lightweight feature encoder: Used to extract compact multidimensional features from photovoltaic panel images, including color / reflectivity, texture / edge, shape / coverage, and stability enhancement features; Channel-aware weighting module: used to calculate the inter-class discriminative power of each feature dimension, and combined with the estimated channel noise, to generate feature importance weights, so as to give greater weight to feature dimensions with high inter-class discriminative power, while maintaining stability against intra-class noise, channel fluctuations or measurement disturbances. Complex value encoding module: Based on the correlation between features and the weight of feature importance, it performs optimal pairing of real value feature pairs and performs coordinate system rotation in each pairing subspace to generate a complex value symbol stream, so as to enhance the separability of features after transmission; Precoder: With the goal of minimizing the weighted mean square error, under the constraint of total transmit power, the optimal precoding matrix is ​​solved, and the transmit resources are tilted towards high-weight features; Transmitting module: Used to transmit pre-coded signals to the data center via a wireless channel; The data center includes a receiving module, an equalization module, and a classifier module; among them, Receiver module: Used to receive signals transmitted from the transmitter module; Equalization module: used to recover the received signal and input it into the pre-trained classifier module; Classifier module: Used to classify the pollution categories of photovoltaic panels.

2. A photovoltaic panel pollution classification method based on unmanned aerial vehicles (UAVs) and semantic communication, characterized in that: Includes the following steps: 1) Use drones to collect images of photovoltaic panels; 2) Extract compact multidimensional features from photovoltaic panel images, including color / reflectivity, texture / edge, shape / coverage, and stability enhancement features; 3) For the extracted multidimensional features, calculate the inter-class discriminant of each dimension feature, and combine it with the estimated channel noise to generate feature importance weights, so as to give greater weight to feature dimensions with high inter-class discriminant and stable intra-class noise, channel fluctuations or measurement disturbances. 4) Based on the correlation and importance weights between features, multidimensional features are paired to form complex arrays, with each pair of features forming the real and imaginary parts of a complex number. During pairing, the pairing is biased towards features with higher correlation and higher weights. Then, coordinate system rotation is performed in each paired subspace to obtain new complex arrays, thereby generating a complex value symbol stream. The purpose of coordinate system rotation is to enhance the separability of features after transmission. And under the global power budget, amplitude and phase resources are preferentially allocated to high-weight coordinates. 5) With the goal of minimizing the weighted mean square error, under the constraint of total transmit power, the optimal precoding matrix is ​​solved by precoding, and transmit resources are allocated according to the weights of the new complex numbers obtained after rotation in step 4; 6) Transmit the pre-coded signal and allocated transmission resources to the data center via a wireless channel; 7) The data center uses a corresponding linear equalizer to recover the received signal and inputs it into a pre-trained classifier. The classifier then outputs the photovoltaic panel pollution category.

3. The photovoltaic panel pollution classification method based on UAV and semantic communication according to claim 2, characterized in that: Steps 1) through 6) are performed by the modules corresponding to each step. The modules corresponding to steps 1) through 6) are image acquisition module, lightweight feature encoder, channel-aware weighting module, complex value encoding module, pre-encoder and transmission module, and are all set on the UAV.

4. The photovoltaic panel pollution classification method based on UAV and semantic communication according to claim 2, characterized in that: In step 3), the feature importance weights are determined as follows: set up Let c be the set of sample indices belonging to category c, where ,and The empirical prior probability is For the k-th normalized feature dimension, its class-conditional sample mean and variance are: The prior weighted global mean is in The k-th feature of sample i represents the normalized value; the intra-class divergence and inter-class divergence for each dimension are defined as follows: The channel-aware Fisher ratio for each dimension is, in, The term C represents the dimension-dependent variance of transmission / measurement noise; it is a small, data-scale value used to gently regularize the denominator. Channel-aware Fisher ratio shortening: in express The percentile, It is the lower / upper percentile hyperparameter; Subsequently, a power-law mapping was applied. Normalization yields the feature importance weights; in It is a precision protection item, which is ignored when the denominator is non-zero; An index used to control the degree of concentration of weight distribution. ; The larger the value, the more concentrated the weight distribution; The smaller the value, the more dispersed the weight distribution.

5. A photovoltaic panel pollution classification method based on UAV and semantic communication according to claim 4, characterized in that: The term C is a small, data-scaled value used to gently regularize the denominator; term C is defined as follows: here Controlling the magnitude of the basis relative to typical intra-class divergence, It is a tiny numerical protection term that ensures that the denominator is strictly positive and the value is stable.

6. The photovoltaic panel pollution classification method based on UAV and semantic communication according to claim 1, characterized in that: In step 4), the complex-valued symbol stream is generated according to the following process: Based on the normalized training matrix The empirical correlation is calculated using the following formula: Where N is the number of samples; then, Ledoit-Wolf shrinkage is applied to shrink towards the identity matrix: in Indicates shrinkage strength. The identity matrix is ​​used; the strength is estimated once on the training set using the Ledoit-Wolf closed-form formula and then kept fixed; a contrast enhancement operation is then applied: in It represents the Hadamah accumulation. It is a comparative index used to highlight strong correlations and weaken weak correlations; Based on two characteristics correlation and weight , The selection is achieved by solving a one-to-one assignment problem using an objective function. Disjoint pairings; objective function is The objective function is solved using the Hungarian algorithm, where Controlling the impact of the learned weight distribution; For each selected pair The weights after pairing are calculated using the following formula: Calculate the between-class and within-class scatter matrices in paired two-dimensional subspaces. and Under a fixed power budget, the allowed energy linear transformations in two-dimensional space are orthogonal; maximizing the one-dimensional projection. Fisherby, of which It is a unit direction vector: in Regularized inner terms; maximized These are the dominant generalized eigenvectors for the following generalized eigenvalue problems: remember The optimal rotation matrix is ​​obtained. in ,and Thus, power-preserving mapping is obtained. Complex value symbol Finally, all complex value symbols are stacked as For subsequent beamforming processing.

7. A photovoltaic panel pollution classification method based on UAV and semantic communication according to claim 2, characterized in that: In step 5), the specific implementation process of the precoding is as follows: The weighted mean square error of the transmitted symbol vector is defined as: in It is a diagonal weight matrix; , These refer to the signals received at the receiving end and the signals transmitted at the transmitting end, respectively. Using a linear equalizer For the received signal Processing is performed to recover the transmitted symbol vector, that is: The weighted mean square error of the transmitted symbol vector is: in express The covariance matrix is ​​estimated using training data samples; The problem of minimizing the weighted mean square error of the transmitted symbol vector is formulated as follows: This constraint ensures that the total transmission power consumed by the drone does not exceed its allowed power budget; This represents the trace operator for the enclosed matrix.

Images