A photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the internet of things

Abstract

The application discloses a photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with an Internet of Things, relates to the photovoltaic electronic fault diagnosis and remote operation and maintenance field, and sequentially connects a global perception acquisition unit, an edge front diagnosis unit, a cloud fusion decision unit, an operation and maintenance closed loop execution unit and a hierarchical authorization multi-terminal interaction unit through a communication link suitable for multiple types of heterogeneous Internet of Things protocols; the global perception acquisition unit is distributedly arranged at photovoltaic full-link nodes, acquires equipment electric parameters, environmental parameters and working condition data, and completes preprocessing; the edge front diagnosis unit is arranged at a local station, loads a lightweight model, completes millisecond-level initial diagnosis of an explicit hard fault, triggers local alarm, and uploads suspected hidden fault features to the cloud. The application adopts a cloud-edge collaborative hierarchical diagnosis architecture, greatly reduces transmission bandwidth and cloud computing load, can independently complete local alarm and data caching at the edge side in a network interruption state, improves emergency hard fault response speed, and avoids safety accidents.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic electronic fault diagnosis and remote operation and maintenance technology, and in particular to a photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things. Background Technology

[0002] As a core component of clean energy, photovoltaic (PV) has seen continuous and rapid growth in both centralized and distributed PV installations in recent years. These sites are mostly located in dispersed areas such as mountains, deserts, industrial and commercial rooftops, and residential rooftops. The sheer number and dispersed deployment of electronic equipment such as PV modules, inverters, and combiner boxes make operation and maintenance coverage challenging. The industry's core needs are focused on improving the accuracy of fault identification, shortening fault response time, reducing overall operation and maintenance costs, and ensuring the power generation efficiency of PV sites.

[0003] Currently, small-scale distributed photovoltaic (PV) power plants generally adopt the traditional manual inspection and maintenance (O&M) model. The working principle involves O&M personnel conducting on-site inspections and recording equipment operating parameters at fixed intervals. When a fault occurs, the site management notifies the O&M personnel to investigate and resolve the issue. Fault diagnosis and repair rely on the practical experience of the O&M personnel. This model requires no additional investment in information systems and has a low deployment threshold. However, it has significant drawbacks: limited inspection coverage density, early latent faults such as module hotspots and potential-induced degradation cannot be detected in a timely manner, resulting in a high rate of missed faults; response time after a hard fault occurs can be delayed by several hours to several days, leading to significant power generation losses; O&M quality is highly dependent on the experience level of personnel, with a high rate of misjudgment by newcomers; and the labor and transportation costs for inspections at remote sites are extremely high, making it difficult to adapt to the needs of large-scale PV O&M.

[0004] Currently, most large and medium-sized photovoltaic (PV) power plants adopt cloud-based centralized PV operation and maintenance (O&M) systems. These systems work by deploying sensors on the equipment to collect all operational data, which is then uploaded to a cloud platform. Fault identification is achieved through fixed threshold judgments or simple machine learning models, enabling remote centralized monitoring of equipment operating status and significantly reducing on-site workload compared to manual inspections. However, several shortcomings remain: Uploading all data requires high network bandwidth; data loss and delays are common at remote sites with poor network stability, leading to a significant decrease in diagnostic accuracy. All diagnostic logic is executed in the cloud, resulting in fault response latency of seconds or even minutes, making it impossible to provide rapid local alarms for urgent hard faults such as short circuits and overheating, potentially causing equipment burnout or fires. Diagnosis relies solely on single-device operational data, failing to achieve multi-source feature cross-fusion, resulting in low accuracy in identifying weak-feature hidden faults and significant missed detections. Furthermore, the system does not streamline the entire O&M process; after fault diagnosis, manual work is still required for dispatching orders, spare parts scheduling, and fault verification, leading to numerous process interruptions and limited improvement in O&M efficiency. Summary of the Invention

[0005] This invention proposes a photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things, in order to solve the problems mentioned in the prior art, such as low accuracy of existing photovoltaic operation and maintenance fault identification, high response latency, non-closed-loop operation and maintenance process, and poor scenario adaptability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things, comprising a global perception acquisition unit, an edge pre-diagnosis unit, a cloud fusion decision unit, an operation and maintenance closed-loop execution unit, and a hierarchical authorization multi-terminal interaction unit, which are sequentially coupled through communication links that support adaptation to multiple heterogeneous Internet of Things protocols. The full-domain perception and acquisition unit is distributed across all nodes of the photovoltaic electronic product chain, covering the installation points of photovoltaic modules, string inverters, combiner boxes, grid-connected cabinets, and energy storage modules. It collects node operating electrical parameters, environmental parameters, and operating status data according to dynamically configured differentiated sampling frequencies. After performing clock drift compensation, bad value removal, and timestamp alignment preprocessing on the collected data, it transmits the data to the edge pre-diagnostic unit through an encrypted breakpoint resume transmission channel. The edge pre-diagnostic unit is deployed on the local industrial control computer of the site. It loads a lightweight fault identification model that has been knowledge distilled and compressed to perform millisecond-level preliminary diagnosis on real-time streaming data. When a visible hard fault is identified, a local hierarchical alarm is triggered. At the same time, the associated feature fragments of suspected latent faults are filtered according to the preset confidence threshold and uploaded to the cloud fusion decision unit along with the preliminary diagnosis confidence results. The cloud-integrated decision-making unit is deployed on a hybrid cloud architecture platform. It performs cross-dimensional multi-source feature fusion on the received edge data, loads a temporal convolutional deep diagnostic model with an attention mechanism to accurately locate faults, quantifies the fault risk level, clarifies the causes and impact boundaries, and generates differentiated operation and maintenance handling strategies that match the fault level. The operation and maintenance closed-loop execution unit connects with the site spare parts storage, personnel positioning, and work order scheduling systems through data interfaces. It automatically generates dispatch orders and spare parts scheduling instructions based on the operation and maintenance strategy, tracks operation and maintenance nodes throughout the entire process, and conducts automated verification after troubleshooting, forming a complete closed-loop link of diagnosis-dispatch-handling-verification. The hierarchical authorization multi-terminal interaction unit provides differentiated operation entry points for operation and maintenance personnel, site administrators, and equipment manufacturer personnel, enabling real-time synchronization and two-way interaction of alarms, work order status, and operation data.

[0007] Furthermore, the comprehensive sensing and acquisition unit includes a high-precision electrical parameter acquisition module, a multi-dimensional environmental parameter acquisition module, an equipment condition acquisition module, and an edge preprocessing module. The high-precision electrical parameter acquisition module has a sampling accuracy of no less than 0.5 level, covering the open-circuit voltage, operating current, and maximum power point power of photovoltaic strings; the DC-side input voltage and current, AC-side output voltage and current, conversion efficiency, and total harmonic distortion of the inverter; the current of each branch of the combiner box and the bus voltage; and the charging and discharging voltage, charging and discharging current, individual cell temperature, and state of charge value of the energy storage module. The multi-dimensional environmental parameter acquisition module acquires data on the irradiance of the tilted surface of the photovoltaic panel, the ambient dry-bulb temperature, the panel backplane temperature, the near-surface wind speed of the site, rainfall indicators, and snow thickness identification data. The equipment condition acquisition module acquires data on the DC switch status, AC circuit breaker status, cooling fan speed, surge protector status, and communication link signal strength. The edge preprocessing module first compensates for the clock drift of each acquisition node to achieve accurate timestamp alignment. The clock drift compensation alignment formula used is: ; in, For the first The data collected by each acquisition node is aligned with the compensated timestamps in milliseconds. For the first The collection timestamp recorded by the local clock of each collection node, in milliseconds; This is the time deviation value between the acquisition node and the edge gateway during the most recent clock synchronization, in milliseconds. The crystal oscillator drift coefficient of the acquisition node is obtained by linear fitting of the deviation values ​​of three consecutive clock synchronizations; The reference time for the most recent clock synchronization between the data acquisition node and the edge gateway, in milliseconds; This compensation method fully considers the cumulative time error caused by the inherent drift of the crystal oscillators of the distributed acquisition nodes, which can control the time alignment error of multi-node data to an extremely low level, providing an accurate data foundation for subsequent time-series matching of fault features and avoiding errors in fault feature extraction caused by time-series misalignment. After completing the timestamp alignment, the preprocessing module uses the improved 3σ criterion to remove outliers and uses the linear interpolation method of adjacent sampling points to fill in short-term missing data. During the removal and filling process, the original data identifier is retained to avoid the preprocessing operation from changing the original fault features.

[0008] Furthermore, the lightweight fault identification model loaded by the edge pre-diagnostic unit is a lightweight convolutional neural network model based on residual structure and compressed through channel pruning and knowledge distillation. The number of model parameters is adapted to the storage and computing capabilities of the edge industrial control computer, and the single inference latency meets the millisecond-level real-time response requirements. The model training phase uses the explicit hard fault sample set accumulated by the field station to complete the training. The types of explicit hard faults identified include five types of faults: photovoltaic module positive and negative short circuit, module output open circuit, inverter over-temperature triggered protection shutdown, combiner box branch fuse blown and current interruption, and complete communication link interruption. The edge pre-diagnostic unit outputs a preliminary diagnostic result for each piece of real-time collected data, and at the same time outputs the edge confidence value of the diagnostic result. When the edge confidence value is higher than 0.95, it is determined to be a confirmed explicit hard fault, which directly triggers on-site audible and visual alarms and marks it as an emergency maintenance level. It can be pushed to the maintenance closed-loop execution unit without waiting for secondary diagnosis in the cloud, thus shortening the response latency of high-risk faults. When the edge confidence value is between 0.6 and 0.95, it is determined that there is a suspected hidden fault risk. The time-series feature segments before and after the fault occurrence are automatically extracted and packaged together with the edge confidence value and point identification information and uploaded to the cloud fusion decision unit. When the edge confidence value is below 0.6, it is determined that there is a data collection disturbance. It is directly marked as normal data and no alarm or upload process is triggered, thereby reducing the computing load and transmission bandwidth usage in the cloud.

[0009] Furthermore, after receiving suspected fault feature fragments uploaded from the edge side, the cloud-based fusion decision unit first performs multi-source feature dimension alignment and fusion. Specifically, it unifies the dimensions of the fault feature fragments collected in real time, the historical normal operation benchmark features of equipment at the same location, the real-time operation features of adjacent strings or equipment of the same specification, the weather forecast data for future periods of the station, and the equipment's full life cycle operation and maintenance record data. This constructs a third-order fault feature tensor containing three dimensions: time step, feature category, and associated equipment, thereby eliminating the interference of single-source data collection bias and environmental disturbances on the diagnostic results. After feature fusion is completed, the cloud-based deep diagnostic model outputs the cloud-based diagnostic results and the corresponding cloud-based confidence scores. The system uses a cloud-edge collaborative confidence-weighted fusion formula to calculate the final diagnostic confidence score, which serves as the basis for fault level determination. The cloud-edge collaborative confidence-weighted fusion formula used is as follows: ;in, This represents the final confidence level of the fault diagnosis result, with a value ranging from 0 to 1. The confidence level of the diagnostic results output by the edge pre-diagnostic unit, with a value ranging from 0 to 1; The confidence level of the diagnostic results output by the cloud-integrated decision unit, with a value ranging from 0 to 1; The weighted value of the edge-side diagnostic results is dynamically adjusted based on the diagnostic accuracy of the edge model at the corresponding site over the past 7 days. The weighting value for the cloud-based diagnostic results is dynamically adjusted based on the global diagnostic accuracy of the cloud model over the past 30 days, and meets the following requirements: and The sum of is 1; This fusion approach eliminates the drawbacks of traditional cloud-edge collaborative diagnosis which uses fixed weights. It can dynamically adjust the weight allocation based on the actual performance of the model, while taking into account the low latency advantage of edge diagnosis and the high precision advantage of cloud diagnosis. It reduces the transmission bandwidth consumption and cloud computing load, while improving the overall fault diagnosis accuracy. When the final confidence level is higher than the preset threshold, the fault diagnosis result is deemed valid and the subsequent operation and maintenance strategy generation process begins. When the final confidence level is lower than the preset threshold, the remote expert review process is automatically triggered to avoid misdiagnosis and missed diagnosis.

[0010] Furthermore, the deep diagnostic model loaded by the cloud-integrated decision unit is a two-layer attention temporal convolutional network model that introduces channel attention mechanism and time attention mechanism. The model automatically assigns weights to different feature dimensions through the channel attention module to improve the feature sensitivity of weak feature latent faults such as hot spots, potential-induced decay, and loose connection of wiring terminals. The time attention module automatically locates the time segment with the highest correlation with the fault in the time series data, reducing the interference of irrelevant time series segments on the diagnostic results. The model output includes the precise location of the fault, the classification of fault causes, the scope of fault spread, and the quantitative value of potential power generation loss. The types of latent faults covered include six early latent faults: hot spots on photovoltaic modules, potential-induced degradation, aging of encapsulation materials leading to decreased light transmittance, abnormal degradation of inverter conversion efficiency, abnormal degradation of energy storage module capacity, and abnormal temperature rise caused by excessive contact resistance of wiring terminals. These faults are difficult to identify by the lightweight model at the edge because their characteristics are not obvious. By integrating multi-source data in the cloud, accurate identification of the early stage of the fault can be achieved, preventing the fault from further expanding and causing equipment damage and power generation loss.

[0011] Furthermore, when the cloud-integrated decision-making unit generates operation and maintenance handling strategies, it automatically calls the structured storage of historical fault case libraries. It performs cosine similarity matching between the feature vector of the current fault and the feature vector of historical faults in the case library. When the matching degree reaches a preset threshold, it directly reuses the standardized handling process, on-site safety operation specifications, required spare parts models and quantities, and fault investigation priority information of the corresponding historical case. When the matching degree is lower than the threshold, it automatically sends a remote assistance request to the manufacturer's technical personnel of the corresponding equipment model. The manufacturer's technical personnel formulate targeted handling strategies based on the fault data. After the strategy is reviewed and approved, it is synchronously stored in the historical fault case library, realizing the dynamic iterative update of the case library. As the system runs for a longer period of time, the fault scenarios covered by the case library continue to enrich, and the accuracy of strategy generation continues to improve.

[0012] Furthermore, the operation and maintenance closed-loop execution unit includes an intelligent dispatch module, an intelligent spare parts scheduling module, a full-process tracking module, and an automated troubleshooting verification module. The intelligent dispatch module calculates the comprehensive suitability of each maintenance personnel based on four dimensions: the real-time location of the maintenance personnel, the current workload of pending work orders, the completion rate of handling similar faults in the past, and the matching degree between their qualifications and the fault type. The personnel with the highest comprehensive suitability are selected as the work order recipient. The formula for calculating the comprehensive suitability of maintenance personnel is as follows: ;in, The overall adaptability score for operations and maintenance personnel to take on the current work order is given, with a value ranging from 0 to 1. The higher the score, the stronger the adaptability. The straight-line distance between the current real-time location of the maintenance personnel and the fault location, in kilometers; The maximum straight-line distance between the dispatchable maintenance personnel and the fault location within the station area, in kilometers; The estimated remaining processing time for the current pending work orders of the operations and maintenance personnel, in minutes; The maximum estimated remaining processing time for pending work orders among all available maintenance personnel, in minutes; The satisfaction rate of maintenance personnel in handling similar faults in the past, with a value ranging from 0 to 1; The degree of matching between the qualifications held by the maintenance personnel and the current fault handling requirements is 1 if the requirements are fully met, 0.5 if the requirements are partially met, and 0 if the requirements are not met. , , , The weighting coefficients for distance, load, historical performance, and qualification dimensions are configured by the site management personnel according to actual operation and maintenance needs, and the sum of the four weighting coefficients must be 1. This quantification method eliminates the dimensional differences of different indicators through multi-dimensional normalization calculations, avoids the subjectivity and arbitrariness of manual dispatching, achieves optimal personnel matching for operation and maintenance tasks, shortens fault response time, and improves operation and maintenance efficiency; the intelligent spare parts scheduling module automatically matches the corresponding spare parts model and quantity in the site spare parts warehouse according to the fault type. When the inventory meets the demand, it automatically locks the corresponding spare parts and sends a requisition reminder. When the inventory is insufficient, it automatically triggers the requisition process from the nearest regional storage point; the full-process tracking module synchronizes the full-node information of operation and maintenance personnel's on-site check-in, fault investigation, and handling operations in real time. After maintenance personnel submit a defect elimination completion application, the automated defect elimination verification module automatically retrieves the continuous preset duration of operation data for the corresponding location and compares it with the normal operation benchmark value during the same period before the fault occurred. If the deviation is within the normal threshold range, the defect elimination is confirmed to be completed and the work order is automatically closed; otherwise, the defect elimination application is rejected and reprocessing is required.

[0013] Furthermore, the hierarchical authorization multi-terminal interaction unit adopts a role-based access control mechanism to achieve hierarchical permission management. It grants site management personnel the right to view all operational data, adjust operation and maintenance strategy parameters, configure system sampling frequency, and allocate permissions. It also supports viewing statistical reports such as power generation data, fault statistics, and operation and maintenance completion rate of the entire site. For front-line maintenance personnel, the system grants them access to receive fault alarms in their assigned area, update work order status, transmit audio and video footage of on-site troubleshooting, and apply for spare parts. They can only view the equipment data and handling requirements associated with the work orders they have undertaken. For technical personnel from equipment manufacturers, access is granted to iterative updates of fault diagnosis models for corresponding equipment models and remote troubleshooting guidance for difficult faults. They can only view desensitized operating data of the equipment models they are responsible for and cannot access sensitive operating information of other equipment in the plant. All operations of all roles are recorded in operation logs, including the operator, operation time, operation content, and parameter changes before and after the operation. The logs are retained to meet the traceability requirements of power system safety management.

[0014] Furthermore, an encrypted incremental model synchronization link is established between the cloud-based fusion decision unit and the edge-based pre-diagnostic unit. After each iteration and optimization of the deep diagnostic model based on newly added fault samples, the cloud automatically generates a lightweight incremental update package adapted to the corresponding site. The update package uses differential compression to reduce the transmission volume and is sent to the edge-based pre-diagnostic unit of the corresponding site through a breakpoint resume mechanism. The model update process adopts a dual-caching mechanism, that is, the old version model continues to run to provide diagnostic services, while the new version model completes loading and local test set accuracy verification in the background. When the diagnostic accuracy of the new version model on the local test set is not lower than that of the old version model, the system automatically and smoothly switches to the new version model without service interruption. If the accuracy of the new version model does not meet the requirements, the system automatically rolls back to the old version model to avoid diagnostic service interruption caused by model updates.

[0015] Furthermore, the cloud-integrated decision-making unit also integrates a predictive operation and maintenance module. Based on continuously collected long-term operating data of the equipment, it predicts the amount of performance degradation of the equipment within a preset period by fitting the equipment performance degradation trend curve. When the amount of equipment performance degradation reaches the preset warning threshold, predictive operation and maintenance tasks are automatically generated, and operation and maintenance personnel are arranged in advance to carry out preventive inspections and cleaning maintenance during low power generation periods. This changes the fault handling mode from post-repair emergency repair to pre-prevention, reducing power generation losses caused by fault downtime. The predictive maintenance module can also combine historical fault occurrence patterns and meteorological disaster early warning information to automatically generate routine inspection plans for station equipment, dynamically adjust inspection points and inspection frequency, increase the inspection frequency for points with high fault risk, and appropriately reduce the inspection frequency for points with stable operating status, thereby improving the targeting of inspections while reducing the amount of ineffective inspection work for maintenance personnel.

[0016] Compared with existing technologies, the beneficial effects of this invention are: This invention adopts a cloud-edge collaborative hierarchical diagnostic architecture, deploying a lightweight fault identification model on the edge side to achieve extremely fast local alarms for explicit hard faults. It only uploads feature fragments of suspected latent faults instead of the full amount of operational data, effectively solving the problems of high bandwidth consumption and delayed or even failed fault response when the network is unstable in existing centralized cloud operation and maintenance systems. In the case of network outage, the edge side can independently complete local alarms and data caching, avoiding security incidents caused by untimely handling of emergency hard faults, while significantly reducing transmission resource consumption and cloud computing load.

[0017] This invention employs a deep diagnostic logic based on multi-source feature fusion to accurately compensate for clock drift in distributed acquisition nodes, achieving time-series alignment of multi-source data. It introduces a cloud-edge diagnostic confidence fusion mechanism with dynamic weights to enhance the ability to capture weak features of early latent faults. This effectively solves the problems of large diagnostic bias in single-source data, high missed detection rate of early latent faults, and high dependence of diagnostic accuracy on the experience of maintenance personnel in existing operation and maintenance systems. It significantly improves the identification accuracy of all types of faults and reduces misdiagnosis and missed diagnosis.

[0018] This invention streamlines the entire operation and maintenance process, from diagnosis and dispatch to spare parts scheduling and defect verification. It employs a multi-dimensional quantitative algorithm for the suitability of operation and maintenance personnel to achieve intelligent dispatch, and is complemented by a standardized historical fault case library for reuse and a hierarchical authorization multi-terminal interaction mechanism. This effectively solves the problems of multiple breakpoints in traditional operation and maintenance processes, strong subjectivity in manual dispatch, non-standard defect verification, and delayed information transmission. It achieves automated closed-loop management of the entire operation and maintenance process, reduces operation and maintenance labor costs, and improves overall handling efficiency. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things proposed in this invention; Figure 2 This is a flowchart of the edge sensing acquisition and pre-diagnosis process of the present invention; Figure 3 This is a flowchart of the cloud-based multi-source feature fusion and deep diagnosis process of this invention; Figure 4 This is a flowchart illustrating the generation and closed-loop execution of the operation and maintenance strategy of this invention. Figure 5 This is a flowchart of the multi-terminal collaborative interaction and model iterative update process of the present invention. Detailed Implementation

[0020] 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.

[0021] Reference Figures 1 to 5This invention discloses a photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things (IoT). It includes a global sensing and acquisition unit, an edge pre-diagnosis unit, a cloud-based fusion decision-making unit, an operation and maintenance closed-loop execution unit, and a hierarchical authorization multi-terminal interaction unit, sequentially coupled through communication links supporting adaptation to multiple heterogeneous IoT protocols. The communication links have a built-in protocol conversion gateway, pre-configured with proprietary protocol parsing plugins from mainstream photovoltaic equipment manufacturers and standard industrial protocol parsing rules. It supports automatic conversion of multiple protocols such as MQTT, CoAP, LoRaWAN, and Modbus-RTU, achieving data interoperability without requiring customized driver development for individual devices. This solves the problems of high equipment adaptation costs and severe data silos in traditional photovoltaic operation and maintenance systems. The link layer uses TLS 1.3 encrypted transmission, and all access devices use a one-device-one-key certificate authentication mechanism. Devices without valid certificates cannot access the network, thus meeting the security protection requirements of the power monitoring system.

[0022] The full-domain sensing and acquisition unit is distributed across all nodes of the photovoltaic electronic product chain, covering the installation points of photovoltaic modules, string inverters, combiner boxes, grid-connected cabinets, and energy storage modules. It uses dynamically configured differentiated sampling frequencies to collect data. Under normal operating conditions, the sampling frequency for electrical parameters is 1 time / minute, and for environmental parameters it is 1 time / 5 minutes. When parameter fluctuations are detected to exceed the normal threshold, the sampling frequency is automatically increased to 1 time / second, balancing the integrity of fault feature acquisition with system load control. After preprocessing such as clock drift compensation, bad value removal, and timestamp alignment, the collected data is transmitted to the edge front-end diagnostic unit through an encrypted channel that supports breakpoint resume. It is equipped with 128G industrial-grade storage locally, which can store the collected data for up to 30 days when the network is down. The data will be automatically resumed after the network is restored, solving the problem of data loss caused by unstable network in remote sites.

[0023] The edge-based pre-diagnostic unit is deployed on a local industrial-grade control computer at the site that is resistant to high and low temperatures and electromagnetic interference. It is equipped with a quad-core ARM processor and 8GB of RAM. It loads a lightweight fault identification model that has been knowledge distilled and compressed to perform millisecond-level preliminary diagnosis on real-time streaming data. When a visible hard fault is identified, it triggers a local hierarchical alarm. At the same time, it filters the associated feature fragments of suspected latent faults according to a preset confidence threshold and uploads them, along with the preliminary diagnosis confidence results, to the cloud-based fusion decision unit. It can independently complete the local alarm function in the absence of network, avoiding the problem of pure cloud architecture failing when the network is down.

[0024] The cloud-integrated decision-making unit is deployed on a hybrid cloud architecture platform, with sensitive operational data stored on the site's private cloud nodes. Model training and big data computing are scheduled using public cloud elastic computing power, balancing data security and computing power support capabilities. The platform performs cross-dimensional multi-source feature fusion on the received edge data, loads a temporal convolutional deep diagnostic model with an attention mechanism to accurately locate faults, quantifies fault risk levels, clarifies causes and impact boundaries, and generates differentiated operation and maintenance handling strategies that match the fault level.

[0025] The operation and maintenance closed-loop execution unit establishes standardized data interfaces with the site spare parts storage ERP system, personnel UWB positioning system, and work order OA system. It automatically generates dispatch orders and spare parts scheduling instructions based on the operation and maintenance strategy, tracks operation and maintenance nodes throughout the entire process, and conducts automated verification after troubleshooting. This forms a complete closed-loop link of diagnosis-dispatch-handling-verification, solving the problems of many breakpoints and low efficiency of manual follow-up in traditional operation and maintenance processes.

[0026] The hierarchical authorization multi-terminal interaction unit covers the Web management backend, the operation and maintenance mobile APP, the on-site audible and visual alarm screen, and the manufacturer's remote diagnostic port. It provides differentiated operation entry points for operation and maintenance personnel, site administrators, and equipment manufacturer personnel, and realizes real-time synchronization and two-way interaction of alarms, work order status, and operation data, solving the problems of lagging information transmission and unclear responsibility boundaries in traditional operation and maintenance.

[0027] This invention also discloses a global sensing and acquisition unit comprising a high-precision electrical parameter acquisition module, a multi-dimensional environmental parameter acquisition module, an equipment operating condition acquisition module, and an edge preprocessing module. The high-precision electrical parameter acquisition module employs a 0.2S-level current transformer and a high-precision voltage divider circuit, with a sampling error not exceeding 0.2%. The acquisition range covers the open-circuit voltage, operating current, and maximum power point power of photovoltaic strings; the DC-side input voltage and current, AC-side output voltage and current, conversion efficiency, and total harmonic distortion of the inverter; the current of each branch of the combiner box and the bus voltage; and the charging and discharging voltage, charging and discharging current, individual cell temperature, and state of charge value of the energy storage module. In the multi-dimensional environmental parameter acquisition module, the irradiance sensor is installed at the same tilt angle as the photovoltaic panel to avoid irradiance measurement deviation caused by horizontal installation. It can collect data such as irradiance of the tilted surface of the photovoltaic panel, ambient dry bulb temperature, panel back plate temperature, near-ground wind speed at the site, rainfall indicators, and snow thickness identification. The equipment status acquisition module connects to equipment status contacts via hard nodes to collect data on the status of DC switches, AC circuit breakers, cooling fan speeds, surge protectors, and communication link signal strength. To achieve accurate time alignment of data from multiple nodes, the system dynamically compensates for clock drift at each acquisition node, ensuring timestamp alignment satisfies the following relationship: ;in, For the first The data collected by each acquisition node is aligned with the compensated timestamps in milliseconds. For the first The collection timestamp recorded by the local clock of each collection node, in milliseconds; This is the time deviation value between the acquisition node and the edge gateway during the most recent clock synchronization, in milliseconds. The crystal oscillator drift coefficient of the acquisition node is obtained by linear fitting of the deviation values ​​of three consecutive clock synchronizations; The reference time, in milliseconds, is the most recent clock synchronization time between the acquisition node and the edge gateway. Through the above compensation logic, the system can fully consider the cumulative time error caused by the inherent drift of the distributed acquisition node's crystal oscillator, controlling the time alignment error of multi-node data to within 1 millisecond. This provides an accurate data foundation for subsequent timing matching of fault features and avoids errors in fault feature extraction due to timing misalignment. After the edge preprocessing module completes timestamp alignment, it uses a sliding window improved 3σ criterion to remove outliers. The window size is set to 1 hour, identifying outliers only within the same operating condition range, avoiding the global threshold from misjudging normal power fluctuations at sunrise and sunset as outliers. For data with a missing duration of less than 1 minute, linear interpolation of adjacent sampling points is used to fill in the missing data. Data with a missing duration of more than 1 minute is marked as invalid and is not filled in by interpolation to avoid introducing false features. All operations during the preprocessing process retain the original data identifier and do not modify the original collected values ​​to ensure the authenticity of the fault features.

[0028] This invention also discloses a lightweight fault identification model loaded by an edge pre-diagnostic unit, which is a lightweight convolutional neural network model based on residual structure and subjected to channel pruning and knowledge distillation compression.

[0029] The model training dataset is derived from the operational fault data of 217 grid-connected photovoltaic power stations across the country accumulated over 6 years. It contains 1.2 million explicit hard fault samples verified on-site. The dataset is divided into training set, validation set and test set in a ratio of 8:1:1. The training process uses the Adam optimizer with an initial learning rate of 0.001 and a cosine annealing strategy to adjust the learning rate. The knowledge distillation process uses a high-precision temporal convolutional network trained in the cloud as the teacher model and an 8-layer lightweight convolutional neural network with channel pruning as the student model. The distillation temperature is set to 4. The student model is trained by soft labels output by the teacher model. Finally, the number of parameters of the student model is compressed to 7.8MB. The explicit fault identification accuracy on the test set reaches 96.2%, and the single inference latency does not exceed 20 milliseconds, which is fully compatible with the computing power conditions of edge industrial control computers.

[0030] The model can identify five types of explicit hard faults: short circuit between the positive and negative electrodes of the photovoltaic module, open circuit at the module output, inverter over-temperature triggering protection shutdown, blown fuse in the combiner box branch circuit, and complete interruption of communication link. The edge pre-diagnostic unit outputs preliminary diagnostic results for each piece of real-time acquired data, along with the edge confidence value of the diagnostic results. When the edge confidence value is higher than 0.95, it is determined to be a confirmed explicit hard fault, directly triggering on-site audible and visual alarms and marking it as an emergency maintenance level. It can be pushed to the maintenance closed-loop execution unit without waiting for secondary diagnosis in the cloud, reducing the response latency of high-risk faults to milliseconds and preventing short-circuit faults from escalating and burning out equipment. When the edge confidence value is between 0.6 and 0.95, it is determined that there is a suspected hidden fault risk. The time-series feature segments before and after the fault occurs are automatically extracted and packaged together with the edge confidence value and point identification information and uploaded to the cloud fusion decision unit. Uploading only the feature segments can reduce the transmission bandwidth usage to 5% of the full data upload, which greatly reduces the network load. When the edge confidence value is below 0.6, it is determined that the data has acquisition disturbances such as bird shadows or transient electromagnetic interference, and is directly marked as normal data without triggering alarms or uploading processes, thus reducing the interference of invalid alarms on operation and maintenance personnel.

[0031] This invention also discloses that after the cloud-based fusion decision unit receives suspected fault feature fragments uploaded from the edge side, it first performs multi-source feature dimension alignment and fusion. Specifically, it unifies the dimensions of real-time collected fault feature fragments, historical normal operation benchmark features of equipment at the same location, real-time operation features of adjacent strings or equipment of the same specification, weather forecast data for future periods of the station, and equipment life cycle operation and maintenance record data. It constructs a third-order fault feature tensor with three dimensions: time step, feature category, and associated equipment. The time step dimension is 60 time points with a 1-second interval. The feature category dimension covers 28 features including electrical parameters, environmental parameters, operating conditions, historical benchmarks, and weather forecasts. The associated equipment dimension covers the remaining 15 strings under the same combiner box or other equipment of the same model in the same compartment. The overall tensor dimension is 60×28×16. Multi-source data cross-validation eliminates the interference of single-source data acquisition bias and environmental disturbances on the diagnostic results.

[0032] To balance the low latency advantage of edge diagnostics with the high accuracy advantage of cloud diagnostics, the system uses a dynamic weighting method to fuse the diagnostic results from both sides. The final diagnostic confidence satisfies the following relationship: ;in, This represents the final confidence level of the fault diagnosis result, with a value ranging from 0 to 1. The confidence level of the diagnostic results output by the edge pre-diagnostic unit, with a value ranging from 0 to 1; The confidence level of the diagnostic results output by the cloud-integrated decision unit, with a value ranging from 0 to 1; The weighted value of the edge-side diagnostic results is dynamically adjusted based on the diagnostic accuracy of the edge model at the corresponding site over the past 7 days. The weight value is the diagnostic result on the cloud side, which is dynamically adjusted based on the global diagnostic accuracy of the cloud model over the past 30 days, and the sum of the two weights is 1.

[0033] By using dynamic weight allocation logic, the system avoids the drawbacks of using fixed weights in traditional cloud-edge collaborative diagnosis. It can continuously optimize weight allocation based on the actual performance of the model, thereby reducing transmission bandwidth consumption and cloud computing load while improving the overall fault diagnosis accuracy.

[0034] The system sets the final confidence threshold to 0.85. When the final confidence score is higher than the threshold, the fault diagnosis result is deemed valid and proceeds to the subsequent operation and maintenance strategy generation process. When the final confidence score is lower than the threshold, the remote expert review process is automatically triggered, pushing the fault data to the port of the corresponding equipment manufacturer's technical personnel for manual review and judgment to avoid misdiagnosis and missed diagnosis. The labeled samples after expert review are automatically stored in the model training dataset to provide data support for subsequent model iterations.

[0035] This invention also discloses a deep diagnostic model loaded by a cloud-based fusion decision unit, which is a two-layer attentional temporal convolutional network model that incorporates channel attention and temporal attention mechanisms. The model employs four dilated convolutional layers with dilation coefficients set to 1, 2, 4, and 8, respectively, and a kernel size of 3. The overall receptive field covers 60 time steps, enabling it to capture temporal change features within a 1-minute timeframe. The channel attention module adopts a compression-excitation structure. First, global average pooling is performed on each feature channel, and then the weight of each channel is learned through two fully connected layers. The weight ratio of feature channels with high correlation to faults is increased, and the interference of irrelevant features is reduced. The temporal attention module employs a scaled dot product self-attention structure to learn the feature weights at each time step, strengthening the feature contribution at the moment of fault occurrence and suppressing redundant features during normal periods. The model training set contains 3.6 million verified latent fault samples, also divided in an 8:1:1 ratio. Training uses the AdamW optimizer with an initial learning rate of 0.0005, a weight decay coefficient of 0.01, and a dropout ratio of 0.2 to prevent overfitting. The model achieves a latent fault identification accuracy of 93.7% on the test set.

[0036] The model output includes the precise location of the fault, the classification of fault causes, the scope of fault spread, and the quantitative value of potential power generation loss. The types of latent faults covered include six early latent faults: hot spots on photovoltaic modules, potential-induced degradation, aging of encapsulation materials leading to decreased light transmittance, abnormal degradation of inverter conversion efficiency, abnormal degradation of energy storage module capacity, and abnormal temperature rise caused by excessive contact resistance of wiring terminals. These faults have weak characteristics in the early stages of development and are difficult to identify using traditional threshold judgment methods. By integrating multi-source data in the cloud and using a dual-layer attention mechanism to enhance the ability to capture weak features, the model can accurately identify anomalies in the early stages of faults, preventing further expansion of faults that could damage equipment and cause power generation losses.

[0037] This invention also discloses that when the cloud-integrated decision-making unit generates operation and maintenance handling strategies, it automatically calls a structured historical fault case library. Each fault entry in the case library stores standardized fault feature vectors, fault types, handling procedures, safe operation specifications, required spare parts models and quantities, historical handling time, and fault elimination verification standards.

[0038] The system performs cosine similarity matching between the 28-dimensional feature vector of the current fault and the feature vector of historical faults in the case library. When the matching degree reaches 0.9, the standardized handling plan of the corresponding historical case can be directly reused without the need for manual reformulation of strategies, thus ensuring the standardization of operation and maintenance handling actions. When the matching degree is lower than 0.9, a remote assistance request is automatically sent to the manufacturer's technical personnel of the corresponding equipment model. The system's built-in audio and video channels enable real-time synchronization of on-site images and equipment data. The manufacturer's technical personnel formulate targeted handling strategies based on the fault data. After the strategies are reviewed and approved by the site administrator, they are simultaneously stored in the historical fault case library, enabling dynamic iterative updates of the case library. As the system's operating time increases, the fault scenarios covered by the case library continue to enrich, and the accuracy of strategy generation continues to improve. This reduces the reliance on the experience of senior maintenance personnel, and newly hired maintenance personnel can complete fault handling by following standardized procedures.

[0039] This invention also discloses an operation and maintenance closed-loop execution unit comprising an intelligent dispatch module, an intelligent spare parts scheduling module, a full-process tracking module, and an automated defect elimination verification module. To achieve optimal personnel matching for operation and maintenance tasks and eliminate the subjective bias of manual dispatch, the system quantitatively evaluates the overall suitability of dispatchable operation and maintenance personnel. The suitability score satisfies the following relationship: ;in, The overall adaptability score for operations and maintenance personnel to take on the current work order is given, with a value ranging from 0 to 1. The higher the score, the stronger the adaptability. The straight-line distance between the current real-time location of the maintenance personnel and the fault location, in kilometers; The maximum straight-line distance between the dispatchable maintenance personnel and the fault location within the station area, in kilometers; The estimated remaining processing time for the current pending work orders of the operations and maintenance personnel, in minutes; The maximum estimated remaining processing time for pending work orders among all available maintenance personnel, in minutes; The satisfaction rate of maintenance personnel in handling similar faults in the past, with a value ranging from 0 to 1; The degree of matching between the qualifications held by the maintenance personnel and the current fault handling requirements is 1 if the requirements are fully met, 0.5 if the requirements are partially met, and 0 if the requirements are not met. , , , The weighting coefficients for distance, load, historical performance, and qualification dimensions are configured by the site management personnel according to actual operation and maintenance needs. The default values ​​are 0.3, 0.2, 0.3, and 0.2, and the sum of the four weighting coefficients is 1.

[0040] Through multi-dimensional normalized quantitative evaluation, the system can eliminate the differences in the dimensions of different evaluation indicators, automatically select the most suitable maintenance personnel to undertake tasks, shorten fault response time, and improve maintenance efficiency.

[0041] The intelligent spare parts scheduling module automatically matches the corresponding spare parts model and quantity in the station spare parts warehouse according to the fault type. When the inventory meets the demand, it automatically locks the corresponding spare parts and sends a requisition reminder to the warehouse manager. When the inventory is insufficient, it automatically triggers the transfer process of the nearest regional storage point, eliminating the need for maintenance personnel to check the spare parts inventory themselves. The full-process tracking module synchronizes all nodes of the operation and maintenance personnel's on-site check-in, fault diagnosis, and handling operations in real time. Managers can view the work order progress in the background without repeated phone communication. The automated defect elimination verification module automatically retrieves 10 consecutive minutes of operating data for the corresponding point after the operation and maintenance personnel submit a defect elimination completion application. It compares the deviation with the normal operating baseline value at the same time before the fault occurred. If the deviation is within the normal threshold range, the defect elimination is confirmed to be completed and the work order is automatically closed. Otherwise, the defect elimination application is rejected and re-handled, avoiding the problems of manual verification being a mere formality and incomplete defect elimination.

[0042] This invention also discloses a hierarchical authorization multi-terminal interaction unit that adopts a role-based access control mechanism to achieve hierarchical permission management. It grants site management personnel the right to view full operational data, adjust operation and maintenance strategy parameters, configure system sampling frequency, and allocate permissions. It also supports viewing multi-dimensional statistical reports such as power generation data, fault statistics, and operation and maintenance completion rate of the entire site. For front-line maintenance personnel, the system grants them access to receive fault alarms in their assigned area, update work order status, transmit audio and video footage of on-site troubleshooting, and apply for spare parts. They can only view the equipment data and handling requirements associated with the work orders they have undertaken. For equipment manufacturer technicians, access is granted to iterative updates of fault diagnosis models and remote troubleshooting guidance for specific equipment models. Sensitive data such as total power generation and grid connection revenue are anonymized. Manufacturer personnel can only view the operating parameters of the equipment models they are responsible for and cannot access sensitive operating information of other equipment in the plant.

[0043] All user actions are recorded in an unalterable log, including the operator, time, content, and parameter changes before and after the operation. Log retention meets the traceability requirements of the power system's Classified Protection Scheme 2.0. Alarm notifications employ a tiered delivery mechanism: emergency faults trigger simultaneous voice calls, strong app alerts, and on-site audible and visual alarms; general faults trigger app message notifications; and early warning information is pushed to the system's internal messaging system, preventing indiscriminate alarms from disrupting maintenance personnel's normal work.

[0044] This invention also discloses the establishment of an encrypted model incremental synchronization link between the cloud-based fusion decision unit and the edge-based pre-diagnosis unit. After each iteration and optimization of the deep diagnostic model based on a new fault sample, the cloud automatically generates a lightweight model incremental update package adapted to the corresponding site. The update package uses the bsdiff differential compression algorithm to reduce the transmission volume and only transmits the differences between the old and new models. The size of a single update package does not exceed 1MB, which greatly reduces the transmission traffic consumption.

[0045] The update package is delivered to the edge diagnostic unit of the corresponding site via a breakpoint resume mechanism. The model update process employs a dual-caching mechanism: the old version model continues to run to provide diagnostic services, while the new version model is loaded in a separate background partition. After loading, accuracy verification is performed using 1000 locally stored labeled test samples. If the diagnostic accuracy of the new version model on the local test set is not lower than that of the old version model, the system automatically and smoothly switches to the new version model. The switching process uses memory hot-loading technology, resulting in no service interruption. If the accuracy of the new version model does not meet the requirements, the update package is automatically deleted, and the old version model continues to run, avoiding diagnostic service interruptions caused by model updates. After the model update is completed, the system automatically sends an update result notification to the administrator, eliminating the need for on-site operation by maintenance personnel and significantly reducing the maintenance costs of model iteration.

[0046] Reference Figure 1This flowchart illustrates the overall operational architecture and data flow of a photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things (IoT). At the start of system operation, the global sensing and acquisition unit collects multi-dimensional operational data in real time from various nodes such as photovoltaic modules, inverters, and energy storage modules, and transmits it efficiently via a heterogeneous IoT protocol. Subsequently, the edge pre-diagnostic unit deployed at the site takes over the data, quickly performs initial diagnosis to identify explicit hard faults and trigger alarms. The cloud-based fusion decision unit receives feature fragments uploaded from the edge side, completes multi-source fusion and in-depth analysis of hidden faults, and generates precise operation and maintenance handling strategies accordingly. Finally, the operation and maintenance closed-loop execution unit automatically follows up with dispatching and fault verification according to the strategy, while the multi-terminal interaction unit provides a hierarchical authorization information synchronization hub for site management, operation and maintenance personnel, and manufacturers throughout the process, forming a complete operation and maintenance closed loop.

[0047] Reference Figure 2 This flowchart details the data acquisition, preprocessing, and pre-diagnostic mechanisms at the edge of the photovoltaic power station. The global sensing acquisition unit first comprehensively acquires electrical parameters, environmental parameters, and operating status data of the photovoltaic equipment. In the preprocessing stage, the edge preprocessing module automatically completes missing data using a linear interpolation algorithm, accurately removes outliers using benchmark criteria, and strictly aligns timestamps at the microsecond level. The high-quality preprocessed data is then input into a lightweight fault identification model that has undergone knowledge distillation and compression. This model can rapidly identify explicit hard faults such as component short circuits or inverter overheating. Once a hard fault is detected, the edge side can directly trigger on-site audible and visual alarms without waiting for cloud commands, simultaneously generating an emergency-level maintenance identifier and pushing it to the execution unit, while smoothly uploading characteristic fragments of normal or suspected latent faults to the cloud.

[0048] Reference Figure 3 This flowchart illustrates the core logic of the cloud-based fusion decision-making unit in performing complex data processing and uncovering hidden faults. After receiving real-time feature fragments uploaded from the edge, the cloud performs multi-dimensional strict alignment with historical features from the same station, features from adjacent locations, and meteorological forecast data to construct a high-dimensional multi-source fault feature tensor. Subsequently, the system loads a temporal convolutional network model incorporating an attention mechanism to process this tensor. Through attention weight allocation, highly correlated features are dynamically strengthened, accurately locating the occurrence points, cause classifications, and impact levels of hidden faults such as component hotspots and potential-induced attenuation. Furthermore, based on a continuous time-series data trend extrapolation algorithm, the system predicts equipment performance degradation trends in advance. When the degradation reaches a warning threshold, predictive maintenance tasks are automatically generated, achieving a shift from passive maintenance to proactive prevention.

[0049] Reference Figure 4This flowchart illustrates the closed-loop execution process from generating maintenance and handling strategies to on-site troubleshooting verification. In the strategy generation phase, the system calculates the cosine similarity between the current fault characteristics and the characteristics in the historical fault case library. If the matching degree meets the standard, the system directly reuses the historical successful handling process and spare parts selection; if it does not meet the standard, the system guides equipment manufacturer experts to remotely intervene and customize the strategy. After the strategy is determined, the intelligent dispatch module automatically dispatches work orders based on the fault location, personnel workload, and skill matching degree, while the spare parts scheduling module simultaneously matches the site inventory or triggers nearby parts dispatch. During on-site operations, the process tracking module synchronously checks and monitors the progress of troubleshooting and handling operations in real time. After the operation is completed, the troubleshooting verification module automatically retrieves the latest operating data for that node and compares it with the normal baseline value. The work order is only completed after the data confirms that it has returned to the normal threshold range.

[0050] Reference Figure 5 This flowchart illustrates the multi-role access control mechanism and the dynamic evolution process of the diagnostic model within the system. The multi-terminal interaction unit constructs a strict hierarchical access control system, granting site managers global data and policy adjustment permissions, providing on-site maintenance personnel with permissions to receive area alarms and transmit data back, and offering equipment manufacturers access to model iteration and remote guidance. All operations are fully recorded and traceable. Regarding continuous system evolution, after the deep diagnostic model is iteratively updated in the cloud, the model update package is securely deployed to the edge front-end diagnostic unit using breakpoint resume technology. This update process is executed silently in the background, without interrupting real-time perception data collection and initial diagnostic services on the edge side. After the update package is deployed, the system automatically performs parallel accuracy verification between the old and new models. Once the accuracy meets the standards, it smoothly switches to the new model, ensuring continuous improvement in diagnostic capabilities.

[0051] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things, characterized in that, include: The global sensing and acquisition unit is deployed at various nodes of photovoltaic electronic products. It is used to collect the operating electrical parameters, environmental parameters and operating status data of photovoltaic modules, inverters, combiner boxes and energy storage modules according to the preset sampling frequency. After the collected data is timestamped and screened for bad values, it is transmitted to the edge front-end diagnostic unit through the heterogeneous Internet of Things protocol. Edge-based pre-diagnostic unit: Deployed on the photovoltaic power station side, it is used to load a lightweight fault identification model to complete the initial diagnosis of the received real-time data, identify obvious hard faults and trigger local alarms, and upload the initial diagnosis results and feature fragments associated with suspected hidden faults to the cloud-based fusion decision unit. Cloud-based fusion decision unit: used to perform multi-source feature fusion on the received edge-uploaded data, load deep diagnostic models to locate the location, cause and impact level of hidden faults, and generate operation and maintenance handling strategies that match the fault level; Operation and maintenance closed-loop execution unit: used to automatically generate dispatch instructions and spare parts scheduling instructions according to operation and maintenance handling strategies, track the execution status of the entire operation and maintenance process, and complete the effect verification after fault elimination; multi-terminal interaction unit is used to provide data viewing, alarm push and operation entry for site managers, operation and maintenance personnel and equipment manufacturers with hierarchical authorization, so as to realize information synchronization of the entire operation and maintenance process.

2. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, The global sensing and acquisition unit includes an electrical parameter acquisition module, an environmental parameter acquisition module, an operating condition acquisition module, and an edge preprocessing module. The electrical parameter acquisition module is used to acquire data on the output voltage and current of the photovoltaic string, the DC and AC power of the inverter, the branch current of the combiner box, and the charging and discharging voltage and current of the energy storage module. The environmental parameter acquisition module is used to acquire data on the irradiance of the photovoltaic panel, the ambient temperature, the panel surface temperature, the wind speed, and the precipitation indicator. The operating condition acquisition module is used to acquire data on the switching status of the equipment, the speed of the cooling fan, and the connectivity status of the communication link. The edge preprocessing module uses a linear interpolation algorithm to complete missing data, uses the 3σ criterion to remove outliers, and completes microsecond-level timestamp alignment of multi-source data according to the unique identifier of the acquisition point.

3. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, The lightweight fault identification model loaded by the edge pre-diagnostic unit is a lightweight convolutional neural network model that has undergone knowledge distillation and compression. During the model training phase, the historical hard fault sample set of the photovoltaic power station is used as the training input. The identified explicit hard faults include module short circuit, module open circuit, inverter over-temperature shutdown, combiner box branch interruption, and communication link interruption. When the edge pre-diagnostic unit identifies an explicit hard fault, it directly triggers on-site audible and visual alarms and simultaneously generates an emergency-level operation and maintenance identifier. It can be pushed to the operation and maintenance closed-loop execution unit without waiting for secondary diagnosis in the cloud.

4. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, The multi-source feature fusion process of the cloud-based fusion decision unit specifically involves dimensionally aligning the received real-time operational feature fragments, historical operational features of equipment of the same model at the same station, operational features of equipment at adjacent locations, and meteorological forecast data to construct a multi-dimensional fault feature tensor.

5. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 4, characterized in that, The deep diagnostic model loaded by the cloud-integrated decision unit is a temporal convolutional network model that introduces an attention mechanism. The model strengthens the weights of feature dimensions that are highly correlated with faults through attention weight allocation, and outputs the specific location of the fault, the classification of fault causes, the scope of fault spread and the risk level. The diagnosed hidden faults include hot spots in components, potential-induced attenuation, abnormal decrease in inverter efficiency, capacity attenuation of energy storage modules, and loose connection faults in wiring terminals.

6. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, When the cloud-based fusion decision-making unit generates an operation and maintenance handling strategy, it calls the historical fault case library and performs cosine similarity matching between the fault characteristics diagnosed at present and the characteristics of historical cases. When the matching degree reaches a preset threshold, the handling process, safe operation specifications and spare parts selection information of the corresponding historical cases are directly reused. When the matching degree is lower than the threshold, experts from the equipment manufacturer remotely intervene to complete the strategy formulation and update it synchronously to the historical fault case library.

7. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, The operation and maintenance closed-loop execution unit includes an intelligent dispatch module, a spare parts scheduling module, a process tracking module, and a defect elimination verification module. The intelligent dispatch module automatically allocates operation and maintenance tasks based on the location of the fault point, the real-time location of the operation and maintenance personnel, the current workload of pending work orders, and the skill matching degree. The spare parts scheduling module automatically matches the corresponding spare parts in the station's inventory according to the fault type. When the inventory is insufficient, it automatically triggers the transfer process to the nearest storage point. The process tracking module synchronizes the check-in, fault investigation, and handling progress of maintenance personnel in real time. After the maintenance personnel submit the defect elimination completion application, the defect elimination verification module automatically retrieves the real-time operation data of the corresponding point and compares it with the normal operation benchmark value before the fault occurred. If the data returns to the normal threshold range, the defect elimination is confirmed to be completed.

8. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, The multi-terminal interaction unit adopts a hierarchical permission control mechanism, granting site management personnel full access to view operational data, adjust maintenance strategies, and configure system parameters; granting maintenance personnel access to receive fault alarms, update work order status, and transmit on-site data for their assigned area; and granting equipment manufacturers access to iterate fault diagnosis models for corresponding equipment models and provide remote guidance on difficult faults. The operation records of different permission holders are fully traceable.

9. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, A model synchronization link is set up between the cloud-based fusion decision unit and the edge-based pre-diagnostic unit. After each deep diagnostic model iteration update is completed in the cloud, the corresponding site's model update package is sent to the edge-based pre-diagnostic unit through a breakpoint resume method. The model update process does not interrupt the real-time data acquisition and diagnostic services on the edge side. After the update is completed, the accuracy verification of the old and new models is automatically completed. If the verification accuracy meets the requirements, the new model is automatically switched to run.

10. The photovoltaic electronic product fault diagnosis and remote operation and maintenance system coupled with the Internet of Things as described in claim 1, characterized in that, The cloud-based fusion decision-making unit is also equipped with a fault prediction module. Based on continuously collected device runtime sequence data, it predicts the performance degradation trend of the device within a preset period through a trend extrapolation algorithm. When the performance degradation reaches the warning threshold, it generates a predictive maintenance task.

Images