A Smart Image Analysis Platform for Photovoltaic Construction Drone Inspection

Abstract

This invention relates to the field of photovoltaic construction inspection technology, and discloses an intelligent image analysis platform for photovoltaic construction drone inspections. The platform collects real-time images, location, and environmental parameter information of the construction target using drones; retrieves historical inspection image datasets, classifies image features, extracts historical defects and analysis data, obtains a historical analysis score set, and optimizes it into a historical analysis score sequence; calculates the rate of change and stability parameters based on sequence analysis trends; and combines real-time environmental parameters with standard environmental parameters that match the feature categories to correct deviations, ultimately generating and outputting an analysis report. This platform combines historical data with real-time environmental parameters for analysis and decision-making, improving the accuracy and effectiveness of photovoltaic construction inspection analysis, and is suitable for quality inspection and analysis during photovoltaic construction.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic construction inspection technology, specifically to an intelligent image analysis platform for photovoltaic construction drone inspections. Background Technology

[0002] With the rapid development of the photovoltaic industry, the scale of photovoltaic construction projects is constantly expanding. Quality control and defect detection during construction have become crucial to ensuring the stable operation of photovoltaic systems. Currently, inspection work in photovoltaic construction areas relies heavily on manual inspections or traditional drone inspections, which have many limitations in practical applications.

[0003] Manual inspections not only consume significant manpower and time, but are also heavily influenced by the inspectors' experience, skill level, and subjective judgment, making them prone to omissions and misidentifications, and compromising the accuracy and consistency of inspection results. Furthermore, manual inspections struggle to provide comprehensive and efficient coverage of the construction area, particularly for photovoltaic construction sites with complex terrain or vast areas, where the problem of low inspection efficiency is even more pronounced.

[0004] While traditional drone inspections have improved coverage and efficiency to some extent, they have significant shortcomings in image analysis and processing. Current image analysis methods largely rely on manual post-processing of drone-captured images, lacking intelligent analysis tools and failing to quickly and accurately identify defects in construction targets. Furthermore, traditional methods often focus only on real-time image data, neglecting the value of historical inspection data. They cannot combine historical data for trend analysis, making it difficult to predict potential problems during construction and hindering timely and effective countermeasures.

[0005] The photovoltaic (PV) construction environment is complex and variable, and different environmental parameters (such as sunlight, temperature, and humidity) can significantly affect construction quality and defect detection. Existing technologies fail to fully consider the correlation between real-time environmental parameters and historical data during the analysis process, resulting in reduced accuracy and reliability of the analysis results, making it difficult to meet the high-quality and high-efficiency inspection requirements of PV construction. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent analysis platform for photovoltaic construction drone inspection images to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an intelligent image analysis platform for photovoltaic construction drone inspections, the platform comprising:

[0008] The image acquisition module is used to collect real-time image data of the construction target in the photovoltaic construction area using a drone, and to obtain the real-time location information and real-time environmental parameter information of the construction target.

[0009] The historical data indexing module is used to retrieve a set of historical inspection image data within a user's historical time period based on the identification information of the construction target.

[0010] The feature classification module is used to classify the image feature information in the historical inspection image data set, generate multiple classification feature categories, and extract the historical defect information set and historical analysis data set under the multiple classification feature categories to obtain multiple historical analysis score sets.

[0011] The analysis score optimization processing module is used to optimize the multiple historical analysis score sets respectively to obtain multiple optimized analysis score sets, and arrange them according to the time series to obtain multiple historical analysis score sequences;

[0012] The trend analysis module is used to perform trend analysis operations based on the multiple historical analysis score sequences, and calculate the rate of change parameter and stability parameter;

[0013] The correction module is used to match the real-time environmental parameter information with the multiple classification feature categories to obtain the matching feature category, and to perform correction calculations on the rate of change parameter and stability parameter based on the deviation between the real-time environmental parameter information and the standard environmental parameter information of the matching feature category to obtain the corrected rate of change parameter and corrected stability parameter.

[0014] The decision output module is used to perform analysis and decision operations based on the corrected rate of change parameter and the corrected stability parameter, generate an analysis report scheme, and output the analysis report.

[0015] Preferably, the image acquisition module acquires real-time image data of the construction target in the photovoltaic construction area using a drone, and obtains the real-time location information and real-time environmental parameter information of the construction target, including:

[0016] The drone's image sensor captures real-time image data of the construction target, while the drone's positioning device records the real-time location information of the construction target. It also integrates environmental sensors to collect real-time environmental parameter information of the construction target, including light intensity data, temperature data, and humidity data.

[0017] Preferably, the feature classification module classifies the image feature information within the historical inspection image dataset, generates multiple classification feature categories, and extracts historical defect information sets and historical analysis data sets under the multiple classification feature categories, processing to obtain multiple historical analysis score sets, including:

[0018] Multiple image feature information is extracted from the historical inspection image data set, and clustering and classification operations are performed to generate multiple classification feature categories;

[0019] Extract the historical defect information set under the multiple classification feature categories, and obtain the historical detection result information set and historical submission time set submitted by the user under the multiple classification feature categories;

[0020] Based on the degree of difference between the historical detection result information set and the historical defect information set, multiple basic analysis score sets are generated.

[0021] By utilizing the ratio between a preset time threshold parameter and the historical submission time set, adjustment calculations are performed on the multiple basic analysis score sets to generate multiple historical analysis score sets.

[0022] Preferably, the analysis score optimization processing module optimizes the multiple historical analysis score sets respectively to obtain multiple optimized analysis score sets, and arranges them according to time series to obtain multiple historical analysis score sequences, including:

[0023] Select one of the multiple historical analysis score sets as the current analysis score set, and determine the benchmark analysis score in the current analysis score set;

[0024] Based on the distance between other analysis scores in the current analysis score set and the benchmark analysis score, weight coefficients are assigned to generate a basic weight distribution, wherein the distance is inversely proportional to the weight coefficient.

[0025] Based on the aforementioned basic weight distribution, an optimization and compression operation is performed on the current set of analysis scores to obtain an optimized set of analysis scores.

[0026] The historical analysis score sequence is generated by sorting multiple analysis scores within the optimized analysis score set according to their timestamp information.

[0027] Repeat the optimization compression and sorting operations on other historical analysis score sets to generate multiple historical analysis score sequences.

[0028] Preferably, the analysis score optimization processing module performs an optimization compression operation on the current analysis score set based on the basic weight distribution to obtain an optimized analysis score set, including:

[0029] A preset number of analysis scores are randomly selected from the current set of analysis scores to form an initial set of optimization scores.

[0030] Based on the distance between the analysis scores within the initial optimized score set and the benchmark analysis scores, weight coefficients are assigned to generate an initial weight distribution;

[0031] Calculate the similarity metric between the initial weight distribution and the basic weight distribution, and use it as the optimization fitness value;

[0032] A preset number of analysis scores are randomly selected from the current set of analysis scores to form a set of candidate optimization scores, and the corresponding candidate fitness values ​​are calculated.

[0033] Continue performing optimization compression operations until the fitness values ​​converge, and output the candidate optimization score set with the largest optimization fitness value as the optimization analysis score set.

[0034] Preferably, the trend analysis module performs trend analysis operations based on the multiple historical analysis score sequences to calculate the rate of change parameter and stability parameter, including:

[0035] Collect sample inspection data sets from multiple users, obtain sample analysis score sequence sets, and calculate sample change rate set and sample stability set based on the analysis score change characteristics of each sample analysis score sequence;

[0036] A trend analysis model is constructed using the set of sample analysis score sequences as input data and the set of sample change rates and the set of sample stability as output data.

[0037] The trend analysis model is applied to process the multiple historical analysis score sequences to obtain multiple characteristic change rate parameters and multiple characteristic stability parameters.

[0038] The similarity values ​​between the real-time environmental parameter information and the multiple classification feature categories are analyzed. Based on the magnitude of the multiple similarity values, a weighted summation calculation is performed on the multiple feature change rate parameters and multiple feature stability parameters to generate change rate parameters and stability parameters.

[0039] Preferably, the correction module performs a matching operation between the real-time environmental parameter information and the multiple classification feature categories to obtain a matching feature category, and performs correction calculations on the rate of change parameter and stability parameter based on the deviation between the real-time environmental parameter information and the standard environmental parameter information of the matching feature category to obtain a corrected rate of change parameter and a corrected stability parameter, including:

[0040] Select the category with the highest similarity value as the matching feature category, and obtain the standard environmental parameter information of the matching feature category;

[0041] An adjustment coefficient is set based on the deviation between the real-time environmental parameter information and the standard environmental parameter information of the matching feature category;

[0042] The adjustment coefficient is applied to perform a correction calculation on the rate of change parameter and the stability parameter to generate a corrected rate of change parameter and a corrected stability parameter.

[0043] Preferably, the decision output module performs analysis and decision operations based on the corrected rate of change parameter and the corrected stability parameter, generates an analysis report scheme, and outputs an analysis report, including:

[0044] Obtain the set of sample corrected rate of change parameters and the set of sample corrected stability parameters. Based on the magnitude of each sample corrected rate of change parameter and sample corrected stability parameter, set a sample reporting scheme and generate a set of sample reporting schemes. The sample reporting scheme includes a prompt level parameter. The magnitude of the sample corrected rate of change parameter and the sample corrected stability parameter are inversely proportional to the magnitude of the prompt level parameter.

[0045] A decision model is constructed using the set of sample corrected rate of change parameters and the set of sample corrected stability parameters as decision input data, and the set of sample reporting schemes as decision output data.

[0046] The decision model is applied to process the corrected rate of change parameter and the corrected stability parameter to generate an analysis report scheme.

[0047] Preferably, the platform further includes a visualization module, which is used to receive the analysis report scheme generated by the decision output module, convert the analysis report scheme into visual chart information, and output it to the user interface;

[0048] The visualization module is connected to the decision output module. The analysis report scheme output by the decision output module serves as the input data for the visualization module. The visualization module processes the analysis report scheme to generate visualization chart information.

[0049] Preferably, the platform further includes:

[0050] An environmental interference filtering module is used to receive real-time environmental parameter information collected by the image acquisition module, identify interference scenarios such as sudden changes in light intensity, abnormal dust concentration, or sudden increase in humidity, and generate environmental interference labeling data.

[0051] The analysis and compensation module is used to locate the contaminated image region based on the environmental interference marker data, calculate the compensation coefficient by calling the correction stability parameter output by the correction module, and perform noise reduction and enhancement processing on the image features of the contaminated region.

[0052] The dynamic weight allocation module compares the noise-reduced image feature data with the historical defect information set generated by the feature classification module, dynamically allocates analysis weights for different defect types according to the construction stage, and outputs the weighted defect analysis results to the decision output module.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] By leveraging drones to acquire real-time image data, location information, and environmental parameters, dynamic capture of the photovoltaic construction area is achieved, allowing for timely presentation of the construction target's real-time status. The historical data index module retrieves historical inspection image data, and combined with the feature classification module's analysis of image features, the system can systematically integrate past defect information and analysis data to form a structured set of historical analysis scores.

[0055] The score optimization module optimizes the score set and arranges the time series, making the historical data present a clear trend of change. This provides well-structured data support for the trend analysis module to calculate the rate of change and stability parameters, facilitating a clear understanding of the long-term changing trends of construction objectives. The correction module incorporates real-time environmental parameters, matching feature categories and adjusting parameters based on deviations to eliminate the interference of environmental differences on the analysis results, making the parameters more closely match the actual construction scenario.

[0056] The decision output module generates analysis reports based on the corrected parameters, transforming complex data processing results into directly applicable reference content. This allows for the accurate identification of potential problems and development trends during construction. The entire process forms a closed loop from data acquisition to analysis and decision-making, organically integrating real-time information with historical data. It takes into account both the current situation and past patterns, making the analysis process more coherent and comprehensive, and the output reports more closely aligned with the actual needs of photovoltaic construction. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the working principle of the photovoltaic construction drone inspection image intelligent analysis platform described in this invention;

[0058] Figure 2 The flowchart for the feature classification module;

[0059] Figure 3 The flowchart for analyzing the score optimization processing module;

[0060] Figure 4 A flowchart for the trend analysis module;

[0061] Figure 5 The flowchart for the correction module. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see Figure 1 This invention provides an intelligent image analysis platform for photovoltaic construction drone inspections, the platform comprising:

[0064] The image acquisition module is used to collect real-time image data of the construction target in the photovoltaic construction area using a drone, and to obtain the real-time location information and real-time environmental parameter information of the construction target.

[0065] The historical data indexing module is used to retrieve a set of historical inspection image data within a user's historical time period based on the identification information of the construction target.

[0066] The feature classification module is used to classify the image feature information in the historical inspection image data set, generate multiple classification feature categories, and extract the historical defect information set and historical analysis data set under the multiple classification feature categories to obtain multiple historical analysis score sets.

[0067] The analysis score optimization processing module is used to optimize the multiple historical analysis score sets respectively to obtain multiple optimized analysis score sets, and arrange them according to the time series to obtain multiple historical analysis score sequences.

[0068] The trend analysis module is used to perform trend analysis operations based on the multiple historical analysis score sequences, and calculate the rate of change parameter and stability parameter.

[0069] The correction module is used to match the real-time environmental parameter information with the multiple classification feature categories to obtain the matching feature category, and to perform correction calculations on the rate of change parameter and stability parameter based on the deviation between the real-time environmental parameter information and the standard environmental parameter information of the matching feature category to obtain the corrected rate of change parameter and corrected stability parameter.

[0070] The decision output module is used to perform analysis and decision operations based on the corrected rate of change parameter and the corrected stability parameter, generate an analysis report scheme, and output an analysis report.

[0071] The above modules are connected via a data bus to ensure data flow and processing continuity; the input source for image data is the UAV sensor array, and the output report is transmitted to the user terminal through a communication interface.

[0072] Example 1: See Figure 2 The image acquisition module captures real-time image data of the construction target using a multispectral imaging device mounted on the UAV. This imaging device includes visible light and infrared channels, acquiring full-view images of the photovoltaic module installation status, cable connection points, and support structures at a rate of 25 frames per second. The positioning device integrates RTK-GPS and an inertial measurement unit to record the real-time location information of the construction target. The location coordinates include three-dimensional geographic coordinates (longitude, latitude, and elevation), attitude angles (pitch, yaw, and roll), and timestamp data. The environmental sensor group includes a photometer, temperature and humidity probes, and a barometer. The collected real-time environmental parameter information is packaged in a standard format: light intensity is continuously recorded in lumens per square meter, temperature data accuracy reaches ±0.5℃, and humidity data is expressed as percentage relative humidity and correlated with the acquisition time. All sensor data is transmitted to the onboard processor via the UAV's CAN bus. The processor executes a data time synchronization algorithm, aligns the image frames, location coordinates, and environmental parameter timestamps, packages them into structured data frames, and transmits them to the platform server via a 4G / 5G link. The hardware layer uses a gimbal stabilizer to stabilize the image sensor, reducing image blur caused by flight vibration; the environmental sensor group is equipped with a windproof and dustproof cover to reduce the interference of external factors on sampling accuracy.

[0073] After receiving the historical inspection image dataset, the feature classification module first performs multi-scale feature extraction: using the SIFT operator to detect key points in the image, quantizing texture information using the SURF descriptor, and extracting edge gradient distribution using the HOG descriptor. For photovoltaic panel crack feature extraction, an adaptive threshold segmentation algorithm is used to divide the crack pixel region, calculating crack length, number of branches, and direction angle. For stain feature extraction, HSV color space analysis is used to identify abnormal regions exceeding the set color gamut threshold and calculate their area proportion. For connector loosening features, template matching is used to locate bolt positions, and the offset distance between the reference template and the real-time image is compared to output a loosening index. Subsequently, a DBSCAN-based clustering operation is performed, setting the neighborhood radius parameter to 5 pixels and dynamically adjusting the minimum sample size to 0.1% of the total data volume, generating three classification feature categories: photovoltaic panel surface defects, support structure deformation, and electrical interface anomalies. Each category is associated with a historical defect information set: the crack feature category stores the average, maximum, and growth trend statistics of crack lengths detected in the past six months; the stain category archives the historical proportion matrix of dirt area in each region; and the connector category retains the time-series record of torque offset distance.

[0074] The user interaction database retrieves historical detection result information sets for three feature categories, including the coordinates of the defect locations marked by the user, the severity level (divided into minor / moderate / severe), and maintenance action records. The historical submission time set is extracted from the operation log, containing the start time and completion timestamp of each detection. The difference calculation adopts a two-way comparison mechanism: the IoU overlap is calculated between the crack locations marked by the user and the crack areas identified by the algorithm; the severity level of the taint as rated by the user is mapped and calibrated with the taint area interval statistically determined by the algorithm. When outputting the basic analysis score set, the score for the crack category is set according to the overlap deviation (overlap > 90% gets 100 points, 70-90% gets 80 points, < 70% gets 60 points); the taint category score is calculated based on the correlation coefficient between the area ratio and the user's level (coefficient > 0.9 gets 90 points, 0.7-0.9 gets 75 points); the connector category score is calculated based on the offset distance and the mean square error of the user's judgment result. The preset time threshold parameter is a 15-day cycle. The adjustment calculation involves two steps: First, a time decay factor is established, setting the difference between the submission time and the current time ΔT, and calculating the decay factor η=1-ln(ΔT+1) / ln(16); then, the factor is applied to adjust the base score. If ΔT≤15 days, the adjusted score = base score × (1+η / 2); if ΔT>15 days, the score = base score × (1-η / 2). The final three sets of historical analysis scores are stored in a distributed database according to the feature category number.

[0075] In terms of data stream processing, real-time images from the image acquisition module are compressed via edge computing nodes and H.265 encoding is used to reduce bandwidth consumption. The clustering and classification operations of the feature classification module are deployed on a GPU cluster, using CUDA parallelism to accelerate feature vector distance calculations. An anomaly detection mechanism is implemented in the generation process of the historical analysis score set: when the user-submitted timestamp deviates from the sensor's recorded time by more than 300 seconds, a data verification process is triggered to recalibrate the time series. In environmental parameter information processing, temperature data uses a moving average filter to eliminate the impact of instantaneous fluctuations, and light intensity data is fused from redundant sampling values ​​from multiple sensors to improve reliability. Data exchange between modules uses the Apache Avro serialization protocol; image data is transmitted as a binary stream, feature vectors are encapsulated in a Float array format, and the analysis score set is stored in a JSON structure. The platform service layer establishes an independent data sandbox for each construction target to prevent data cross-interference during multi-task processing. The hardware infrastructure adopts containerized deployment; the feature classification module runs within a Docker container, accelerating image feature transmission through shared memory.

[0076] Example 2: See Figure 3When the analysis score optimization processing module starts, it loads multiple historical analysis score sets generated by the feature classification module from the distributed database. Each set corresponds to defect analysis records for a specific construction stage or component type. During module initialization, a data selection strategy is set: it iterates through the median distribution of all score sets and selects the set with the narrowest fluctuation range as the current analysis score set. The benchmark analysis score calculation employs robust processing logic, first removing the highest and lowest 10% extreme values ​​from the current set, and then setting the arithmetic mean of the remaining data as the benchmark value.

[0077] The weight allocation process performs a discreteness mapping transformation, calculating the absolute difference between each analysis score in the current set and the benchmark value. A difference quantification table is pre-set in the module's memory: if the absolute difference is less than 5% of the benchmark value, the weight coefficient is set to 1.0; a difference between 5% and 10% of the benchmark value is assigned a coefficient of 0.8; a difference between 10% and 15% corresponds to a coefficient of 0.6; a difference between 15% and 20% is assigned a coefficient of 0.4; and a difference exceeding 20% ​​reduces the weight to 0.1. The weight vector construction follows a linear decreasing principle, and all coefficients are normalized to form a basic weight distribution matrix with dimensions consistent with the capacity of the current analysis score set.

[0078] The optimization and compression process initiates the Monte Carlo iteration flow, with an initial screening ratio of one-fifth of the total sample size. A random number generator produces an initial index sequence, extracting analysis scores from corresponding positions to form an initial optimized score set. Within this set, the absolute difference between each score and the baseline value is recalculated, and an initial weight distribution is generated according to the same rules. The similarity measure employs the vector angle principle, converting the basic weight distribution and the initial weight distribution into unit vectors. The dot product result is defined as the optimization fitness value. This value is stored in the iteration buffer, rounded to four decimal places.

[0079] Multiple rounds of optimization and screening are performed consecutively, with each round generating a new random index sequence to construct a candidate optimization score set. After the weight distribution of the candidate set is reconstructed, its vector similarity is repeatedly compared with the basic weight distribution. Parallel processing units monitor the trend of optimization fitness values, and convergence is determined when the change in fitness value is less than one-thousandth for two consecutive optimizations. Finally, the candidate set with the highest fitness value during its existence is selected as the optimization analysis score set, and the elimination process compares the results of eight candidate sets cumulatively.

[0080] The time series generation stage employs a timestamp processing engine to parse the collection time label associated with each score from the optimized analysis score set. Timestamps are stored in UNIX millisecond timestamp format, which is then converted to the standard ISO8601 date format by a converter. The sorting algorithm uses monotonically increasing logic to process all dates, arranging data from the same year, month, and day in millisecond-level time precision. The generated historical analysis score sequence is stored in a linked list structure, with nodes containing date and time strings and corresponding analysis score values. The module includes an internal self-verification mechanism to check the continuity and logical order of timestamps within the sequence; if a time reversal is detected, a re-sorting process is triggered.

[0081] The module hardware implementation is based on a heterogeneous computing architecture: the control unit uses a dual-core ARM processor to schedule the overall process, weight calculation and vector operations are deployed on an NVIDIA Jetson GPU unit, and the iterative process of optimization compression is accelerated by FPGA logic circuits. The data channel includes two types of buses: a control bus for transmitting module instruction status and a high-speed data bus for transmitting the floating-point matrix of historical analysis score sets. A threshold is set for the processing cycle, with optimization compression of a single set limited to 300 milliseconds. The storage interface is compatible with SATA solid-state drives for caching intermediate data during the iteration process. An exception handling unit monitors the running status and automatically switches to redundant data processing mode to continue execution when it detects anomalies in the score set data or missing timestamps.

[0082] The output layer encapsulates historical analysis score sequences into a specific data structure, with each sequence appended with metadata descriptions including the corresponding classification feature category number, construction target area code, and sequence generation time. Data packets are transmitted to the input queue of the trend analysis module via fiber optic channel, using an industrial-grade time-sensitive network standard to ensure timing accuracy. A resource reclamation mechanism automatically releases temporary storage space after successful sequence transmission, freeing up computing resources for processing the next score set. Module initialization parameters support dynamic configuration, allowing adjustment of baseline value calculation rules and weight allocation gradients based on different photovoltaic scenarios.

[0083] Example 3: See Figure 4 and Figure 5The trend analysis module begins with the sample data preprocessing stage, loading sample inspection data sets from multiple users from a distributed storage system. Each sample contains analysis score records for no fewer than 30 consecutive time points. The data cleaning stage removes outlier samples, using a sliding window method to detect data points deviating from the mean by more than three standard deviations, and replacing them with linear interpolation of adjacent time points. The sample analysis score sequence set is grouped by photovoltaic module type, and each group of sequences undergoes normalization, mapping the original scores to the [0,1] interval. The rate of change feature extraction uses a piecewise linear fitting method, dividing each sample sequence into five equal-length time periods and calculating the slope of each segment as the local rate of change. Stability features are characterized by the moving average coefficient of variation, with the window width set to 20% of the total sequence length, calculating the ratio of the standard deviation to the mean of the data within the window.

[0084] The trend analysis model employs a two-layer LSTM neural network structure. The input layer receives fractional sequence segments of length 10, the hidden layer has 64 memory units, and the output layer contains two independent fully connected branches, corresponding to the prediction of the rate of change parameter and the stability parameter, respectively. During model training, an adaptive moment estimation optimizer is used, with an initial learning rate of 0.001 and a batch size of 32 samples. To prevent overfitting, a random deactivation layer with a dropout rate of 0.2 is added after the hidden layers. The loss function is designed as a composite function, using Huber loss for the rate of change branch and log-cosine loss for the stability branch. The early stopping metric on the validation set is continuously monitored during training; training is terminated when the validation loss fails to decrease for five consecutive epochs.

[0085] In the real-time analysis phase, the module receives historical analysis score sequences from the analysis score optimization processing module. Each sequence first undergoes the same normalization process as the training data. The model inference engine loads the pre-trained weight file and performs forward propagation calculations on the input sequences. The rate of change parameter of the output layer is converted to its original dimension, multiplied by 100, and expressed as a percentage rate of change; the stability parameter retains a dimensionless value within the interval [0,1]. In the similarity calculation stage, the matching of real-time environmental parameter information and classification feature categories adopts a multi-dimensional spatial distance metric:

[0086]

[0087] in, Represents the similarity value. The first representing real-time environmental parameters Dimensions (light intensity, temperature, humidity). The first environmental parameter for classifying characteristic categories One dimension, These are the weighting coefficients for each dimension (light intensity 0.5, temperature 0.3, humidity 0.2). The calculation results generate a similarity vector with the same number of dimensions as the number of categories in the classification feature.

[0088] In the parameter fusion stage, multiple feature change rate parameters and feature stability parameters output by the model are weighted and aggregated. The weight allocation is determined based on the ranking of similarity values: the parameter corresponding to the highest similarity category is assigned a weight of 0.6, the second highest similarity category a weight of 0.3, and the remaining categories share the remaining weight of 0.1. The weighted summed change rate parameters and stability parameters are then calibrated for range. The change rate parameter is limited to the range of [-50%, 50%], and boundary values ​​are taken when it exceeds the threshold; the stability parameter is adjusted to the range of 0.2-0.8 using the Sigmoid function.

[0089] The correction module's operation process involves identifying the index corresponding to the maximum value in the similarity vector and setting the classification feature category pointed to by that index as the matching feature category. Standard environmental parameter information is extracted from the feature category's metadata, including baseline values ​​for light intensity (klx), temperature (°C), and humidity (%RH). Deviation is calculated using the relative difference method, taking the absolute difference between the real-time value and the baseline value for each of the three environmental parameters, and then dividing by the baseline value to obtain the normalized deviation. Adjustment coefficients are generated using a piecewise function strategy: the coefficient is 1 when the total deviation is less than 0.1; 0.9 when the deviation is between 0.1 and 0.3; 0.7 for deviations between 0.3 and 0.5; and 0.5 when the deviation exceeds 0.5.

[0090] The correction calculation process performs differentiated processing on the rate of change parameter and the stability parameter. The correction of the rate of change parameter uses a multiplicative model, multiplying the original parameter by an adjustment coefficient and then adding a linear compensation term for the deviation. The correction of the stability parameter introduces a nonlinear transformation, using the adjustment coefficient as the base of the exponent term and the original parameter as the power in the calculation. The corrected parameter values ​​retain three decimal places of precision, and the data verification module verifies the rationality of the values; values ​​exceeding the theoretical range trigger a recalculation process.

[0091] In terms of hardware implementation, the module is deployed on a dedicated inference server, equipped with an NVIDIA T4 GPU to accelerate model computation. The environmental parameter processor adopts a low-power ARM architecture and independently processes sensor data streams. Memory management employs a double-buffering mechanism to ensure continuous data supply during real-time analysis. An exception handling unit monitors the entire process and automatically switches to a standby analysis mode when sensor data interruption or model inference timeout is detected. The communication interface uses a time-triggered Ethernet protocol to ensure the timing determinism of correction parameter transmission.

[0092] The data persistence layer records a complete correction log, including original parameter values, environmental deviations, adjustment coefficients, and correction results. Log entries are appended with timestamps and operator identifiers to support subsequent audit trails. The intermediate data visualization function allows viewing similarity calculation details and weight allocation processes in debug mode, aiding in algorithm optimization and verification. Module configuration parameters support hot updates, and deviation thresholds and correction strategies can be dynamically adjusted through the management interface.

[0093] The output interface encapsulates the corrected rate of change and corrected stability parameters into structured messages, appending data quality identifiers and confidence scores. The message queue producer pushes data to the subscribed topics of the decision output module and simultaneously writes it to the time-series database for long-term trend analysis. The resource reclamation thread cleans up temporary data and resets the model inference state after each analysis, preparing for the next processing task.

[0094] Example 4: The operation of the decision output module begins with parameter reception. The correction rate of change parameter records the trend of construction defects as a percentage (positive values ​​indicate deterioration, negative values ​​indicate improvement), and the correction stability parameter describes the degree of fluctuation with a value between 0 and 1 (the larger the value, the more stable). The module's built-in decision model is trained and generated based on the support vector machine algorithm. The training dataset contains various combinations of correction parameters for different scenarios and corresponding manually labeled decision schemes. The following table shows the mapping relationship between five typical input parameters and output schemes:

[0095] Table 1: Mapping Relationship between Five Typical Input Parameters and Output Schemes

[0096]

[0097] During model training, 200 sets of sample parameters from historical maintenance records are collected as input feature vectors, and the output labels correspond to five reporting schemes. The support vector machine kernel function adopts the radial basis function, the penalty coefficient is set to 1.0, and the decision boundary is optimized through five-fold cross-validation. During real-time decision-making, the input vector [corrected rate of change parameter, corrected stability parameter] is preprocessed by feature scaling and then input into the model. The probability distribution of the output layer triggers the selection of the corresponding scheme. When the probability difference between two classes is less than 0.2, the multi-scheme fusion mechanism is activated to generate a combined report.

[0098] After receiving the report scheme from the decision-making output, the visualization module executes the data-to-graphics conversion logic. For red warning reports in the table, the module generates a heat map overlaid on the 3D model of the construction area: crack propagation areas are marked with flashing red dots, and a sidebar displays a line chart comparing historical data. Regular reports are converted into interactive dashboards, where users can click on the cable trough number to expand the loosening trend over twelve months. Email text reports embed dynamic bar charts showing the ratio of the corrosion area of ​​the support to the standard threshold. The 3D positioning map uses layered rendering technology, with electrical short circuit points displayed with pulsed light effects, and the related equipment list displays insulation resistance parameters in a floating display. The visualization dashboard is integrated into the web interface, dividing the drilling views into multiple levels according to the construction area, with green checkmarks indicating areas that have passed the stain cleaning standard.

[0099] The hardware support environment consists of a dual-socket server cluster. The decision-making model runs on compute nodes equipped with 128GB of memory, and visualization rendering tasks are distributed across four GPU accelerator cards. When processing 3D localization maps, the OpenGL pipeline generates a photovoltaic power station model containing over 20,000 triangular faces in real time, and texture maps are loaded from UAV orthophotos. The network transport layer employs a layered compression strategy: structured report data is transmitted in Protobuf format with an average size of 8KB; heatmap data streams are compressed to 30% of their original size using Delta encoding; and the 3D model uses a LOD hierarchical loading mechanism, initially transmitting only a simplified topology.

[0100] The message routing system selects the delivery channel based on the report type: red alert reports trigger push notifications simultaneously via SMS gateway, email server, and mobile app; regular reports are written to a relational database and then retrieved via polling by the platform's message center. The security mechanism employs two-factor authentication; when a shutdown / maintenance command is sent, the engineering supervisor's biometrics are required for secondary verification. Data archiving establishes independent storage volumes based on photovoltaic project numbers, and all historical reports are associated with corresponding environmental parameter snapshots at specific times, supporting the reconstruction of decision-making scenarios during retrospective analysis.

[0101] The user interface design incorporates multi-level control functions: in the visual dashboard, hovering the mouse over a compliant stain area brings up a cleaning record timeline; in the 3D model view, a slider controls the playback of the defect evolution process over time; and the report export menu provides a complete evidence chain download from raw inspection data to decision reports. All graphic elements comply with the WCAG 2.0 accessibility standard, and colorblind mode automatically converts red-green warning signs to zebra-striped patterns. The real-time monitoring dashboard refresh rate is set to 5 seconds per refresh, triggering a pop-up alert when the correction change rate parameter exceeds 15%.

[0102] Version control mechanisms ensure the traceability of decision rules: after each model update, the new decision logic is imported into a 10% end-user environment for testing via a canary release; historical decision records are stored alongside current rules, allowing for comparison of decision differences for the same parameters across different versions in audit mode. The failover solution is designed with a three-tier degradation strategy: automatic switching to a backup model service when the primary server is interrupted; generation of basic reports using offline decision tables in the event of a system-wide failure; and cached data on a local SSD in network unavailability scenarios, with batch retransmission upon network reconnection.

[0103] Example 5: The environmental interference filtering module continuously receives real-time environmental parameter data streams transmitted from the image acquisition module. The input channel samples light intensity, temperature, and humidity data at a frequency of 100Hz. Light intensity abrupt change detection employs a bidirectional threshold judgment mechanism: when the light intensity change exceeds a baseline value by more than 20% within three consecutive sampling periods, and fails to return to the baseline range in the subsequent two periods, an abnormal light intensity marker is triggered. Dust concentration monitoring relies on a scattering optical sensor; a dust marker is activated when the standard deviation of the detected increase in suspended particle count exceeds three times the historical average. Humidity surge judgment is based on a humidity gradient algorithm; a humidity anomaly is recorded if the change rate exceeds 10% per minute. The marker data structure includes interference type encoding, start timestamp, duration, and spatial location coordinates, derived from the location information associated with the image acquisition module. The generated environmental interference marker data is embedded in the message header and transmitted to the analysis and compensation module.

[0104] The analysis and compensation module polls the marker message queue in real time, and immediately loads the original image of the corresponding time window when an interference marker is detected. The region localization engine delineates rectangular detection boxes in the UAV orthophoto coordinate system based on the coordinate parameters in the markers. Contaminated area identification employs adaptive mask generation technology: for scenes with sudden changes in illumination, it segments pixel clusters in the image that exceed the ambient brightness mean by ±30%; for dust interference, it extracts hazy areas with color saturation below 25%; for humidity interference, it locates abnormal blocks whose hue shifts to the blue spectrum. Segmentation boundaries are stored as vector polygons, with vertex coordinates mapped to the construction target entity. The compensation coefficient calculation module calls the correction stability parameters from the correction module, scaling them to the input variables of an exponential function. Image enhancement processing uses a channel-specific operation: histogram equalization is performed on the red channel of the contaminated area to compensate for color distortion; bilateral filtering is applied to the blue channel to suppress water mist noise; and a nonlocal mean algorithm is used to restore detail texture in the green channel. The processed local image undergoes feathering edge processing and is then reconstructed into the original image using alpha blending technology to output the compensated image.

[0105] The dynamic weight allocation module initiates the defect comparison process, loading the compensated image feature data and the historical defect information set stored by the feature classification module. The feature extractor separates the structural crack morphology vector, the electrical component location topology map, and the surface stain distribution matrix from the compensated image. The comparison unit performs three-stage matching: the structural crack features are compared with the largest crack morphology in the historical database using Hausdorff distance calculation; the electrical component location matching is based on topological similarity to evaluate the offset; and the stain distribution is calculated using matrix convolution to calculate the spatial correlation coefficient. The construction stage identifier parses the UAV inspection task code: pile foundation construction stage codes 01-04, support installation stage codes 05-08, and component debugging stage codes 09-12. The preset defect weight strategy for each stage is: 0.8 for structural deformation in the pile foundation stage and 0.7 for electrical defects in the component stage. The dynamic allocation algorithm reads the stage code matching strategy table, loads the baseline weight according to the defect type, and then superimposes the influence coefficient of the comparison result. Crack defect weights are adjusted according to the Hausdorff distance reduction ratio, increasing by 0.15 when the distance is less than 20% of the historical value; electrical defect weights are adjusted according to the topology offset using a gradient; stain weights are multiplied by the square root of the correlation coefficient. The weighted defect analysis results are packaged into a structured dataset, with fields including the defect type code, original score, dynamic weight value, and final weighted score for each construction point. Data messages with timestamps are transmitted to the pre-analysis interface of the decision output module.

[0106] The system implementation employs multi-level redundancy control. A hot standby node for the environmental interference filtering module takes over the data stream when the host response latency exceeds 50ms. Image processing tasks for the analysis and compensation module are distributed to a heterogeneous computing cluster: CPUs handle region segmentation tasks, GPUs perform channel enhancement operations, and FPGAs accelerate hybrid recombination operations. The dynamic weight allocation module uses a distributed in-memory database to store stage strategy tables, supporting thousands of concurrent read requests. Communication transmission uses a fiber optic ring topology, establishing point-to-point encrypted data channels between modules, and the transmission protocol adheres to industrial IoT security standards. The anomaly handling unit monitors the entire workflow; when interference marker loss or compensation calculation timeout is detected, it activates the default compensation mode based on historical averages.

[0107] The persistent storage system records the entire operation chain: from the original sampled values ​​of environmental interference to the final weighted defect score, the complete data chain is stored in partitions according to the construction area. A visual debugging interface allows maintenance personnel to view the segmentation effect of the interference area, image comparison before and after compensation, and the weight allocation calculation tree. System maintenance supports online policy updates; after changes to the engineering construction plan, a new stage weight comparison table can be imported through the configuration interface, and change records are generated with independent version numbers and archived in the engineering knowledge base. Execution efficiency optimization measures include a preprocessing caching mechanism, preloading frequently occurring interference area templates within the last three months to accelerate the localization process. The lifecycle management module automatically archives historical policy datasets that have not been accessed for more than three years, retaining policies compressed into minimal feature snapshots for audit traceability.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent image analysis platform for photovoltaic construction drone inspections, characterized in that, The platform includes: The image acquisition module is used to collect real-time image data of the construction target in the photovoltaic construction area using a drone, and to obtain the real-time location information and real-time environmental parameter information of the construction target. The historical data indexing module is used to retrieve a set of historical inspection image data within a user's historical time period based on the identification information of the construction target. The feature classification module is used to classify the image feature information in the historical inspection image data set, generate multiple classification feature categories, and extract the historical defect information set and historical analysis data set under the multiple classification feature categories to obtain multiple historical analysis score sets. The analysis score optimization processing module is used to optimize the multiple historical analysis score sets respectively to obtain multiple optimized analysis score sets, and arrange them according to the time series to obtain multiple historical analysis score sequences; The trend analysis module is used to perform trend analysis operations based on the multiple historical analysis score sequences, and calculate the rate of change parameter and stability parameter; The correction module is used to match the real-time environmental parameter information with the multiple classification feature categories to obtain the matching feature category, and to perform correction calculations on the rate of change parameter and stability parameter based on the deviation between the real-time environmental parameter information and the standard environmental parameter information of the matching feature category to obtain the corrected rate of change parameter and corrected stability parameter. The decision output module is used to perform analysis and decision operations based on the corrected rate of change parameter and the corrected stability parameter, generate an analysis report scheme, and output the analysis report; The analysis score optimization processing module optimizes each of the multiple historical analysis score sets to obtain multiple optimized analysis score sets, and arranges them according to time series to obtain multiple historical analysis score sequences, including: Select one of the multiple historical analysis score sets as the current analysis score set, and determine the benchmark analysis score in the current analysis score set; Based on the distance between other analysis scores in the current analysis score set and the benchmark analysis score, weight coefficients are assigned to generate a basic weight distribution, wherein the distance is inversely proportional to the weight coefficient. Based on the aforementioned basic weight distribution, an optimization and compression operation is performed on the current set of analysis scores to obtain an optimized set of analysis scores. The historical analysis score sequence is generated by sorting multiple analysis scores within the optimized analysis score set according to their timestamp information. Repeat the optimization compression and sorting operations on other historical analysis score sets to generate multiple historical analysis score sequences.

2. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 1, characterized in that, The image acquisition module uses a drone to collect real-time image data of the construction target in the photovoltaic construction area, and obtains the real-time location information and real-time environmental parameter information of the construction target, including: The drone's image sensor captures real-time image data of the construction target, while the drone's positioning device records the real-time location information of the construction target. It also integrates environmental sensors to collect real-time environmental parameter information of the construction target, including light intensity data, temperature data, and humidity data.

3. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 1, characterized in that, The feature classification module classifies the image feature information within the historical inspection image dataset, generating multiple classification feature categories, and extracts historical defect information sets and historical analysis data sets under the multiple classification feature categories, processing them to obtain multiple historical analysis score sets, including: Multiple image feature information is extracted from the historical inspection image data set, and clustering and classification operations are performed to generate multiple classification feature categories; Extract the historical defect information set under the multiple classification feature categories, and obtain the historical detection result information set and historical submission time set submitted by the user under the multiple classification feature categories; Based on the degree of difference between the historical detection result information set and the historical defect information set, multiple basic analysis score sets are generated. By utilizing the ratio between a preset time threshold parameter and the historical submission time set, adjustment calculations are performed on the multiple basic analysis score sets to generate multiple historical analysis score sets.

4. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 1, characterized in that, The analysis score optimization processing module performs an optimization and compression operation on the current analysis score set based on the basic weight distribution to obtain an optimized analysis score set, including: A preset number of analysis scores are randomly selected from the current set of analysis scores to form an initial set of optimization scores. Based on the distance between the analysis scores within the initial optimized score set and the benchmark analysis scores, weight coefficients are assigned to generate an initial weight distribution; Calculate the similarity metric between the initial weight distribution and the basic weight distribution, and use it as the optimization fitness value; A preset number of analysis scores are randomly selected from the current set of analysis scores to form a set of candidate optimization scores, and the corresponding candidate fitness values ​​are calculated. Continue performing optimization compression operations until the fitness values ​​converge, and output the candidate optimization score set with the largest optimization fitness value as the optimization analysis score set.

5. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 1, characterized in that, The trend analysis module performs trend analysis operations based on the multiple historical analysis score sequences, calculating the rate of change parameter and stability parameter, including: Collect sample inspection data sets from multiple users, obtain sample analysis score sequence sets, and calculate sample change rate set and sample stability set based on the analysis score change characteristics of each sample analysis score sequence; A trend analysis model is constructed using the set of sample analysis score sequences as input data and the set of sample change rates and the set of sample stability as output data. The trend analysis model is applied to process the multiple historical analysis score sequences to obtain multiple characteristic change rate parameters and multiple characteristic stability parameters. The similarity values ​​between the real-time environmental parameter information and the multiple classification feature categories are analyzed. Based on the magnitude of the multiple similarity values, a weighted summation calculation is performed on the multiple feature change rate parameters and multiple feature stability parameters to generate change rate parameters and stability parameters.

6. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 5, characterized in that, The correction module matches the real-time environmental parameter information with the multiple classification feature categories to obtain matching feature categories, and calculates corrections for the rate of change parameter and stability parameter based on the deviation between the real-time environmental parameter information and the standard environmental parameter information of the matching feature categories, to obtain corrected rate of change parameters and corrected stability parameters, including: Select the category with the highest similarity value as the matching feature category, and obtain the standard environmental parameter information of the matching feature category; An adjustment coefficient is set based on the deviation between the real-time environmental parameter information and the standard environmental parameter information of the matching feature category; The adjustment coefficient is applied to perform a correction calculation on the rate of change parameter and the stability parameter to generate a corrected rate of change parameter and a corrected stability parameter.

7. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 1, characterized in that, The decision output module, based on the corrected rate of change parameter and the corrected stability parameter, performs analytical decision operations, generates an analytical report scheme, and outputs an analytical report, including: Obtain the set of sample corrected rate of change parameters and the set of sample corrected stability parameters. Based on the magnitude of each sample corrected rate of change parameter and sample corrected stability parameter, set a sample reporting scheme and generate a set of sample reporting schemes. The sample reporting scheme includes a prompt level parameter. The magnitude of the sample corrected rate of change parameter and the sample corrected stability parameter is inversely proportional to the magnitude of the prompt level parameter. A decision model is constructed using the set of sample corrected rate of change parameters and the set of sample corrected stability parameters as decision input data, and the set of sample reporting schemes as decision output data. The decision model is applied to process the corrected rate of change parameter and the corrected stability parameter to generate an analysis report scheme.

8. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 1, characterized in that, The platform also includes a visualization module, which is used to receive the analysis report scheme generated by the decision output module, convert the analysis report scheme into visual chart information, and output it to the user interface. The visualization module is connected to the decision output module. The analysis report scheme output by the decision output module serves as the input data for the visualization module. The visualization module processes the analysis report scheme to generate visualization chart information.

9. The intelligent image analysis platform for photovoltaic construction drone inspection according to claim 1, characterized in that, Also includes: An environmental interference filtering module is used to receive real-time environmental parameter information collected by the image acquisition module, identify interference scenarios such as sudden changes in light intensity, abnormal dust concentration, or sudden increase in humidity, and generate environmental interference labeling data. The analysis and compensation module is used to locate the contaminated image region based on the environmental interference marker data, call the correction stability parameter output by the correction module to calculate the compensation coefficient, and perform noise reduction and enhancement processing on the image features of the contaminated region. The dynamic weight allocation module compares the noise-reduced image feature data with the historical defect information set generated by the feature classification module, dynamically allocates analysis weights for different defect types according to the construction stage, and outputs the weighted defect analysis results to the decision output module.

Images