A new energy station safety inspection management method and system

Abstract

The application relates to the technical field of photovoltaic power station intelligent operation and maintenance, and discloses a new energy field station safety inspection management method and system. In view of the problem of high false alarm rate of traditional infrared thermal imaging under complex meteorological conditions, the method obtains multi-source heterogeneous data such as unmanned aerial vehicle double-light images, inverter time sequence electric power and meteorological sequences; the equivalent heat conduction time constant of a calculation component is solved, dynamic hysteresis compensation is performed on the electric heat micro-derivative characteristics to align the physical phase; by solving the first-order partial correlation coefficient, the environmental common-mode interference such as irradiance is stripped, and the intrinsic causal relationship between thermodynamic mutation and electrical attenuation is judged; a residual neural network is combined to extract local topological features to filter physical artifacts; finally, a device topological correlation matrix fusing multi-dimensional degradation features is constructed, and a graded diagnosis result is output. The application effectively eliminates transient thermal noise caused by cloud shadow shielding and meteorological fluctuations, and improves the accuracy and engineering robustness of photovoltaic defect diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for photovoltaic power plants, specifically to a method and system for safety inspection and management of new energy power plants. Background Technology

[0002] With the large-scale development of the photovoltaic power generation industry, intelligent operation and maintenance and equipment defect diagnosis of large-scale photovoltaic power plants have become key links in ensuring power generation efficiency and operational safety. Currently, aerial inspection methods based on drones equipped with infrared thermal imaging and visible light dual-light pods have been widely used for hot spot detection and physical condition assessment of photovoltaic modules.

[0003] In practical engineering applications, existing UAV visual inspection generally adopts a routine scanning strategy that covers the entire field without discrimination. This traversal acquisition method based on fixed routes results in UAVs engaging in ineffective flights and redundant loitering in a large number of fault-free areas. This not only consumes a significant amount of power battery life but also generates a massive amount of image features with no diagnostic value. Sudden changes in operating parameters captured by the underlying electrical system (such as combiner boxes or inverters) fail to establish a real-time mapping and linkage with the inspection scheduling in the upper physical space, resulting in a lack of data-driven and precise targeting in the dispatch of inspection terminals.

[0004] In identifying latent defects in specific photovoltaic modules, conventional diagnostic methods typically rely on static infrared absolute temperature extremes or regional temperature difference thresholds under a single shutter trigger for fault classification. However, the physical encapsulation structure of photovoltaic modules (such as the tempered glass panel and ethylene-vinyl acetate copolymer film) objectively possesses high thermal inertia. Electro-induced Joule heating caused by microcracks in the cells, potential-induced decay, or high-resistance solder joints requires a specific thermal conduction relaxation time to be conducted from the module's interior to the surface and form a macroscopic temperature gradient that can be captured by sensors. Static single-frame image extraction or simple temperature difference comparison completely severs the dynamic physical process of thermodynamic conduction and lacks a mechanism for actively stimulating and continuously observing internal electro-induced heat sources over a time window.

[0005] Due to neglecting the high thermal inertia limitations imposed by the physical structure of components, existing technologies struggle to align the physical phase difference between the underlying electrical signal excitation and the surface thermal signal response on the data processing timeline. When photovoltaic power plants face non-electrical environmental interferences such as bird droppings, uneven dust accumulation, optical reflections, or complex cloud shadow movements, these external factors also exhibit significant thermal anomalies in the infrared field of view. Traditional diagnostic logic cannot establish a mathematical causal model between dynamic temperature changes and the underlying current derivative, making it difficult to separate real internal electrical defects that change with electrical operating conditions from non-electrical static heat sources that do not change with current. This lack of causal verification and multi-source feature decoupling mechanism ultimately leads to the system being prone to misjudging hot spots under complex outdoor weather conditions, severely restricting the engineering robustness of automated defect diagnosis systems and the efficiency of on-site maintenance and troubleshooting. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a safety inspection and management method and system for new energy power stations. It solves the problem that existing photovoltaic defect diagnosis schemes rely solely on static thermodynamic extreme values ​​and lack multidimensional electrothermal causality verification, which can easily lead to large-scale false alarms under complex weather conditions (such as cloud shadows and sudden changes in irradiance) and surface foreign object interference.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a safety inspection and management method for new energy power plants. This method establishes a cross-domain control mechanism between the backend system and the frontend power plant equipment. Firstly, based on the dispersion anomalies of the real-time electrical operating parameters of the photovoltaic array, it generates inspection trigger events with electrical topology identifiers. Then, the site's mapping database is invoked to convert the abnormal pure logical topology node into a three-dimensional physical space coordinate set, and the inspection equipment is dispatched to the target airspace. After the inspection equipment is aligned with the target photovoltaic string, the system actively intervenes in the inverter's operating status, sends a transient bias command to it to forcibly change the output current of the target photovoltaic string, and simultaneously records thermal video streams and continuous current sequence data within the time window of the current change. During the data processing stage, the thermal derivative of the spatial pixel temperature over time and the electrical derivative of the internal heating power over time are calculated respectively. A longitudinal heat conduction time constant for the physical structure of the photovoltaic module is introduced, and the partial correlation coefficient between the two is calculated after phase hysteresis compensation is performed on the electrical derivative. Finally, based on the combination of the partial correlation coefficient and the absolute temperature of the pixel, the real electro-induced high-resistivity hot spots and the pseudo hot spots formed by external environmental interference are separated at the physical property level.

[0008] The core innovation of this invention lies in the following: According to Joule's law, the heat generated by electrically induced high-resistivity defects such as microcracks and loose connections within a photovoltaic module is proportional to the square of the current flowing through it. When the system actively forces a transient change in the output current of the photovoltaic string, the heat generated in the actual electrical defect area will inevitably undergo a synchronous transient change, leading to a change in surface temperature. In contrast, static heat sources generated by external environmental factors such as bird droppings and reflections do not change their thermodynamic state with fluctuations in internal current. This invention extracts the derivatives of temperature and electrical power over time and introduces a thermal conduction time constant to offset the physical time delay of internal heat conduction to the module surface, establishing a mathematical model for the partial correlation of the spatial thermodynamic response with the underlying electrical excitation. This model objectively distinguishes between actual defects with electrothermal coupling characteristics and environmental interference sources without such characteristics.

[0009] Preferably, when performing electrical operation anomaly identification and event triggering, this invention extracts the real-time DC current of all photovoltaic strings under the same maximum power point tracking loop, calculates the relative deviation of the current of each string from the average operating current of the system, constructs a string current deviation matrix based on this deviation data, and continuously calculates the time rate of change of the Frobenius norm of this matrix within a fixed time sliding window. When the time rate of change is detected to exceed a preset trigger threshold, it is identified as an electrical transient anomaly event. This process quantifies multi-dimensional current fluctuations into a single matrix energy change index, providing an objective triggering basis directly driven by electrical characteristics for subsequent inspection actions.

[0010] In one specific embodiment, for the spatial positioning of abnormal sequences, the system internally calls a mapping database that stores the correspondence between the inverter logic ports and the actual three-dimensional geographic coordinate system of the site, directly converting the logical identifiers into a set of three-dimensional spatial polygon coordinates. The inspection and scheduling module reads the current GPS coordinates of the mobile terminal and the site's elevation obstacle map, generates the shortest unobstructed directional cruise trajectory, and sends it out. This step eliminates the need for a full-site blind scan in traditional inspections, allowing the inspection terminal to directly reach the physical location where electrical anomalies exist.

[0011] Preferably, the implementation process of actively intervening in the inverter's operating state is specifically manifested as follows: A transient bias command based on a voltage ramp control model is issued to the inverter, forcing it to interrupt its original maximum power point optimization process. The inverter is controlled to continuously increase or decrease the output current of the target photovoltaic string within a set thermal conduction relaxation time window, according to a preset fixed slope. The total duration of this thermal conduction relaxation time window is set to be greater than the longitudinal thermal conduction time constant from the internal conductive layer to the surface glass layer of the photovoltaic module. The combination of this time window and the ramp load control mechanism provides sufficient physical time for the Joule thermal changes inside the module to be conducted to the surface, overcoming the technical deficiency that conventional millisecond-level current scanning cannot generate an effective temperature gradient at the infrared sensor end due to the large thermal inertia of the photovoltaic glass layer.

[0012] In one specific embodiment, the synchronous acquisition and analysis calculation process of multimodal data includes: within the aforementioned thermal conduction relaxation time window, the infrared sensor captures images at a constant frame rate, and the site monitoring and data acquisition system synchronously records the corresponding DC current data using a network time protocol, ensuring that each frame of the thermal video stream is aligned with the current sampling sequence in terms of timestamps. In the derivative extraction stage, the partial derivatives of the temperature value and the square of the current value at each pixel are calculated with respect to time. During the pairing calculation, the thermal derivative of the current temperature is paired with the thermal derivative of the electrical power at a historical moment shifted forward by one thermal conduction time constant on the time axis to complete phase hysteresis compensation and eliminate data lag errors caused by the physical heat transfer process.

[0013] Preferably, the specific logic of the defect identification and pseudo-hotspot filtering mechanism is as follows: First, the system extracts a set of high-temperature pixels with absolute temperatures higher than the surrounding normal area from the original infrared image of the spatial pixel-level thermal video stream as the test samples. Then, it iterates through this set and evaluates the partial correlation coefficient of each high-temperature pixel. For pixels with a partial correlation coefficient greater than a preset threshold, it is determined that their surface temperature rise transient characteristics and changes in internal heating power are physically causally coupled, confirming them as genuine electrically induced high-resistivity hotspots. For pixels with a partial correlation coefficient less than or equal to the preset threshold, it is determined that their temperature rise does not change with current, confirming them as non-electrically induced environmental interference pseudo-hotspots and removing them from the system record. Finally, the system constructs a device topology correlation matrix based on the coordinates of the actual defective pixels and derives a device defect diagnosis report without environmental interference information.

[0014] A second aspect of this invention provides a safety inspection and management system for new energy power stations, the physical and functional architecture of which includes: The event-driven triggering module is used to collect real-time electrical operating parameters of the photovoltaic array in the power station through the data interface, perform calculations of current dispersion and matrix norm change rate, and output an inspection triggering event carrying an abnormal string logic identifier when an abnormal parameter limit is detected. The reverse parsing scheduling module is used to call the built-in geographic information mapping database, translate the above logical identifiers into the corresponding physical space coordinate set, and plan the route based on the positioning data to schedule the inspection mobile terminal. The active bias and acquisition module is used to send transient bias commands to the specified inverter through the station communication network to interfere with its maximum power point tracking algorithm, control the target photovoltaic string to perform long-cycle time window load changes, and control the infrared camera and current acquisition device to synchronously acquire thermal image video stream and actual continuous current sequence under the same timestamp protocol. The phase compensation calculation module is used to perform partial derivative mathematical operations to extract the thermal derivative of temperature over time and the time derivative of electrical power, respectively. It performs misalignment matching compensation on the time axis through the set heat conduction time constant and calculates the Pearson partial correlation coefficient between the two. The defect identification and judgment module is used to execute comparison logic, combine the absolute temperature characteristics of infrared pixels and calculate the partial correlation coefficient of the output, perform threshold screening operation, filter out false hot spots caused by static environmental heat sources, and output the physical spatial coordinates of electrically induced high-resistivity hot spots.

[0015] This invention provides a method and system for safety inspection and management of new energy power stations. It has the following beneficial effects: 1. This invention generates inspection trigger events by calculating the time rate of change of the Frobenius norm of the photovoltaic string current deviation matrix, and uses an electrical and physical topology mapping database to parse logical identifiers into three-dimensional spatial coordinates, directly dispatching the inspection mobile terminal to the target area. This technical feature directly drives spatial positioning through the physical quantification of the underlying electrical operating parameters, replacing the indiscriminate video acquisition process covering the entire field, and reducing the drone flight time and ineffective dwell time of the inspection terminal in fault-free areas.

[0016] 2. This invention controls the output current of the target photovoltaic string to continuously and monotonically change at a preset slope within a thermal conduction relaxation time window that is greater than the longitudinal thermal conduction time constant of the photovoltaic module by issuing transient bias commands to the inverter. This technical feature provides the physical continuity required to satisfy the thermodynamic conduction equation, enabling the change in Joule heat caused by internal electrogenic defects to be conducted to the module surface and generate a causally related macroscopic temperature gradient at the infrared sensor end, overcoming the technical limitation of lag in surface temperature change caused by the high thermal inertia of the photovoltaic module's physical structure.

[0017] 3. This invention extracts the spatial pixel thermal quotient and the internal heating electrical power quotient, and introduces a heat conduction time constant to compensate for the phase hysteresis of the electrical power quotient, thereby calculating the partial correlation coefficient between the two to perform defect identification. This technical feature aligns the excitation of the underlying electrical signal and the response of the surface thermal signal on the data processing time axis, establishes a mathematical calculation model between temperature change and current change, and isolates the actual electro-induced high-resistivity hot spots that change with current from static heat sources that do not change with current and are not affected by electrical environmental interference, thus eliminating misjudgments of hot spots caused by external factors such as bird droppings or surface reflection. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the overall application scenario and physical architecture of an embodiment of the present invention; Figure 2 This is a flowchart illustrating the overall method of an embodiment of the present invention; Figure 3 This is a comparison chart of the performance evaluation of various diagnostic schemes under complex weather conditions according to the present invention; Figure 4 The present invention provides subject operating characteristic (ROC) curves for different protocols. Detailed Implementation

[0019] 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.

[0020] See attached document Figure 1 A photovoltaic array comprises multiple independent photovoltaic strings. These strings are electrically connected in parallel to the same maximum power point tracking (MPPT) loop of the string inverter. The MPPT loop is commonly referred to as such in the art. The DC input of the string inverter is equipped with a current sensor. This current sensor, typically a Hall effect current sensor or a shunt, is used to independently sample the branch current of each connected photovoltaic string to obtain real-time electrical operating parameters. These real-time electrical operating parameters are specifically represented in actual physical measurements as the real-time DC current values ​​of each photovoltaic string.

[0021] To achieve upward aggregation of the underlying physical state, this embodiment is equipped with a monitoring and data acquisition system at the data communication and aggregation layer. String inverters are connected to data acquisition units within the photovoltaic subarray area via RS485 communication cables or power line carrier communication technology. The data acquisition units continuously transmit the aforementioned real-time DC current values ​​to the monitoring and data acquisition system through a fiber optic Ethernet ring network laid within the site. The monitoring and data acquisition system is equipped with a time-series database responsible for continuously recording and persistently storing the underlying electrical operating status of the entire site.

[0022] As a specific implementation of a lower-level technical feature, the inspection mobile terminal employs a rotary-wing drone with vertical takeoff and landing (VTOL) and hovering capabilities. An infrared sensor is mounted on the multi-axis stabilized gimbal under the drone's fuselage. Specifically, the infrared sensor is an uncooled vanadium oxide microbolometer, used to perform two-dimensional scanning and imaging of the target photovoltaic module surface from the air, acquiring a spatial pixel-level thermal image video stream containing both spatial pixel planar coordinates and the temporal dimension. The inspection mobile terminal also integrates a Global Positioning System (GPS) receiver module supporting real-time dynamic differential positioning technology, used to output real-time three-dimensional spatial position coordinates with centimeter-level accuracy.

[0023] For the basic flight attitude calculation and motor drive control of rotary-wing UAVs, those skilled in the art can use existing open-source flight control algorithms or commercial flight control systems. The specific aerodynamic balance principles and basic navigation logic are well-known technologies in this field and will not be elaborated here.

[0024] The inspection management server establishes a communication connection with the monitoring and data acquisition system via fiber optic Ethernet, continuously retrieving electrical data from the photovoltaic array to perform anomaly detection calculations. When the inspection management server detects an electrical operational anomaly, it generates an inspection trigger event and extracts the anomaly string logical identifier from the corresponding data packet. This anomaly string logical identifier is specifically represented in the underlying data structure as a string field in the form of "equipment area code-inverter MAC address-DC port number". The inspection management server's storage medium has a pre-installed mapping database. The program calls this database to convert the above-mentioned purely numeric or alphanumeric logical identifier into a target spatial coordinate set. The target spatial coordinate set is represented as a three-dimensional spatial polygonal region boundary formed by connecting multiple latitude, longitude, and elevation points. The vertex data of this polygonal boundary is based on the design coordinates entered from the construction drawings during the site construction period and has been actually calibrated.

[0025] Based on the technical objective of establishing a space-ground coordinated control link, the inspection management server establishes a wireless data communication link with the mobile inspection terminal after takeoff via a wireless broadband communication base station built within the site. The inspection management server packages and sends flight path data containing the target spatial coordinate set to the mobile inspection terminal. After the mobile terminal is in place, the inspection management server issues a transient bias command to the inverter corresponding to the target spatial coordinate set. The specific path for this transient bias command is as follows: the inspection management server sends it via a wired network to the monitoring and data acquisition system, which converts it into a standard industrial control protocol format and then writes it to a specific register address on the corresponding inverter control board via the RS485 bus, thereby forcibly intervening in the control algorithm of the inverter's internal digital signal processor.

[0026] The network time protocol server maintains clock synchronization connections with the monitoring and data acquisition system and the wireless broadband communication base station, broadcasting standard time source signals to the electrical acquisition nodes and inspection mobile terminals within the site. The existence of this network time protocol server ensures that every current sampling point in the actual continuous current sequence recorded by the monitoring and data acquisition system is strictly aligned with every frame of the spatial pixel-level thermal imaging video stream recorded by the inspection mobile terminal, under a unified global timestamp benchmark. This eliminates timing misalignment errors caused by spatial physical location offsets in multi-source heterogeneous data.

[0027] The aforementioned physical devices and communication links together constitute the foundational platform supporting the joint analysis of thermodynamic and electrical multimodal data. The underlying electrical data transmitted and processed within this architecture can be fundamentally defined in matrix form. Let a specific maximum power point tracking loop have a number of parallel connections... For a given photovoltaic string, the real-time DC current sequence of the circuit acquired by the monitoring and data acquisition system at any given time point can be represented as a column vector as shown in the following code block: In the formula, represent The overall current state vector of the tracking loop at the maximum power point at that moment; The natural number index variable for the string, with values ​​ranging from 1 to... ; Representing the A photovoltaic string in The actual DC current value constantly collected by the underlying current sensor; superscript This represents the transpose operation of a matrix.

[0028] See attached document Figure 2 The overall workflow for implementing this safety inspection management method unfolds sequentially according to time sequence and physical causal relationship, specifically covering the following processing steps: The inspection and management system continuously receives real-time DC current values ​​from inverters distributed throughout the site. The system performs array-level comparison calculations. In this embodiment, the root mean square error is used as an evaluation index for the degree of current dispersion. When the current data of a specific photovoltaic string within the same maximum power point tracking loop deviates from the system's average operating current by a dispersion exceeding a set threshold, it is determined that an electrical operating abnormality has occurred in the equipment.

[0029] As a preferred approach, the threshold value is set to 5% to 10% of the system's current average operating current. The system constructs a data object for the inspection trigger event in memory, extracting a combined code of the device section, inverter number, and branch port number from the communication message where data anomalies occur. This combined code constitutes an anomaly string logical identifier representing the abnormal topology node.

[0030] The system reads a pre-built topology mapping table in a relational database, mapping logical identifiers representing electrical connections to a sequence of polygon vertex coordinates in the actual three-dimensional geographic coordinate system of the site. This vertex coordinate sequence delineates specific areas in physical space suspected of having defects, forming a target spatial coordinate set. The system packages this target spatial coordinate set into a flight path mission file with latitude and longitude information and sends it via wireless network to a mobile inspection terminal equipped with an infrared sensor, instructing it to take off and autonomously cruise over the physical airspace corresponding to this coordinate set.

[0031] After confirming its position above the target area via its onboard GPS module, the mobile inspection terminal sends a ready signal back to the system. The system then modifies the power control register of the corresponding inverter via the site's underlying industrial control bus, writing a transient bias command. This command forces the inverter to halt its internal maximum power point tracking calculations and instead adjust its input DC voltage boundary at a set fixed slope. This intervention forces the output current of the target photovoltaic string to exhibit a continuous, monotonous increase or decrease within a preset time window. During this period of actively altering the underlying electrical parameters, the onboard infrared sensor continuously captures images of the target photovoltaic string's surface at a constant frame rate, generating a spatial pixel-level thermal video stream. The underlying data acquisition equipment simultaneously records the current sampling data of the corresponding branch. The thermal video frames and current sampling points are timestamped based on a network time protocol.

[0032] Before extracting the thermal derivative of temperature versus time for each pixel in the spatial pixel-level thermal image video stream, and the electrical power derivative of internal heating power versus time for the actual continuous current sequence, the system needs to ensure that the frame rate time interval of the infrared sensor is not zero to satisfy the boundary conditions for derivative calculation. The system parses the infrared image matrix, extracts the absolute temperature value of each pixel at its spatial coordinate position within a preset time window to form time series data, and calculates the partial derivative of this time series data to obtain the thermal derivative of each pixel. The system extracts continuous current data, and based on the physical relationship in Joule's law that heating power is proportional to the square of the current, calculates the time partial derivative of the square of the acquired DC current value to obtain the electrical power derivative corresponding to the power fluctuation. The basic calculus definitions of the thermal derivative and the electrical power derivative are represented by the mathematical model shown in the following code block: ; ; In the formula, The spatial pixel coordinates in an infrared image are The physical point at The absolute temperature value at any given moment; The thermal derivative of the pixel at that spatial location is calculated and output. Represents the equivalent heat generation power inside the target photovoltaic string; The derivative of the electric power obtained from the derivation; This represents the combined proportionality constant characterizing the internal equivalent impedance and heat dissipation coefficient of the component. This constant is treated as a linear scalar when calculating the correlation of partial derivatives. Represents the time increment.

[0033] In the discretization process of the engineering system, the time increment in the above formula is directly replaced by the actual sampling period of the monitoring and data acquisition system. The specific derivative value is calculated using the first-order backward difference method.

[0034] Since the Joule heat generated by the internal cells of a photovoltaic module requires a certain physical time to conduct to the surface glass, the system shifts the electrical power derivative data on the time axis towards the historical time direction by an order of magnitude of the heat conduction time constant. The specific value of this constant is determined by the product of the total heat capacity of the photovoltaic module's encapsulation materials and the average thermal resistance of each layer of materials. This shift operation mathematically aligns the internal electrical power change excitation with the surface temperature response, completing phase hysteresis compensation. The system executes a statistical calculation program and outputs the Pearson partial correlation coefficient between the aligned thermal derivative time series and the electrical power derivative time series. To avoid calculation errors caused by singular data matrices, the program introduces a variance check mechanism before calculating the correlation coefficient. When the temperature variance within the effective time window is detected to be infinitely close to zero, the correlation coefficient is forcibly output as zero.

[0035] To avoid the biased judgment problem caused by relying solely on single temperature rise data in conventional methods, this embodiment constructs a joint comparison mechanism. The system extracts high-temperature pixels in the infrared image whose absolute temperature values ​​are higher than those of the surrounding normal area. The system compares the correlation coefficient corresponding to the high-temperature pixel with a preset correlation coefficient threshold. This correlation coefficient threshold is typically set in the range of 0.6 to 0.85 based on the statistical distribution of historical measured data. If the correlation coefficient of a pixel is greater than the threshold, the system determines that its temperature changes with internal current through physical coupling, confirming the presence of electrically induced high-resistivity hotspot defects such as hidden cracks or loose connections at that spatial location. If the correlation coefficient of a pixel is less than or equal to the threshold, the system determines that its temperature rise is not affected by current excitation, classifying it as a non-electrical heating pseudo-hotspot caused by bird droppings or external environmental reflections. The system removes the pseudo-hotspot coordinate data from the defect set and outputs a clean spatial location result for electrical defects.

[0036] To support subsequent electrical anomaly identification and multidimensional matrix operations, high-precision discretization and mathematical quantization of the analog electrical signals from the power generation equipment side are required. This process, at the physical level, relies on high-frequency sampling and digital filtering techniques, and its specific implementation includes the following steps: In this embodiment, a Hall current sensor configured at the DC input terminal of the string inverter continuously measures the DC current flowing through each branch at a set sampling frequency. To satisfy the Nyquist sampling theorem and effectively suppress electromagnetic interference generated by the high-frequency switching of the insulated-gate bipolar transistors inside the inverter, the sampling frequency is typically set to more than twice the inverter's switching frequency. The analog signal, after being low-pass filtered by the front-end hardware, is sent to the analog-to-digital converter (ADC), which outputs the raw current sample values ​​with discrete timestamps. The specific selection of the ADC chip and the construction of the peripheral hardware filtering circuit can be designed by those skilled in the art based on the inverter's rated voltage and current specifications; the basic hardware architecture is well-known in the field and will not be elaborated upon here.

[0037] Considering the several orders of magnitude time scale difference between the high-frequency sampling rate of the inverter's underlying control board and the low-frequency polling mechanism of the monitoring and data acquisition system in industrial settings, directly transmitting the original high-frequency sequence would lead to communication bus congestion. Therefore, the inverter's internal digital signal processor first performs internal averaging downsampling on the original current sample values ​​at a preset second-level cycle, generating low-frequency current data messages before uploading them. Due to potential transmission delays and data loss on the RS485 communication bus or power line carrier channel in industrial settings, the current data received by the system from each branch may exhibit slight misalignment on the time axis. In this process, the system extracts the timestamp field from each inverter's data message and sets an allowable jitter window of one to five seconds based on the local time server. The system assigns a unified globally aligned timestamp to data points falling within the same window, thereby ensuring strict alignment of multiple current data streams on the time axis. For random pulse noise generated by strong electromagnetic interference during transmission, the system employs median filtering or moving average filtering algorithms to remove outliers exceeding normal physical extremes.

[0038] Based on the cleaned and aligned high-stability data source, the system enters the physical defect characterization calculation stage. Under ideal physical environment and equipment conditions, multiple photovoltaic strings connected to the same maximum power point tracking loop should maintain highly consistent output current due to the same irradiance and being clamped to the same DC operating voltage by the same controller. When a string experiences a loose connection, microcrack, or partial shading, the increase in its equivalent series resistance inevitably leads to a decrease in current in that branch that differs from other normal branches. Based on this physical causal relationship, the system needs to quantify this degree of dispersion. The system obtains the average operating current of the current loop through summation and averaging, and then calculates the relative deviation reflecting electrical anomalies. The specific mathematical expression adopts the model shown in the following code block: ; ; In the formula, Represents in the given At that moment, the average operating current of the system for all connected photovoltaic strings under the maximum power point tracking loop; The channel index number representing the photovoltaic string; Represents the first after data cleaning A photovoltaic string in Real-time DC current value at any given moment; The first one represents the calculated output. The relative deviation of each photovoltaic string; the larger the absolute value of this value, the greater the degree to which the operating state of the string deviates from the normal operating condition. This represents a minimum positive constant that is preset to prevent division by zero errors.

[0039] To ensure the robustness of industrial-grade software algorithms under extreme conditions, the above relative deviation calculation formula forcibly introduces a constant into the denominator. As a preferred method, The value is set to be between one ten-thousandth and one thousandth of the rated input current of the corresponding inverter's DC port. This physical value range is selected to effectively prevent division-by-zero crashes caused by the system's average operating current approaching zero at night when there is no light, while also avoiding significant numerical interference with calculation accuracy under normal high-irradiance conditions. Furthermore, in the execution logic of the system's underlying code, to avoid frequent algorithm oscillations caused by a single criterion under critical low-light conditions in the early morning or evening, the system establishes a multi-dimensional joint judgment mechanism for current and voltage. The system not only compares in real time... Alongside the inverter's startup current threshold, the system simultaneously checks whether the current DC bus voltage exceeds the set dead zone voltage lower limit. Only when both current and voltage physical indicators are within a consistently stable and effective output range does the system determine that the physical illumination conditions for effective diagnostics are met, and allows the calculation of the relative deviation, thus avoiding the accumulation of invalid data and false alarms. The relative deviation data output from the above calculation constitutes a one-dimensional state vector, providing a standardized underlying data source for subsequently constructing the current deviation matrix and extracting event trigger signals.

[0040] Based on the single-point relative deviation data extracted in the above steps, the system needs to further quantify the dynamic offset state of the entire maximum power point tracking loop in both spatiotemporal dimensions. Considering that the electrical parameters at a single sampling moment are easily affected by transient interferences such as cloud cover or oscillations in the inverter's maximum power point tracking algorithm optimization, this embodiment introduces a time sliding window mechanism. This mechanism constructs a two-dimensional matrix to characterize the comprehensive fluctuation state of the underlying electrical topology, thereby filtering high-frequency disturbances and preserving the evolution trend of real physical defects. The specific implementation process of this part includes the following steps: The system allocates a circular queue buffer area in memory, tracing back along the historical timeline from the current moment, and extracting a relative deviation sequence of a preset sampling period length. To address situations where the buffer queue is not full due to system startup or recovery from a low-level communication interruption, the program incorporates a data readiness verification mechanism. The system only allows matrix concatenation operations when the number of continuously written valid sampling points in the queue reaches the set length of the sliding window, thus preventing memory read out-of-bounds errors or null pointer exceptions in the low-level code. After data readiness, the program arranges the one-dimensional state vectors of the relative deviations of all parallel strings under the same maximum power point tracking circuit within the time sliding window by column or row, forming a string current deviation matrix representing spatiotemporal fluctuations. The specific mathematical construction of this matrix is ​​shown in the following code block: ; In the formula, Representative at The string current deviation matrix generated at each time step; This represents the total number of discrete sampling points contained within the set time window, and its value is a positive integer greater than 1. This represents the historical time step index within the time sliding window, with values ​​ranging from 0 to... _n_ natural numbers; Representing the Each string at a historical moment The relative deviation value.

[0041] To reduce the dimensionality of the multi-dimensional matrix state to a scalar index that facilitates threshold comparison, the system calls the linear algebra library to solve for the Frobenius norm of the aforementioned deviation matrix. This norm, calculated by taking the square root of the sum of the squares of all elements in the matrix, effectively amplifies anomalous abrupt changes and suppresses uniform system background noise. The specific calculation model uses the formula shown in the following code block: ; In the formula, Represents the calculation output The Frobenius norm value at time t; the definitions of the remaining variables are consistent with the aforementioned matrix construction formula.

[0042] To separate static disturbances from dynamic sudden faults, the system calculates the time partial derivative of the continuously calculated norm scalar sequence. In engineered discrete control systems, this derivative operation is implemented through first-order backward difference discretization. The system extracts the norm value at the current moment and the norm value from the previous calculation cycle, subtracts them, and divides by the calculation step size to generate the evaluation index. The relevant difference calculation formula is shown in the following code block: ; In the formula, represent The rate of change of the norm of the string current deviation matrix at any given time; This represents the time step between two norm calculation operations, used to prevent division by zero errors in the program. Strictly limited to a constant greater than zero, in this embodiment its value is equal to the system's sampling period. .

[0043] The core of this time-varying rate of change lies in extracting the transient acceleration of the system's discrete state. A significant spike in the differential value only occurs when transient physical events, such as sudden glass breakage in a photovoltaic module, diode breakdown and short circuit in a junction box, or the formation of a new high-resistivity hot spot, cause a sharp drop in current. For slowly evolving dust accumulation or light-induced degradation of the photovoltaic module, the resulting norm change is extremely gradual, approaching zero after differentiation. This feature extraction logic filters out gradual environmental disturbances at the underlying mathematical level.

[0044] To avoid false triggering caused by transient current changes across the entire field due to large-area, rapidly moving clouds, this embodiment constructs a two-dimensional joint judgment logic based on the norm change rate and static relative deviation. The system monitors in real time. The value. When When the set dynamic mutation threshold is exceeded, the program not only records the corresponding trigger timestamp, but also synchronously backtracks to check the absolute value of the steady-state relative deviation of each string in the corresponding circuit at that moment. .

[0045] In this embodiment, the dynamic mutation threshold is set based on 3 to 5 times the standard deviation of the system's historical norm background noise under normal fault-free operating conditions. Only when... The system will only confirm a real sudden anomaly when the mutation condition is met and the absolute value of the relative deviation of at least one specific photovoltaic string exceeds the steady-state discrete threshold mentioned above.

[0046] Based on the aforementioned extracted multi-dimensional joint physical state indicators, the system enters the final decision-making and event instantiation stage. This process aims to accurately isolate the fault source topology nodes and generate a standardized data carrier that can be directly invoked by the upper-layer scheduling system. Its specific implementation includes the following steps: When the dual-dimensional joint judgment logic confirms the existence of a genuine sudden anomaly, the program needs to accurately locate the specific faulty branch from the loop containing multiple photovoltaic strings. In actual industrial sites, sudden shading or equipment damage is highly likely to affect multiple adjacent photovoltaic strings simultaneously. If only the single extreme value with the largest absolute value is relied upon for optimization, concurrent fault nodes are easily missed. Therefore, this embodiment constructs a multi-objective extraction logic based on adaptive thresholds. The system traverses all relative deviation values ​​under the current maximum power point tracking loop, extracts the channel indices of all channels whose absolute values ​​exceed the set discrete judgment threshold, and then generates a set of abnormal branches. Its mathematical optimization logic adopts the expression shown in the following code block: ; In the formula, This represents the set of physical port indices of the target photovoltaic string that has been identified by the system as potentially having physical defects. Representing the A photovoltaic string in The relative deviation of time; This represents the set discrete decision threshold.

[0047] As a preferred approach, in order to adapt to changes in background noise under different irradiance levels, The value can be dynamically and adaptively calculated based on the current maximum value of the absolute value of the relative deviation under the same circuit.

[0048] After obtaining the physical port index set, in order for the central control layer's inspection and management server to overcome the communication protocol barriers between different underlying equipment manufacturers and uniquely identify the faulty node, the system concatenates the underlying attributes according to a preset data dictionary. The system retrieves the site equipment area code to which the current inverter belongs and the media access control address of the communication network interface from the local relational database, combines them with the aforementioned determined physical port index, and generates an abnormal string logical identifier.

[0049] Furthermore, considering that the underlying program may be unable to query the media access control address in real time due to communication packet loss, the system is configured with an active fault tolerance mechanism. In this case, the program will automatically retrieve the inverter's factory hardware serial number for string replacement, thereby preventing data encapsulation interruption caused by null pointer exceptions. The generated logical identifier will serve as the data primary key, used subsequently to match the corresponding real geographic latitude and longitude coordinates in the site's static topology database, thus providing unambiguous spatial source input for UAV scheduling.

[0050] The system instantiates a data object for the inspection trigger event in memory. This data object not only contains the coded anomaly string logical identifier mentioned above, but also encapsulates on-site characterization data such as the globally aligned timestamp at the time of the fault, the extreme value of the norm time rate of change, and the average system operating current at that time. This structured data carrier is then pushed into a message queue and continuously uploaded to the inspection management server via Ethernet.

[0051] Based on the abnormal string logical identifiers generated in the aforementioned steps, the system needs to convert them into geospatial coordinates that the inspection equipment can directly execute. Considering that the underlying communication network can only transmit the logical codes of electrical ports, this embodiment introduces a spatial mapping mechanism. By establishing a static association dictionary between logical nodes and physical locations, cross-domain data alignment is achieved. The construction and query mechanism of this mapping database includes the following steps: During the construction or final acceptance phase of a photovoltaic power station, construction personnel utilize carrier phase differential measurement equipment to perform high-precision point mapping of the inverters, combiner boxes, and the apexes of the supports for each photovoltaic string throughout the entire station. For the operating specifications of the carrier phase differential equipment and the configuration of the static reference station, those skilled in the art can refer to current engineering surveying standards; the specific implementation details are well-known in the field and will not be elaborated upon here.

[0052] Since engineering surveying typically uses a planar projected coordinate system, while UAV autopilots rely on the Earth's inertial coordinate system for navigation, the system, after reading discrete survey vertex data, first calls a seven-parameter coordinate transformation model to uniformly map all drawing coordinates to standard WGS84 three-dimensional latitude and longitude coordinates. After coordinate system alignment, the system obtains multiple vertex three-dimensional coordinate data for each photovoltaic string under a globally unified reference. These points spatially constitute the external geometric outline of the physical device.

[0053] After obtaining the physical boundary coordinates, the system extracts features from the region enclosed by the array strings based on spatial analytical geometry principles. Considering the significant elevation difference between the front and rear of the photovoltaic array in complex mountainous terrain, to absolutely avoid collisions between the aircraft and the tilted support during approach, the system uses geometric mean calculations in the latitude and longitude plane, while extracting the coordinates of the highest vertex of the array as a safety benchmark in the elevation dimension, and then superimposing a flight offset on top of this. The specific calculation model uses the formula shown in the following code block: ; ; ; In the formula, , , These represent the three-dimensional coordinate components of the spatial geometric center of the photovoltaic string, respectively, corresponding to the longitude, latitude, and absolute cruising elevation of the target waypoint; This represents the total number of valid boundary vertices of the photovoltaic string involved in the calculation, and its value is a natural number greater than or equal to 3. Represents the vertex traversal index; , , Representing the Measured 3D coordinates of each vertex; A mathematical operator that extracts the maximum value from a dataset; This represents the set safe flight offset altitude.

[0054] As a preferred method, The value is typically set between five and twelve meters. This threshold is determined based on the field of view coverage of the UAV's downward-facing camera and the maximum elevation of any protruding structures within the site, such as lightning rods and weather station masts. In the underlying computational logic, if the denominator is determined to be zero or lower than the minimum value of 3 constituting the plane during the data reading phase, it indicates that the data set lacked mapping data in the early stages. At this point, the program will trigger an anomaly detection logic, proactively halting the calculation of the center point of that specific data set to prevent division by zero crashes, and marking the abnormal node as having misaligned coordinates.

[0055] After calculating all spatial feature points, the system deploys a relational database on the server side to establish a mapping table between electrical and physical topologies. The program uses the logical identifier generated earlier as a unique primary key field, and stores the calculated baseline target waypoint coordinates, string boundary vertex arrays, and array installation azimuth angles as non-primary key fields. To improve retrieval efficiency during high-volume concurrent events, the system establishes a B-tree clustered index on the logical identifier field.

[0056] In actual operation scenarios, when the inspection and management system receives a trigger event containing an abnormal string logical identifier, the backend program directly calls the retrieval interface of the mapping library to query the corresponding spatial waypoint coordinates. To ensure the integrity of the scheduling link, this embodiment configures coordinate retrieval compensation logic in the database query engine. When the query result is empty due to human error or storage corruption in the fine mapping coordinates of a specific photovoltaic string, the system will extract the upper-level device code (i.e., the media access control address of the inverter) from the logical identifier, and instead query and return the coarse coordinates of the physical location of the inverter, while simultaneously sending a preset 30-meter area search radius parameter to the UAV.

[0057] Based on the aforementioned topological mapping, the scheduling system needs to transform discrete physical points into continuous flight trajectories that can be directly executed by the UAV flight control equipment. This process involves spatial geometric collision avoidance calculations and optimization solutions under multidimensional constraints, aiming to generate flight commands that balance response time and equipment safety. Its specific implementation includes the following steps: Before scheduling execution equipment, the availability status of all UAVs in the field must be comprehensively assessed to avoid causing low-battery equipment to be forced to take off and crash due to simply pursuing the shortest spatial distance, or interrupting equipment performing higher-priority tasks. Therefore, this embodiment introduces a multi-dimensional weighted optimization logic. When the system receives an inspection trigger event containing spatial coordinates, it simultaneously extracts real-time telemetry data from all online UAVs in the entire field and calculates the comprehensive task cost of the candidate terminals through weighted logic. The mathematical expression of this cost function is shown in the model in the following code block: ; In the formula, Representing the The overall mission cost of using Taiwan's alternative drones to perform current emergency missions; This represents the real-time three-dimensional spatial coordinates of the drone's current location; Represents the reference waypoint coordinates of the target string extracted from the mapping database; This represents a mathematical operator for calculating the Euclidean distance between two objects. This represents the standard cruising speed of a drone in an empty hangar. To prevent division by zero errors, the compensation value for extremely small positive numbers is usually set to 0.01 m / s; This represents the current percentage of the drone's remaining battery state of charge. This represents the minimum safe battery level for mandatory return to base, typically set between 20% and 30%. The maximum value operator is used to significantly increase the cost function output for devices with battery levels below the minimum threshold. This represents the task status flag. The value is 0 when the device is idle, 1 when performing low-priority routine inspections, and 10 when performing high-priority fault reviews. , , These are the normalized weight coefficients for time cost, power risk, and task conflict, respectively, and satisfy the following conditions: As a preferred approach, to emphasize security attributes, The value is usually not lower than 0.5.

[0058] The system iterates through and calculates the comprehensive mission cost of all candidate drones, and selects... The terminal with the lowest value that does not exceed the system's maximum allowable cost threshold is selected as the final execution vehicle. This multi-dimensional joint judgment logic ensures that the selected device possesses both spatial proximity advantages and sufficient physical energy to complete round-trip flight.

[0059] After locking onto the execution terminal, the system needs to generate a discrete sequence of waypoints in the digital twin space. The shortest paths generated by conventional heuristic algorithms often get too close to ground structures, making them susceptible to physical collisions due to sudden crosswinds. To enforce sufficient spatial isolation margin, this embodiment introduces a repulsive potential field model from artificial potential field theory during the path node generation stage. The system reads the 3D elevation point cloud map of the station and calculates the environmental repulsive force on the currently searched node, as shown in the following code block: ; In the formula, The first in the representative space One route point to be expanded The repulsive potential energy encountered by the obstacle; This represents the potential field repulsion gain coefficient, used to adjust the sensitivity of obstacle avoidance behavior, and its value is usually between 0.5 and 2.0. This represents the coordinates of the nearest known static obstacle to the current waypoint, extracted from the point cloud map using the nearest neighbor search algorithm. This represents the shortest Euclidean distance between the two. This represents the minimum physical safety isolation radius. This threshold is usually set to eight to fifteen meters, depending on the fuselage offset that may be caused by the maximum gust level at the airfield. A zero-insignificant constant is introduced to prevent numerical divergence when nodes coincide with obstacles. This repulsive potential field mechanism mathematically constructs a physical spatial isolation zone, forcing Niu Cheng's global navigation to automatically deviate from high-risk defense zones.

[0060] Once the UAV hovers at the target waypoint, to acquire an orthogonal infrared thermal image of the device surface and eliminate emissivity measurement errors, the system needs to reverse-calculate the control compensation angle of the gimbal motor based on the physical installation attitude of the photovoltaic modules. The program extracts the panel tilt and azimuth parameters for this specific string from the underlying mapping database and derives the control commands required for the vertical alignment of the gimbal's optical axis based on three-dimensional coordinate transformation. The specific gimbal yaw and pitch action command formulas are shown in the following code block: ; ; In the formula, This represents the gimbal target yaw angle issued to the optoelectronic pod; This represents the physical installation azimuth angle of the photovoltaic array recorded in the database; This represents the modulo operation, ensuring that the angle value is limited to the standardized closed-loop range; Represents the gimbal target pitch angle. The degree is horizontal forward. Viewed vertically downwards; The fixed installation tilt angle of the photovoltaic panel relative to the horizontal ground; This represents the dynamic compensation margin for pitch angle.

[0061] The global waypoint sequence and local gimbal action commands generated by the reverse analysis are serialized into a standard JSON format task file. The system sends this file to the target UAV's flight control computer via an UHF wireless communication link, driving it to autonomously detach from its nest to perform on-site image acquisition tasks, thus completing a fully automated scheduling closed loop from electrical anomaly detection to physical location verification.

[0062] Based on the global trajectory and gimbal pose commands generated by the aforementioned reverse analysis, the UAV will initiate infrared thermal imaging and visible light dual-band image acquisition after flying over the target photovoltaic string. However, under the conventional maximum power point tracking control mode, the operating current of the photovoltaic string is in a dynamic optimization state, and the power dissipation of some early latent defects is insufficient to form a significant infrared temperature gradient on the component surface. To highlight the physical characteristics of the fault area during image acquisition, this embodiment introduces an active takeover and electrical bias mechanism for the inverter's underlying control. The specific implementation process of this control logic includes the following steps: When the dispatch system confirms that the UAV has entered the hovering and shooting preparation state of the target defense zone, the server sends a communication message containing the target register address and write function code to the target inverter via the industrial Ethernet of the site. This message forces the inverter to exit its internal autonomous optimization state and switch its DC-side control strategy from the conventional power maximization mode to a constant voltage operation mode based on an externally given voltage reference value. For the data frame encapsulation and underlying register address mapping of the industrial communication protocol, those skilled in the art can refer to the communication protocol manuals of the relevant equipment manufacturers. Its internal data flow is well-known technology in the field and will not be elaborated upon here.

[0063] After gaining control of the inverter's underlying layers, the system needs to purposefully alter the operating point of the target circuit to enhance the infrared radiation characteristics of latent defects. According to Joule's law of heating in semiconductor physics, the local thermal power in a defective region is proportional to the square of the DC current flowing through it. Therefore, the system reduces the string's operating voltage, forcing its operating point to shift towards the short-circuit current direction, thereby significantly increasing the actual operating current in the circuit. This operation allows the thermal accumulation rate at high-impedance defect points to far exceed that of normal cells, thus exhibiting high-contrast characteristics in infrared images.

[0064] Because instantaneous fluctuations in ambient light intensity can significantly alter the volt-ampere characteristic curve of photovoltaic modules, directly applying a fixed bias voltage can easily lead to inverter undervoltage and grid disconnection. Therefore, this embodiment constructs a bias calculation model that incorporates adaptive compensation based on meteorological environment, while ensuring strict alignment of timestamps for multi-source heterogeneous data. The specific calculation model for the target bias voltage uses the formula shown in the following code block: ; In the formula, This represents the target bias voltage reference value calculated from the output. This represents the moment when the inverter takes over control. Locked steady-state maximum power point voltage; This represents the set voltage bias depth coefficient; This represents the irradiance compensation gain coefficient, and the specific value is determined based on the open-circuit voltage temperature coefficient specified at the factory for this batch of photovoltaic modules. The timestamp of the current ambient irradiance, collected in real time by the weather station at the site, has been mandated by the program to be consistent with... Strict synchronization; The reference irradiance under standard test conditions is represented by a value of 1000 watts per square meter. This represents the real-time ambient temperature collected synchronously with the parameters mentioned above; This represents the standard ambient reference temperature, typically taken as 25 degrees Celsius. The value is a very small positive compensation value actively introduced by the system, and is set to 0.01.

[0065] The technical purpose of introducing the absolute value of the temperature difference as a penalty term in the denominator of this formula is that when the ambient temperature deviates significantly from the reference operating conditions (such as in extremely cold or hot environments), the internal resistance characteristics of the photovoltaic cell deteriorate nonlinearly. In this case, increasing the denominator weakens the magnitude of irradiance compensation, preventing the final calculated reference voltage from being too low and falsely triggering the inverter's DC-side undervoltage protection logic. Simultaneously, The introduction of this method effectively avoids the division-to-zero overflow crash caused when the ambient temperature is exactly equal to the reference temperature from a mathematical perspective.

[0066] The system writes the calculated target bias voltage reference value into the inverter's target register via a communication link. To prevent transient current spikes caused by bias from damaging the bypass diodes inside the components, this embodiment employs dual hardware and software boundary protection logic. At the software scheduling level, the program monitors the return current during bias in real time; if it reaches 90% of the rated upper limit, a voltage boost command is immediately issued. At the hardware control level, the inverter's local digital signal processor is equipped with microsecond-level response overcurrent protection firmware, achieving low-level hard interception independently of the external network.

[0067] Based on the aforementioned electrical bias command, the inverter changed the operating point of the photovoltaic string, resulting in a significant increase in the localized heat generation power in the latent defect area. However, photovoltaic modules are typical multi-layered composite physical structures. There is an inherent physical time delay between the generation of abnormal Joule heating at the internal cells and the heat passing through the encapsulation film and the front surface covering glass to form a stable temperature gradient that can be captured by an infrared camera. Simultaneously, at the electrical control level, directly issuing a step-like bias voltage reference value to the inverter's DC side may trigger charging and discharging surges in the DC bus capacitor, and even cause severe oscillations in the active power on the AC grid-connected side. To address the asynchronous coordination problem of the aforementioned thermodynamic conduction delay and electromechanical transient impact, this embodiment introduces a thermal conduction relaxation time window and a voltage ramp control model. This process aims to ensure a smooth transition of the underlying electrical intervention and guide the UAV to acquire images at the optimal moment for thermal characteristic exposure. The specific implementation process includes the following steps: Before officially issuing the bias voltage, the system needs to consider the transient withstand capability of the inverter hardware. According to the capacitor's volt-ampere characteristic equation, drastic changes in DC voltage will generate huge transient charging and discharging currents. To prevent this current from triggering the underlying hardware overcurrent protection, the system dynamically calculates a safe voltage regulation slope based on the physical parameters of the inverter's internal topology. The system extracts the nominal electrical parameters of the target inverter and comprehensively evaluates its DC bus energy storage status and grid-connected power regulation capability. The specific calculation model for the target slope uses the formula shown in the following code block: ; In the formula, This represents the maximum allowable rate of change of the slope of the bias voltage in the calculated output. This represents the absolute maximum voltage change rate that inverter hardware manufacturers are required to specify in their technical specifications. This represents a mathematical optimization operator that extracts the smaller of two values ​​to ensure that the calculation result never exceeds the physical limits of the hardware. The electromechanical transient safety margin factor represents the set value. As a preferred method, it is used to reserve sufficient buffer space for grid-connected current. The value of this factor is usually set between 0.60 and 0.80. This represents the total physical support capacitor capacity of the inverter's DC bus. This represents the nominal operating voltage on the DC side under the current power grid environment; This represents the rated grid-connected output power of the inverter on the AC side. This represents the control cycle time of the inverter's underlying voltage loop. The value of the small positive number introduced into the system to prevent the division of zero is set to 1×10. -6， This is used to prevent program crashes caused by the loss of control cycle parameters due to underlying communication parsing errors.

[0068] After the inverter begins voltage biasing according to the aforementioned ramp rate, the system synchronously starts its internal timing state machine. According to the principles of solid-state heat transfer, the one-dimensional diffusion time of an internal heat source in a multi-layered medium is directly proportional to the square of the physical thickness of each layer and inversely proportional to its thermal diffusivity. To ensure that the UAV captures images when the thermal contrast on the front surface of the photovoltaic module reaches its peak, the system establishes a relaxation time calculation model based on the heat transfer equation of the multi-layered physical medium. Simultaneously, considering that outdoor wind convection significantly removes surface heat and delays the establishment of the thermal field, this model introduces a wind speed compensation term based on real-time meteorological data. The specific calculation formula is shown in the following code block: ; In the formula, This represents the length of the calculated thermal conduction relaxation time window; This represents the total number of physical heat transfer layers on the front side of a photovoltaic module. For a conventional double-glass module, The value of is 2; The traversal index representing the heat transfer level; Representing the The measured physical thickness of the layer medium; Representing the The factory-calibrated thermal diffusivity of the medium; The additional zero-prevention parameter in this embodiment is used to prevent the denominator from being zero due to missing material database properties; Represents the natural logarithm operator; The thermal response threshold ratio representing the desired surface temperature difference extreme value is typically limited to between 0.80 and 0.95. This represents the time compensation gain coefficient resulting from forced convection cooling. Its physical meaning is the number of extra waiting seconds required to compensate for each additional unit of wind resistance.

[0069] Through the above multi-dimensional thermodynamic modeling, the system accurately derives the optimal heat dissipation waiting time under the coupling of a specific material system with the current complex meteorological conditions, thus eliminating the blind spot of thermal field sampling in the traditional timed snapshot mode.

[0070] After obtaining the accurate thermal conduction relaxation time window, the dispatch system issues a cooperative hovering command to the UAV hovering over the target defense zone. Upon receiving the command, the UAV's flight control system locks the current wind-resistant hovering three-dimensional coordinates and gimbal pitch angle, and starts a local software timer. During the waiting period, if the inertial navigation module at the flight control level detects that the spatial drift caused by wind exceeds 0.5 meters, the system will autonomously trigger fine-tuning compensation to maintain the line-of-sight alignment lock.

[0071] When the timer reaches the calculated relaxation time threshold, it means that the thermal characteristics on the surface of the photovoltaic module have reached the theoretical extreme point excited by the electrical bias. At this time, the mission computer inside the UAV sends a synchronization trigger pulse to the photovoltaic pod, driving the high-resolution infrared thermal imager and the visible light camera to simultaneously complete the multimodal image acquisition of the current string.

[0072] For the encapsulation of the underlying communication protocol for closed-loop hovering control of UAV spatial position and camera shutter triggering, those skilled in the art can refer to existing aerial surveying standards and open-source micro-aircraft link protocols. The specific data link timing and register mapping mechanisms are well-known technologies in the field and will not be elaborated upon here. After image acquisition is completed, the system immediately enters the image transmission and inverter control release phase, restoring the station's normal grid-connected power generation process.

[0073] Based on the aforementioned established thermal conduction relaxation time window, the UAV will perform image acquisition at specific moments. To achieve accurate mapping between the infrared temperature distribution on the photovoltaic module surface and its transient electrical operating conditions and meteorological environment, it is essential to ensure that the image shutter time, the voltage / current time relayed by the inverter, and the irradiance time from the weather station are strictly homogeneous on the physical time axis. Since the local crystal oscillators of different hardware terminals have initial phase differences and frequency offsets, this embodiment constructs a global clock synchronization and local dynamic interpolation alignment mechanism based on the Network Time Protocol (NTP). The specific implementation process of this mechanism includes the following steps: Before UAV takeoff and within the set mission duration, the dispatch center server, acting as the global Layer 1 time reference source, periodically broadcasts time synchronization messages to the inverter communication management unit, weather station data collector, and UAV flight control terminal within the site. The system calculates the system clock offset of each terminal relative to the server by recording four key timestamps of the message's round-trip time in the network link. For the underlying data packet encapsulation and round-trip delay calculation of the network time protocol, those skilled in the art can refer to the RFC5905 standard protocol; its message structure and basic offset calculation are well-known technologies in the field and will not be elaborated upon here.

[0074] Considering the drastic temperature changes faced by drones during flight, the oscillation frequency of their onboard hardware clocks will experience temperature drift. Simply relying on low-frequency time synchronization messages cannot guarantee millisecond-level alignment accuracy. Therefore, this embodiment embeds a clock drift compensation algorithm within each terminal. The system calculates the dynamic drift rate of the local clock by fitting the offsets obtained from multiple historical synchronization handshakes and performs real-time compensation on the local hardware tick timer. The formula for calculating the dynamic compensation time is shown in the following code block: ; In the formula, This represents the globally aligned timestamp output after drift compensation. Represents the current absolute time output by the terminal's local hardware timer; This represents the static clock offset calculated from the most recent network time synchronization handshake. The local timestamp representing the most recent successful time synchronization. The dynamic drift rate of the local clock is represented by a value obtained by least-squares linear fitting of the most recent ten synchronization offsets.

[0075] Under the premise of a unified global time base, when the UAV triggers the camera shutter according to the relaxation time window, the onboard mission computer will extract the current precise time. The image file's metadata tag (EXIF) is then sent back to the server. Simultaneously, the inverter and weather station, according to their inherent low-frequency sampling period (usually once per second or every two seconds), append their respective compensated timestamps to the electrical and meteorological characteristic data they collect and continuously push them to the scheduling server's circular memory queue.

[0076] Since the triggering time for UAV image acquisition is random and continuous, its timestamp is highly unlikely to fall exactly at the discrete sampling time points of the inverter and the weather station. To obtain the precise electrical and physical parameters at the moment of image acquisition, the system retrieves the two nearest valid sampling data points before and after the image timestamp in a circular memory queue, and reconstructs the feature values ​​of the target time using a linear interpolation algorithm. The mathematical model for reconstruction interpolation uses the formula shown in the following code block: ; In the formula, This represents the interpolated and aligned values ​​corresponding to the drone shutter speed. The equipment operating parameters include string operating voltage, current, ambient irradiance, and wind speed, etc. and These represent the timestamps retrieved by the system from the memory queue. The actual sampled feature values ​​before and after; and Represents the globally aligned timestamps associated with each of the two sampling points; A zero-reduction protection constant is implanted into the system to avoid program crashes caused by the retransmission of redundant underlying data, resulting in identical timestamps.

[0077] After completing the above data interpolation, the scheduling server structures and packages the infrared / visible light image files, the reconstructed inverter voltage / current values, and the synchronized ambient irradiance and temperature data. This structured data package is stored in a distributed database using a unified identifier, thus providing a high-quality multimodal data foundation for subsequent photovoltaic latent defect diagnosis models that is absolutely homogeneous in the time dimension and highly coupled in the operating condition dimension.

[0078] Based on the aforementioned multimodal dataset strictly aligned via network time protocol, the system acquired structured files containing infrared thermal images, visible light images, and synchronous electrical and meteorological parameters of the photovoltaic strings. Conventional defect diagnosis schemes often rely directly on the absolute temperature extreme values ​​in infrared images for judgment. This reliance on a single extreme value easily leads to numerous false alarms and false negatives when faced with sudden changes in wind speed or cloud cover. To delve deeper into the physical degradation mechanisms within photovoltaic modules, this embodiment introduces a joint extraction technique of spatial thermal derivatives and temporal electrical power derivatives. This extraction process aims to mathematically decouple and quantify the apparent thermal phenomena from the underlying electrical impedance characteristics. The specific implementation process includes the following steps: The system reads the raw thermal radiation data stream generated by the airborne infrared camera and, combined with the lens focal length and the relative flight altitude of the drone during the shooting, transforms it into a two-dimensional absolute temperature matrix with a mapping relationship to actual physical dimensions. Since the drone inevitably undergoes pitch and yaw maneuvers while hovering in the air, the acquired component images often exhibit perspective distortion. To eliminate the errors caused by distortion in subsequent spatial gradient calculations, the program extracts the edge contours of the photovoltaic panels from the visible light images, calculates the homography matrix, and performs perspective transformation correction on the temperature matrix.

[0079] During this process, because the edge extraction algorithm may be affected by reflections or dirt, leading to collinearity of the extracted point sets, the program introduces Singular Value Decomposition (SVD) and condition number threshold verification before solving to prevent the calculated homography matrix from becoming singular and resulting in no solution. As a preferred method, if the calculated condition number is greater than 1 × 10⁻⁶... 4 The system will determine that the current matrix is ​​close to singular, actively discard the reconstruction of the frame and reuse the transformation parameters of the previous valid frame, thereby ensuring the robustness of the perspective correction algorithm.

[0080] In the reconstructed absolute temperature matrix, due to uneven illumination or wind-cooled convection, the overall temperature of a normally operating photovoltaic panel may experience macroscopic drift, but the temperature distribution in adjacent areas within the panel remains smooth. The core physical characteristic of latent defects lies in their localized anomalous Joule heating, which creates a significant temperature gradient within a very short physical space. To accurately locate such anomalies, the system solves for the partial derivatives of the temperature matrix in two-dimensional space, calculating and extracting the spatial pixel thermal quotient. The specific solution model uses the formula shown in the following code block: ; In the formula, The coordinates in the image plane are The spatial pixel thermal quotient value corresponding to each pixel, with its physical unit being degrees Celsius per meter, is used to quantify the temperature change rate within a unit physical distance. This represents the absolute physical temperature of the coordinate point after perspective correction. These are the pixel temperatures in the four directions (up, down, left, and right) immediately adjacent to the pixel. This represents the ground sampling resolution of the current infrared thermal image, i.e., the actual physical length corresponding to a single pixel.

[0081] This calculation step eliminates large-area common-mode temperature background drift through differential operations, resulting in a more accurate final output. The matrix can highly purify and highlight the high-frequency thermal characteristics caused by internal impedance anomalies.

[0082] Besides the abnormal thermal distribution in the spatial dimension, severe internal degradation inevitably leads to a decrease in the overall electrical output capability of the photovoltaic string. Based on the aforementioned step-down bias command, the operating point of the photovoltaic string is forced to deviate from the maximum power point. The system extracts the electrical transient data sequence synchronized with the infrared image timestamp and calculates the dynamic power derivative of the string. When a normal, defect-free photovoltaic string approaches the short-circuit current region, the slope of its power decrease with voltage follows a defined semiconductor theoretical curve. For strings with bypass diode breakdown or significantly increased series parasitic resistance, the slope will be significantly distorted. To isolate the interference of real-time irradiance fluctuations on electrical characteristics, this embodiment constructs an electrical power derivative calculation model incorporating meteorological normalization logic. The specific formula is shown in the following code block: ; In the formula, Represents the moment the drone shutter is triggered. The calculated transient power derivative of the photovoltaic string is physically represented as the normalized rate of change of active power caused by a unit voltage change. and Representing time respectively and preceding time The actual operating power of the string DC side transmitted back by the inverter; This represents the set time step for the micro-business calculation, which is used to capture the dynamic response during the bias period. This step is usually set between 1 and 3 seconds. and These represent the real-time environmental irradiance synchronously transmitted back from the weather station; The reference irradiance under standard test conditions is represented by a value of 1000 watts per square meter. and The DC voltage of the string at the corresponding moment; To prevent the irradiance from being zero, the constant is set to 0.1 to prevent calculation divergence caused by zero irradiance at night or during test shutdowns. The minimum positive lower limit of the voltage difference set for the system; This is a function to extract the sign of a variable.

[0083] The system will generate a high-resolution two-dimensional spatial pixel thermal matrix. With the one-dimensional transient electric power derivative A structured combination is performed to form a comprehensive feature tensor for subsequent latent defect diagnosis. In this embodiment, this cross-dimensional feature fusion has a clear physical causal exclusion logic.

[0084] Specifically, the system sets up a multi-dimensional joint criterion based on a preset physical attribution matrix. For example, when The local extrema of the matrix exceed the preset abnormal thermal gradient threshold (e.g., greater than 15 degrees Celsius per meter), but synchronously... If the rate of change remains within ±5% of the healthy baseline curve, the system determines that the hotspot is a non-structural anomaly caused by external physical obstruction (such as bird droppings or leaves); otherwise, if... The matrix exhibits dispersed thermal gradient anomalies, and synchronously... If the rate of numerical degradation exceeds 20 percent of the baseline slope, the system, based on multidimensional weighted logic, determines with high confidence that irreversible electrical degradation has occurred inside the component.

[0085] Based on the aforementioned extracted spatial pixel thermal quotient and transient power quotient, the system obtains the component's characteristic data in both spatial and temporal dimensions. However, the changes in the electrical output characteristics of the photovoltaic module and the transfer and dissipation of internal heat are at different time scales in terms of physical mechanisms. Specifically, the power quotient on the DC side of the inverter responds to environmental radiation or bias commands at the millisecond level, while the Joule heat generated by internal abnormal impedance is transferred through the silicon wafer, encapsulation film, and tempered glass to the surface and captured by the infrared camera. This process is constrained by the thermal inertia of the multilayer dielectric, resulting in significant phase hysteresis. Therefore, although the preceding data achieves absolute alignment in the timestamps of acquisition and recording, in terms of physical causality, the thermal characteristics displayed by the infrared image at the current moment are actually the final manifestation of the accumulated electrical conditions over a continuous period of time. To eliminate this cross-domain misalignment between electromechanical and thermodynamic response rates, this embodiment constructs a dynamic phase hysteresis compensation model, the specific implementation of which includes the following steps: To quantitatively describe the hysteresis effect of heat in a multi-layered physical structure, the system extracts the physical properties of the target component based on a lumped-parameter thermodynamic model and dynamically calculates the equivalent thermal time constant under the current operating conditions by combining real-time wind speed and temperature data from the site. This time constant reflects the rate at which the component surface temperature responds to changes in the internal heat source. The specific solution model uses the formula shown in the following code block: ; In the formula, The equivalent heat conduction time constant of the photovoltaic module, as calculated, is expressed in seconds. Represents the total number of physical material layers on the cross-section of a photovoltaic module; The traversal index representing each material layer; , and Representing the first The density, specific heat capacity, and physical thickness of the layer materials (such as tempered glass, ethylene-vinyl acetate copolymer, battery cells, and backsheet) as specified by the manufacturer; molecular part The physical meaning is the total equivalent heat capacity of the component per unit area; Represents the natural convection heat transfer coefficient in a windless environment; The wind speed heat transfer gain coefficient represents the wind speed under forced convection conditions, and its value is usually between 3.0 and 5.0. This represents real-time wind speed data synchronously transmitted back from the weather station; The zero-prevention constant introduced for the system is used to prevent the solution from crashing when the denominator heat transfer parameters are all zero due to sensor failure.

[0086] After obtaining the accurate thermal time constant, the system inputs the discrete transient power derivative time series into a digital filter, artificially applying delay and smoothing effects to the high-frequency electrical characteristics, forcing them to align with the slower thermal characteristics on the time axis. Based on the discretized derivation of Newton's law of cooling, this embodiment uses a first-order inertial low-pass filter model for phase hysteresis mapping. The specific discrete-time compensation formula is shown in the following code block: ; In the formula, This represents the derivative of the equivalent electrical power output after phase hysteresis compensation, and this value characterizes the electrical state that the current physical thermal field should correspond to. This represents the derivative of the actual electric power at the current moment, calculated in the preceding steps. This represents the historical value of the equivalent power microquote after compensation at the previous sampling time; This represents the discrete sampling time step set by the system, which in this embodiment is usually equal to the inverter data update cycle (e.g., 1.0 second); Represents the natural exponential function; To prevent division by zero of extremely small positive numbers, it is used in Protect the validity of exponential calculations when the solution approaches zero abnormally.

[0087] The system performs pixel-level deep fusion of the acquired spatial pixel thermal quotient matrix and the phase-aligned equivalent electrical power quotient to calculate the spatiotemporal coupling degradation index corresponding to each spatial coordinate point. This index aims to accurately quantify the severity of irreversible degradation within the component through mutual corroboration of electrothermal causality. The specific calculation model is shown in the following code block: ; In the formula, Represents image coordinates The spatiotemporal coupling degradation index at the location; The spatial heat map of the current pixel; The background thermal gradient threshold represents the normal and healthy components and is used to filter out thermal noise caused by normal physical structures such as frame heat dissipation. The value is usually 5.0 degrees Celsius per meter. The operator ensures that only abnormal heating exceeding the baseline threshold is included in the penalty calculation; The equivalent power derivative after phase compensation; The theoretical healthy electrical power microquotient benchmark value for the target component under the same irradiation and temperature conditions; The zero-prevention constant set for the system; and The feature normalization weighting coefficients are determined based on historical data and experience from specific stations. The non-linear penalty index is set, typically ranging from 1.5 to 2.0.

[0088] Based on the aforementioned multidimensional joint solution model, the determination of the output result no longer depends on a single temperature extreme point in the infrared image. Only regions exhibiting an abnormally steep thermal gradient in space and strictly matching an abnormal attenuation of electrical output capability in time phase are considered as having an abnormally steep thermal gradient. Only then will the level significantly increase. By setting a comprehensive degradation threshold, the system not only achieves high-precision spatial positioning of latent defects such as microcracks and desoldering in batteries, but also effectively suppresses false alarms induced by transient non-structural factors such as environmental reflection and bird obstruction, greatly improving the level of automation confidence in the drone inspection of photovoltaic power stations.

[0089] Based on the aforementioned phase hysteresis compensation model, the system obtains the spatial thermal derivative and equivalent electrical power derivative, which are strictly aligned in terms of time axis and physical causality. Under ideal constant environmental conditions, the electromechanical degradation degree of the photovoltaic module can be quantified by directly calculating the conventional correlation coefficient between these two factors. However, in actual outdoor operation scenarios, drastic fluctuations in ambient irradiance and sudden changes in wind speed act as strong common-mode interference sources, simultaneously driving changes in the module's electrical output power and surface thermal field distribution. If these environmental common-mode interferences are ignored and the correlation of electrothermal characteristics is directly calculated, collinearity errors can easily occur due to the synchronous driving of external meteorological conditions, leading the diagnostic algorithm to misjudge normal cloud shadow fluctuations as structural damage within the module. To isolate the common-mode coupling effect caused by meteorological variables and restore the true intrinsic electrothermal relationship within the module, this embodiment introduces a partial correlation measurement mechanism. Its specific implementation process includes the following steps: Before performing statistical measurements, the system uses the shutter trigger moment of the infrared image captured by the drone as a reference point to extract a continuous time series back into the historical time domain. The system extracts the spatial thermal derivative time series set of the target pixel region, the equivalent electrical power derivative time series set after phase compensation, and the irradiance time series set synchronously collected by the weather station.

[0090] To prevent computational divergence under extremely stable operating conditions, the system performs variance checks on the environmental irradiance and power sequences within the window before formally entering the correlation calculation. A timing window is marked as valid and flows into the next calculation stage only if the standard deviation of the irradiance sequence is greater than a set threshold (typically 10 watts per square meter) and the inverter is not under a power-limited shutdown command. For steady-state windows with excessively small variances, the system skips partial correlation calculations and directly reuses the measurement results from the previous valid window.

[0091] After obtaining an aligned and valid time-series sample window, the system needs to quantify the degree of direct linear correlation between each physical variable. The program calculates the zero-order correlation coefficients between the thermal derivative and the electrical derivative, the thermal derivative and irradiance, and the electrical derivative and irradiance, respectively. The specific calculation model uses the formula shown in the following code block: ; In the formula, Representative input feature sequence and The zero-order Pearson correlation coefficient between them; This represents the total number of discrete samples contained within the time-series sliding window; This is the traversal index for the sample sequence; and Each represents a parameter in the first place. The physical values ​​of each sampling point; and These represent the arithmetic mean of the corresponding variables within that time window; The system introduces a feature to prevent overflow and crashes in underlying floating-point operations when data fluctuations are minimal.

[0092] Based on the zero-order correlation coefficients of each parameter, the system uses partial correlation dimensionality reduction projection to remove the common variation caused by changes in environmental irradiance from the total electrothermal correlation. The physical essence of this calculation logic lies in quantifying the net correlation between localized abnormal heating of photovoltaic modules and the dynamic decay of overall electrical power when the environmental irradiance is mathematically locked to an equivalent constant. The specific solution model is shown in the following code block: ; In the formula, Represents the control of environmental irradiance Following the impact, space heat micro-business With equivalent power derivative The first-order partial correlation coefficient between them; This represents the zero-order correlation coefficient calculated between thermal micro-commerce and electrical micro-commerce; The zero-order correlation coefficient between the thermal derivative and irradiance; The zero-order correlation coefficient between electrical micro-quota and irradiance; The zero-constant set for the system.

[0093] After obtaining the partial correlation coefficient, to avoid spurious correlations in small samples caused by random meteorological noise, the system introduces a statistical significance T-test mechanism. The program calculates the T-statistic of the judgment index, and its formula is as follows: ; In the formula, The calculated test statistic; To control for a partial variable and the remaining sample degrees of freedom; It is a set minimum positive number.

[0094] Under the premise of passing the obviousness test, the system jointly weights the partial correlation coefficient with the extreme value of the absolute thermal gradient extracted in the previous step. When the extreme value of the spatial thermal derivative of a specific pixel area in the image exceeds the preset noise floor threshold, and the absolute value of its corresponding partial correlation coefficient exceeds the set threshold (usually set between 0.60 and 0.85), the system determines that there is a highly definite internal physical structure degradation in that area (such as hidden cracks, abnormal surges in series resistance, or bypass diode breakdown). Conversely, if a local area exhibits abnormally high temperature, but the calculated partial correlation coefficient is below 0.30 or fails the T-test, it indicates that the heating behavior is not intrinsically correlated with the electrical dynamic characteristics of the string. Based on this, the system determines that the anomaly belongs to an external non-structural factor (such as local leaf shading) and automatically filters out such alarms at the software level. This judgment mechanism, which combines multidimensional variable stripping and statistical verification, effectively suppresses the interference of environmental noise on the diagnosis of latent defects.

[0095] Based on the aforementioned spatiotemporal multidimensional features and partial correlation metric calculations, the system can effectively isolate common-mode interference caused by environmental meteorological variables. However, in the actual operating environment of photovoltaic power plants, the extreme high-temperature regions captured by infrared cameras are not always caused by internal electrical degradation. Factors such as normal current accumulation and heat generation from the module backplane junction box, physical obstruction caused by surface foreign objects, and specular reflection of sunlight by nearby metal frames will all appear as obvious high-temperature areas in the original thermal image. To accurately separate real electrical defect hotspots from the pseudo-hotspots caused by normal physical structures or environmental reflections, this embodiment constructs a comprehensive filtering mechanism combining viewing angle emissivity compensation and a deep convolutional neural network. Its specific implementation process includes the following steps: Based on the Stefan-Boltzmann law of thermal radiation, the infrared emissivity of the high-transmittance tempered glass covering the surface of a photovoltaic panel is not a constant, but rather decreases nonlinearly with changes in the zenith angle observed by the camera. During high-angle shooting, the uncompensated temperature data will be significantly lower than the actual physical temperature. To obtain accurate absolute temperature characteristics, the system dynamically calculates the actual emissivity of each pixel based on the pitch and yaw angles of the airborne gimbal and performs a physical inverse mapping of the temperature matrix. The compensation model uses the formula shown in the following code block: ; In the formula, Represents the image coordinates after viewpoint compensation. The true absolute thermodynamic temperature at that location; Represents the raw, uncompensated temperature output by the infrared detector; The equivalent radiation temperature representing the sky background was synchronously collected by airborne environmental sensors. Represents the zenith angle observed by the current camera. The actual infrared emissivity of the tempered glass; To prevent division by zero positive constants, used to avoid division by zero overflow crash when the observation angle tends to be horizontal, which would cause the calculated emissivity to be minimized.

[0096] In this formula, the system introduces a maximum value operator. To prevent negative values ​​in the molecular part due to transient measurement noise from the infrared sensor, this mathematically avoids complex number anomalies or program crashes caused by taking the fourth root of a negative number. This calculation step restores the true Joule heating intensity on the component surface, providing a physical foundation with a unified quantitative standard for subsequent model input.

[0097] After obtaining accurate absolute temperature features, the system introduces a deep learning mechanism to perform pattern recognition on the morphology, topology, and multimodal context of the hotspots. In this embodiment, the network model is improved from the ResNet-18 architecture to a dual-branch feature extraction network. Before constructing the input tensor, to adapt to the static graph computation characteristics of convolutional networks, the system uses a bilinear interpolation algorithm to uniformly resample the irregular hotspot regions cropped from the original image to a fixed pixel dimension (usually set to 224×224), and performs min-max normalization on the absolute temperature matrix and the visible light grayscale image respectively. This preprocessing logic ensures that the input data is scaled to the [0,1] interval, avoiding the large numerical temperature gradient from overwhelming the subtle texture features of the visible light image during backpropagation. The preprocessed multimodal tensor dimension is defined as 224×224×2.

[0098] The network comprises a backbone feature extraction stream and a skip connection stream. The backbone stream consists of four cascaded residual stages, each containing multiple 3×3 two-dimensional convolutional kernels, batch normalization layers, and leaky corrected linear units. Through forward propagation, the network extracts geometric features such as edge gradients and shape contours of the hotspot in shallow layers, and abstracts topological semantic features of the relative positions of the hotspot with component borders and junction boxes in deeper layers. At the network's end, a global average pooling layer compresses the multi-dimensional feature map into a one-dimensional dense vector, which is then connected to a fully connected layer with one output node. This layer uses a sigmoid activation function to output a probability value between 0 and 1, which, in practical terms, represents the confidence level in determining whether the current high-heat area belongs to a pseudo-hotspot (such as normal heat from the junction box or specular reflection).

[0099] To ensure the model's generalization ability under different lighting and environments, the system uses a multi-source labeled dataset for gradient backpropagation training during the offline phase. Training samples are derived from over 10,000 hotspot slices in a historical inspection database containing clear physical attributions. The labeling system is defined as follows: genuine defect hotspots confirmed as physically degraded through manual or electroluminescence retesting are labeled 0; pseudo-hotspots caused by junction boxes, reflections, or external foreign objects are labeled 1. The training process uses binary cross-entropy with numerical stability protection as the loss function, the specific formula of which is shown in the following code block: ; In the formula, This represents the total network loss value for a single training batch. This represents the total number of samples in the current batch. Index for sample traversal; For the first Each sample has a real, manually labeled tag; Predicted probabilities output by the fully connected layer of the network ; A very small positive constant is used to prevent logarithmic divergence, which is used to prevent the logarithmic operation from approaching negative infinity when the network outputs absolute zero or one in the early stages. These are the L2 regularization weight coefficients; This is the sum of squares of the weights of all convolutional kernels in the network.

[0100] During the online inference phase, the system obtains the absolute temperature extreme value of the target area and subtracts it from the background arithmetic mean temperature of the surrounding normal area to obtain the absolute temperature difference feature. As a preferred approach, when the probability of a false hotspot output by the judgment network is greater than the classification threshold (usually set to 0.85), the system directly marks the area as normal physical structure heating or optical artifact, blocking the subsequent defect alarm link. When the probability of a false hotspot is in the ambiguous zone between 0.3 and 0.85, the system introduces the absolute temperature difference as an auxiliary weighted judgment condition: only when the absolute temperature difference significantly deviates from the empirical threshold (such as exceeding the preset 12-degree Celsius temperature rise limit of a normal junction box) will the system forcibly overturn the neural network's conclusion of a false hotspot and reclassify it as a suspected physical degradation point that requires attention. This fusion judgment mechanism, combining data-driven and mechanism-driven boundary analysis, effectively compensates for the potential misjudgments and omissions that may occur when a single model faces rare extreme operating conditions.

[0101] Based on the absolute temperature features, spatial thermal derivatives, and the actual defect areas identified through the aforementioned steps, the system achieves deep decoupling from the apparent image to the internal features. However, the infrared and visible light images acquired by UAVs are unstructured pixel arrays based on the camera's perspective, while the actual operation and maintenance scheduling of photovoltaic power plants heavily relies on standardized electrical topologies. To transform diagnostic results into executable maintenance work orders and avoid maintenance personnel blindly searching for fault points on-site, this embodiment constructs a device topology association matrix that integrates multi-dimensional degradation features and performs final alarm classification and result output based on weighted logic. Its specific implementation process includes the following steps: High-risk pixel clusters in the image must be precisely anchored to specific physical component entities. The system reads the high-precision RTK (Real-Time Kinematic Differential) latitude and longitude coordinates and gimbal attitude angle at the moment the drone shutter is triggered, and combines this with the Building Information Model (BIM) data entered during the initial construction of the site to perform coordinate system transformation. The program maps the centroid coordinates of the real physical degraded pixel clusters, filtered for pseudo-hot spots, to a discrete topological space containing three dimensions. The coordinate axes of this topological space are defined as the combiner box / inverter number. String number and photovoltaic module number Through this mapping, disordered pixels are structurally converted into specific device hardware address identifiers. This provides accurate positional indices for subsequent matrix filling.

[0102] After completing the topology mapping, the system extracts corresponding multi-dimensional quantitative indicators from the historical feature cache for each component entity suspected of having defects and performs mathematical fusion. To avoid determining failure based solely on infrared temperature difference under low irradiance, this embodiment constructs a comprehensive degradation metric tensor that considers thermal abrupt changes, electromechanical causal relationships, and environmental boundary conditions. The specific calculation model for the correlation matrix elements adopts the formula shown in the following code block: ; In the formula, The topological coordinates are The comprehensive degradation diagnostic metric corresponding to the photovoltaic module has its numerical range normalized to characterize the severity of irreversible physical losses within the module. This represents the maximum absolute temperature difference extracted within the region of this component; This is the theoretical normal temperature difference benchmark value of the component under the current environmental conditions, which is usually obtained by looking up a table in real time based on the ambient temperature and irradiance returned by the weather station. This represents the maximum value of the spatial pixel thermal quotient within the region, through... Operators filter out negative fluctuation interference; The partial correlation coefficient of electrothermal energy, calculated in the aforementioned steps and passed the T-test, directly reflects the strength of the intrinsic causal relationship between thermodynamic anomalies and the decrease in electrical output. , , These are the normalized weight coefficients for the three dimensions mentioned above.

[0103] After obtaining all elements of the correlation matrix, the system no longer relies on a rigid single threshold cutoff line, but instead introduces a comprehensive hierarchical judgment logic that includes hysteresis intervals. The program traverses the matrix elements and compares them with the system's preset three-level alarm boundaries (attention, severe, and critical). Specifically, when the calculated comprehensive degradation metric value is between 0.4 and 0.6, the system marks it as "mild degradation (attention recommended)," which usually corresponds to a slight PID (potential-induced degradation) effect in the early stages of the component; when the metric value is between 0.6 and 0.85, it is judged as "moderate degradation (requires periodic re-inspection)," indicating that the bypass diode may have intermittent conduction or microcracks; when the metric value is greater than 0.85, it is judged with high confidence as "severe structural damage," and a high-level maintenance alarm is immediately triggered. This hierarchical judgment based on continuously weighted metrics avoids frequent changes in alarm status caused by defects in a critical state during multiple inspections.

[0104] After completing the matrix traversal and defect classification of the entire site, the system encapsulates the diagnostic results into a standardized data payload. The output includes, but is not limited to: the precise topology address of the damaged component, the comprehensive degradation metric score, the physical attribution type of the defect (such as internal hotspots, diode breakdown, etc.), the on-site infrared absolute temperature slice, and repair suggestions. The structured alarm data is packaged and encapsulated in JSON or XML format and then pushed and parsed to the site's SCADA (Supervisory Control and Data Acquisition) central control center and the handheld terminals of maintenance personnel via an industrial Ethernet or 5G wireless network based on the TCP / IP protocol.

[0105] Ultimately, this correlation output matrix provides traceable and quantifiable data support for predictive maintenance of photovoltaic power plants, completely opening up a closed loop from aerial drone inspections to precise ground maintenance.

[0106] See attached document Figure 3 and attached Figure 4 To assist those skilled in the art in gaining a deeper understanding of the collaborative operation mechanism of the aforementioned functional modules in the real physical world, this embodiment uses the daily operation and maintenance inspection of a 50 MW mountain photovoltaic power station in a certain location as a specific application scenario. During a specific time period of the inspection, frequent movement of fragmented clouds occurred above the station, causing severe high-frequency fluctuations in ambient irradiance between 300 watts per square meter and 850 watts per square meter. Such meteorological conditions are typical interference conditions that easily lead to large-scale false alarms in traditional infrared thermal imaging diagnostic algorithms.

[0107] In the specific implementation process, the UAV platform, equipped with a dual-light pod, performed a close-range scan of the No. 12 inverter array at the site. Edge computing nodes extracted the absolute temperature tensor from the infrared video stream in real time. In a certain array area, the system detected a significant localized thermal anomaly, with a view-compensated absolute temperature reaching 68 degrees Celsius, while the background arithmetic mean temperature of surrounding normal components was only 42 degrees Celsius. If based on traditional static threshold diagnostic logic based on absolute temperature difference (e.g., a threshold set at 15 degrees Celsius), this area would be directly identified as a serious physical defect.

[0108] To verify the true physical cause of the thermal anomaly, the system extracted multi-source time-series data from the cloud-based control center, corresponding to a sliding window of 60 sampling points with a 1-second sampling interval. Based on the environmental meteorological sequence of the site, the system calculated the equivalent heat conduction time constant of the components under the current wind speed and ambient temperature to be 42.5 seconds. The program then performed phase hysteresis compensation on the equivalent electrical power derivative timing of the array, forcibly calibrating the physical time difference between the thermodynamic response and the electrical output. After completing the spatiotemporal alignment, the system calculated a zero-order Pearson correlation coefficient of 0.82 between the thermal and electrical derivatives.

[0109] However, the system did not directly accept this result, but further introduced ambient irradiance as a control variable. After partial correlation dimensionality reduction and projection calculation, the system found that the spatial thermal derivative and equivalent electrical power derivative of the high-heat region were highly correlated with ambient irradiance. After removing the common variation caused by irradiance, the system calculated that the true first-order partial correlation coefficient dropped sharply to 0.18. Combined with the significance t-test, the p-value of this partial correlation measure was greater than 0.05, failing the statistical validity check.

[0110] Based on the aforementioned multidimensional feature metrics, the system determined that there was no intrinsic causal relationship between the localized abnormal heating and the dynamic power decay of the inverter. Combining the false hotspot confidence score output by the ResNet residual network deployed at the edge, the system ultimately suppressed the comprehensive degradation metric value of this area to 0.12, clearly classifying it as a transient thermal shadow caused by external non-structural factors, successfully preventing a high-level false alarm caused by the superposition of complex meteorological conditions and surface foreign objects.

[0111] To further quantify and evaluate the technical effectiveness of the present invention in large-scale field applications, this embodiment conducted a comparative experiment on 2000 photovoltaic modules. The experiment acquired internal electroluminescence (EL) images and offline IV curves of all tested modules as ground truth calibration benchmarks. The experimental evaluation metrics adopted industry-standard precision and recall, and their specific calculation models are shown in the following code block: ; ; In the formula, Represents the accuracy of the diagnostic algorithm, used to measure the proportion of actual physical defects among all components that trigger alarms; Recall rate is used to measure the proportion of all physical defects that actually exist in the site that are successfully identified by the algorithm. This represents the number of true positive samples, i.e., the number of components for which the algorithm alarms and EL tests confirm the presence of defects; This represents the number of false positive samples, i.e., the number of false alarms triggered by the algorithm but with normal EL test results; This represents the number of false negative samples, i.e., the number of missed detections where the algorithm did not trigger an alarm but where there were actually hidden cracks or attenuation. The introduced constant for preventing division by zero is used to protect the validity of the underlying operations of the evaluation script when the algorithm fails in an extreme way (such as not outputting any positive results at all) and the denominator becomes zero.

[0112] In the comparative experiment, the system set up three control groups: the first group adopted the traditional static absolute temperature difference threshold model (baseline A); the second group adopted the multimodal baseline model that only integrates zero-order correlation and has no temporal phase compensation (baseline B); and the third group adopted the complete technical solution of the present invention (this embodiment solution) that includes dynamic hysteresis compensation, partial correlation stripping and neural network filtering.

[0113] Experimental results show that under clear, cloudless, steady-state conditions, the recall rates of all three schemes can reach above 0.90. However, under complex meteorological conditions with irradiance fluctuations greater than 20%, the technical performance shows significant divergence. Baseline A, unable to distinguish between cloud shadows and actual microcracks, generates a large number of false alarms, causing its precision to plummet to 0.42. Although baseline B incorporates electrical characteristics, it neglects the phase difference between thermodynamic and electrical responses in physical space, resulting in misalignment in its temporal covariance calculation and its inability to withstand common-mode meteorological interference, thus maintaining a precision of only 0.61.

[0114] In contrast, the solution presented in this embodiment, relying on its underlying time-series alignment logic and partial correlation calculation mechanism, exhibits higher robustness under complex meteorological conditions. While maintaining a recall rate of no less than 0.94, it stably keeps the precision within the range of 0.88 to 0.92. This diagnostic paradigm, based on deep decoupling of multi-source heterogeneous data, effectively filters out high-frequency noise caused by external environment and physical artifacts. This allows the output defect logical coordinates and correlation matrix measurements to be directly used as reliable evidence for on-site defect elimination, verifying the technical effectiveness of this solution for large-scale deployment in complex industrial environments.

Claims

1. A method for safety inspection and management of new energy power stations, characterized in that, Includes the following steps: The system acquires real-time electrical operating parameters of the photovoltaic array, and when an electrical abnormality is detected, it generates an inspection trigger event and extracts the corresponding abnormal string logic identifier. According to the preset electrical and physical topology mapping relationship, the abnormal string logical identifier is parsed into a target space coordinate set, and the inspection mobile terminal is dispatched to the target space coordinate set; After the inspection mobile terminal arrives at the target spatial coordinate set, it sends a transient bias command to the inverter corresponding to the target spatial coordinate set, so that the output current of the target photovoltaic string corresponding to the target spatial coordinate set undergoes a transient change within a preset time window, and simultaneously acquires the spatial pixel-level thermal image video stream and the actual continuous current sequence of the target photovoltaic string within the preset time window. Thermal timing features and electrical timing features corresponding to the actual continuous current sequence are extracted from the spatial pixel-level thermal image video stream, respectively. Based on the degree of electrothermal physical coupling between the thermal timing characteristics and the electrical timing characteristics, it is determined whether the target photovoltaic string has electrically induced high-resistivity hot spot defects, and an inspection and diagnosis report containing the spatial coordinates of the defects is generated based on the determination results.

2. The method for safety inspection and management of new energy power stations according to claim 1, characterized in that, The step of extracting the corresponding abnormal string logical identifier specifically includes: The real-time DC current value of each photovoltaic string under the same maximum power point tracking loop is obtained, and the dispersion of the real-time DC current value of each photovoltaic string from the system average operating current is calculated by the relative deviation algorithm. A string current deviation matrix is ​​constructed using the degree of dispersion corresponding to each photovoltaic string, and the time rate of change of the Frobenius norm of the string current deviation matrix within a set time window is extracted. When the time change rate is greater than the preset dynamic mutation threshold, the electrical operation abnormality is identified, the inspection trigger event is generated, and the logical identifier of the corresponding photovoltaic string that caused the norm mutation is used as the logical identifier of the abnormal string.

3. The method for safety inspection and management of new energy power stations according to claim 1, characterized in that, The steps of parsing the abnormal string logical identifier into a target spatial coordinate set and dispatching the inspection mobile terminal to the target spatial coordinate set specifically include: The mapping database storing the correspondence between the inverter logic ports and the polygon vertex coordinate sequences in the three-dimensional geographic coordinate system of the actual site is invoked as the electrical and physical topology mapping relationship; The mapping database is used to convert the abnormal string logical identifier into a corresponding three-dimensional space polygon coordinate set as the target space coordinate set. The system obtains the current GPS location coordinates of the inspection mobile terminal, generates the shortest unobstructed directional cruise trajectory from the GPS location coordinates to the target spatial coordinate set based on the site elevation obstacle map, and issues an execution command to fly along the shortest unobstructed directional cruise trajectory.

4. The method for safety inspection and management of new energy power stations according to claim 1, characterized in that, The step of causing a transient change in the output current of the target photovoltaic string within a preset time window specifically includes: Based on the voltage ramp control model, the transient bias command is generated and sent to the inverter, causing the inverter to suspend the conventional maximum power point optimization algorithm. The output current of the target photovoltaic string is made to decrease or increase continuously at a preset slope within a set thermal conduction relaxation time window. The thermal conduction relaxation time window is the preset time window, and the duration of the thermal conduction relaxation time window is greater than the longitudinal thermal conduction time constant from the internal conductive layer to the surface glass layer of the photovoltaic module.

5. The method for safety inspection and management of new energy power stations according to claim 1, characterized in that, The steps for simultaneously acquiring the spatial pixel-level thermal image video stream and the actual continuous current sequence of the target photovoltaic string specifically include: Within the preset time window, the infrared sensor mounted on the inspection mobile terminal is controlled to continuously capture images of the target photovoltaic string at a constant frame rate, thereby acquiring the spatial pixel-level thermal image video stream containing spatial pixel plane coordinates and time dimension; Based on the Network Time Protocol, the control monitoring and data acquisition system synchronously records the DC current data of the target photovoltaic string within the preset time window, as the actual continuous current sequence; This ensures that each frame of the spatial pixel-level thermal imaging video stream is aligned with the current sampling points in the actual continuous current sequence on the timestamp.

6. The method for safety inspection and management of new energy power stations according to claim 1, characterized in that, The steps of extracting the thermal time-series features from the spatial pixel-level thermal image video stream and the electrical time-series features corresponding to the actual continuous current sequence specifically include: For each pixel in the spatial pixel-level thermal imaging video stream, the partial derivative of the pixel temperature value with time is calculated as the thermal derivative of the pixel, and the thermal derivative of each pixel is used as the thermal time series feature. Based on the physical principle that heating power is proportional to the square of current in Joule's law, the partial derivative of the square of the current value at each moment in the actual continuous current sequence with respect to time is calculated as the power derivative at the corresponding moment, and the power derivative is used as the electrical time sequence feature.

7. The method for safety inspection and management of new energy power stations according to claim 1, characterized in that, The steps based on the degree of electrothermal-physical coupling correlation between the thermal time series characteristics and the electrical time series characteristics specifically include: Phase hysteresis compensation is performed on the power derivative corresponding to the electrical timing characteristics by introducing a preset heat conduction time constant. The thermal derivative at the current moment within the preset time window is paired with the electrical power derivative at a historical moment after shifting the heat conduction time constant forward from the current moment to complete the phase hysteresis compensation. Calculate the arithmetic mean of the thermal derivative within the effective time window and the integral mean of the electrical power derivative after phase hysteresis compensation, wherein the starting time of the effective time window is the time after the start time of the preset time window is shifted backward by the heat conduction time constant. Based on the difference between the thermal derivative and its corresponding arithmetic mean at each moment within the effective time window, and the difference between the electric power derivative after phase hysteresis compensation and its corresponding integral mean, the partial correlation coefficient between the two is calculated using the Pearson partial correlation metric method as the degree of electrothermal physical coupling.

8. The method for safety inspection and management of new energy power stations according to claim 1, characterized in that, The step of determining whether the target photovoltaic string has electrically induced high-resistivity hot spot defects specifically includes: Extract the set of high-temperature pixels whose absolute temperature features are higher than those of the surrounding area from the spatial pixel-level thermal imaging video stream; Traverse the set of high-temperature pixels. If the correlation coefficient corresponding to the electrothermal physical coupling degree of the target pixel is greater than the preset correlation coefficient threshold, then it is determined that the actual electro-induced high-resistivity hot spot defect occurs at the spatial location corresponding to the target pixel. If the correlation coefficient of the target pixel is less than or equal to the correlation coefficient threshold, it is determined that the temperature rise of the target pixel is not physically coupled with the change in internal heating power, and the target pixel is identified as a pseudo hot spot caused by environmental interference that is not electrically heated and is filtered out.

9. A method for safety inspection and management of new energy power stations according to claim 8, characterized in that, The safety inspection and management methods for new energy power stations also include: The correlation coefficients corresponding to all target pixels that are determined to have real electro-induced high-resistivity hot spot defects are constructed into a device topology correlation matrix; Based on the equipment topology association matrix, the specific physical space coordinates of the defects are extracted, the specific physical space coordinates are bound to the logical identifier of the target photovoltaic string, and the generated inspection and diagnosis report is pushed to the site operation and maintenance management system.

10. A safety inspection and management system for new energy power stations, executing the safety inspection and management method for new energy power stations as described in any one of claims 1-9, characterized in that, include: The event-driven triggering module is used to obtain the real-time electrical operating parameters of the photovoltaic array. When an electrical abnormality is detected, it generates an inspection trigger event and extracts the corresponding abnormal string logical identifier. The reverse parsing scheduling module is used to parse the abnormal string logical identifier into a target space coordinate set according to the preset electrical and physical topology mapping relationship, and to schedule the inspection mobile terminal to go to the target space coordinate set. The active bias and acquisition module is used to send a transient bias command to the inverter corresponding to the target spatial coordinate set after the inspection mobile terminal arrives at the target spatial coordinate set, so that the output current of the target photovoltaic string undergoes a transient change within a preset time window, and simultaneously acquires the spatial pixel-level thermal image video stream and the actual continuous current sequence of the target photovoltaic string within the preset time window. The phase compensation calculation module is used to extract the thermal quotient of temperature versus time for each pixel in the spatial pixel-level thermal image video stream, and the electrical power quotient of internal heating power versus time corresponding to the actual continuous current sequence. A preset heat conduction time constant is introduced to perform phase hysteresis compensation on the electrical power quotient, and the correlation coefficient between the thermal quotient and the electrical power quotient after phase hysteresis compensation is calculated. The feature extraction module is used to extract the thermal timing features in the spatial pixel-level thermal image video stream and the electrical timing features corresponding to the actual continuous current sequence, respectively. The defect identification and judgment module is used to determine whether the target photovoltaic string has electrically induced high-resistivity hot spot defects based on the degree of electrothermal physical coupling between the thermal time series characteristics and the electrical time series characteristics, and to generate an inspection and diagnosis report containing the spatial coordinates of the defects based on the judgment results.

Images