AI tracker intelligent platform

Abstract

The invention relates to the field of photovoltaic intelligent operation and maintenance, and particularly discloses an AI tracker intelligent platform, which comprises a data acquisition module, a data analysis module, a control strategy generation module and a remote control execution module which are connected in sequence, and is characterized in that the data acquisition module acquires operation and environment data from a tracker NCU / TCU unit; the data analysis module performs fault diagnosis and health assessment through an AI model; the control strategy generation module generates a control instruction according to the health report and the time-of-use electricity price; and the remote control execution module completes instruction issuing and feedback closed loop. The system supports private local deployment, public cloud deployment and local and public cloud mixed deployment of a power station, integrates Beidou communication, has an active protection strategy based on prediction and a full-localization software system, solves the problems of large control delay, passive operation and maintenance mode, insufficient safety protection, inflexible deployment and the like of an existing system, and is suitable for popularization and application. And intelligent, high-reliability and income-maximized operation of the photovoltaic tracker is realized.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic intelligent operation and maintenance technology, specifically to an AI tracker intelligent platform. Background Technology

[0002] With the advancement of market-based trading in photovoltaic power plants, the operational objective has shifted from maximizing power generation to maximizing revenue. Tracking systems, which can flexibly adjust the power generation curve based on time-of-use pricing, have become a key means of enhancing the asset value of power plants. While some photovoltaic monitoring systems exist, the following prominent technical issues remain:

[0003] In terms of control performance, existing systems mostly adopt a centralized architecture and a common communication protocol, which makes it difficult to achieve high-precision control of a large number of tracker units (NCU / TCU) at the second level. Especially in areas with complex network conditions or remote areas, the problems of control delay and insufficient accuracy are significant, which limits the ability to flexibly adjust the power generation curve.

[0004] In terms of system deployment and data security, existing solutions often fail to meet the diverse needs of different customers: large-scale ground-mounted power plants have extremely high requirements for data sovereignty and real-time performance, necessitating localized deployment; while distributed power plants tend to favor low-cost, easily scalable cloud services. Existing systems mostly support a single mode, making it difficult to strike a balance between security and cost-effectiveness.

[0005] In terms of operation and maintenance, most systems still rely on regular inspections and post-incident repairs, lacking predictive maintenance capabilities based on artificial intelligence, resulting in high operation and maintenance costs, slow fault response, and unnecessary power generation losses.

[0006] In terms of security protection mechanisms, existing systems typically only have simple threshold alarms and passive protection functions, lacking proactive risk avoidance strategies based on weather forecasts that are deeply integrated with tracker hardware. As a result, the equipment faces higher security risks in severe weather.

[0007] Therefore, there is an urgent need for an intelligent tracker management platform that can integrate AI diagnostics, support flexible deployment, and possess high-precision control and proactive safety protection capabilities to comprehensively improve the profitability, safety, and operational efficiency of photovoltaic power plants. Summary of the Invention

[0008] The purpose of this invention is to provide an AI tracker intelligent platform to solve the comprehensive problems of existing photovoltaic tracker systems in terms of control accuracy, operation and maintenance mode, security protection and deployment flexibility.

[0009] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0010] An AI tracker intelligent platform includes:

[0011] The data acquisition module is used to collect operational and environmental data from the NCU or TCU unit of the photovoltaic tracker;

[0012] The data analysis module is used to receive the operational data and environmental data, perform fault diagnosis and status assessment through AI models, and generate a health assessment report.

[0013] The control strategy generation module is used to generate tracker control instructions based on the health assessment report and time-of-use electricity price data;

[0014] The remote control execution module is used to send the tracker control commands to the corresponding NCU or TCU unit to adjust the tracker angle.

[0015] As a preferred embodiment of the present invention, the data acquisition module specifically includes:

[0016] The equipment data acquisition unit is used to collect real-time operating status data and electrical parameters, including tracker angle, motor current and voltage, drive status and fault codes, through the sensors built into the NCU or TCU unit, and to perform preliminary structured packaging of the collected data.

[0017] The environmental data acquisition unit is used to connect to the meteorological station deployed in the photovoltaic power station or access the authoritative meteorological data service interface to collect environmental data and geographic information data, including irradiance, ambient temperature, wind speed and direction, precipitation and snowfall, and component temperature.

[0018] The protocol adaptation unit is used to dynamically select or enable communication protocols in parallel according to the power plant deployment mode and network conditions: in the local area network environment, Zigbee or LoRa protocol is given priority for low-latency and high-security intranet communication; when remote public network communication is required, 4G or 5G mobile network protocol is enabled; in scenarios with no public network signal or strict requirements for communication autonomy, BeiDou short message communication protocol is enabled for data transmission.

[0019] As a preferred embodiment of the present invention, the data analysis module specifically includes:

[0020] The data preprocessing unit is used to clean, normalize, and extract features from the operational data and environmental data from the data acquisition module to form a structured analysis dataset.

[0021] The AI ​​fault diagnosis unit is used to receive the structured analysis dataset and identify potential fault types and abnormal operating modes of the NCU or TCU unit through a pre-trained AI analysis model.

[0022] The status assessment unit is used to assess the real-time health status and performance degradation trend of the equipment based on the potential fault types and abnormal operating modes, combined with the equipment's historical operating data.

[0023] The report generation unit is used to summarize the potential fault types, abnormal operating modes, real-time health status and performance degradation trends, generate a health assessment report including a fault warning list, a health status overview and maintenance priority suggestions, and output it to the control strategy generation module.

[0024] As a preferred embodiment of the present invention, the workflow of the AI ​​fault diagnosis unit specifically includes:

[0025] The system receives a structured analysis dataset from the data preprocessing unit and calls a pre-trained machine learning model, which is trained using historical fault data and integrates equipment operation mechanisms with data-driven modes.

[0026] The structured analysis dataset is input into the machine learning model. The machine learning model first performs sliding window feature extraction on the time series data to obtain time series feature vectors. Then, it performs pattern matching and probability calculation based on the time series feature vectors, outputs the confidence probability for each type of preset fault, and marks faults with probabilities exceeding a set threshold as potential fault types.

[0027] For operational data that does not directly match the preset fault type but whose feature vector deviates significantly from the normal baseline, it is marked as an abnormal operation mode.

[0028] As a preferred embodiment of the present invention, the workflow of the state assessment unit specifically includes:

[0029] The system receives potential fault types, corresponding confidence probabilities, and abnormal operating modes from the AI ​​fault diagnosis unit. Based on the potential fault types and abnormal operating modes, it queries a pre-set equipment impact weight library to determine the quantitative impact coefficient of each fault or abnormality on the overall health of the equipment.

[0030] The system retrieves the historical performance baseline of the device, compares the current operating data with the corresponding historical performance data under the same conditions, and calculates the performance deviation rate.

[0031] Based on the quantified impact coefficient, confidence probability, and performance deviation rate, the data are input into a health index-based degradation model for calculation, outputting a quantified real-time health status score, and generating a performance degradation trend prediction curve based on time-series data analysis.

[0032] As a preferred embodiment of the present invention, the control strategy generation module specifically includes:

[0033] The strategy input unit is used to receive health assessment reports, synchronously access the time-of-use electricity price data stream released by the electricity market, and perform timestamp alignment and formatted encapsulation on both.

[0034] The active protection strategy unit is used to generate active risk avoidance instructions, including strong wind protection strategy and heavy snow protection strategy, based on real-time or predicted meteorological data.

[0035] The power generation revenue optimization unit is used to integrate the equipment status information and time-of-use electricity price data in the health assessment report, and calculate and generate tracker angle adjustment instructions for optimizing power generation revenue through the revenue maximization model.

[0036] The command fusion and issuance unit is used to perform priority arbitration and conflict resolution on the active risk avoidance command and angle adjustment command, generate the final tracker control command, and output it to the remote control execution module.

[0037] As a preferred embodiment of the present invention, the working process of the active protection strategy unit specifically includes:

[0038] It continuously receives formatted meteorological data from the strategy input unit, as well as short-term forecast meteorological data from the meteorological server;

[0039] The real-time wind speed data is compared with the gale protection threshold, and the real-time snowfall data or predicted snowfall data is compared with the heavy snow protection threshold. When the real-time wind speed exceeds the gale protection threshold, or the real-time snowfall or predicted snowfall exceeds the heavy snow protection threshold, it is determined that active protection needs to be triggered.

[0040] Based on the type of protection triggered, the corresponding preset safety angle is retrieved from the preset safety policy library. The safety angle is determined based on the tracker's structural strength and local wind pressure and snow load model calculations; a forced execution command containing the target safety angle and with the highest execution priority is generated.

[0041] The system monitors weather conditions in real time until wind speed or snowfall data falls below a safe threshold and remains below it for a preset duration. Then, it generates a command to release the protection status and returns control to the power generation revenue optimization unit.

[0042] As a preferred embodiment of the present invention, the workflow of the power generation revenue optimization unit specifically includes:

[0043] The system continuously monitors whether any forced execution commands are triggered. If not, it receives a health assessment report and real-time and predicted time-of-use electricity price data from the policy input unit. It extracts the real-time health status score and performance degradation trend from the health assessment report and determines the allowable tracking angle range and maximum angular velocity limit under the current device status based on a predefined mapping relationship.

[0044] The allowable tracking angle range, maximum angular velocity limit, real-time irradiance data, and time-of-use electricity price data are used as boundary conditions and input parameters, and then input into the profit maximization model.

[0045] The revenue maximization model aims to maximize the expected power generation revenue within a future scheduling cycle. Based on the photovoltaic power generation physical model and electricity price time series, it performs rolling optimization calculations to solve for the optimal tracker angle time series.

[0046] The optimal tracker angle time series is converted into specific, timestamped angle adjustment instructions, and the angle adjustment instructions are output to the instruction fusion and distribution unit.

[0047] As a preferred embodiment of the present invention, the remote control execution module specifically includes:

[0048] The instruction receiving and verification unit is used to receive tracker control instructions from the instruction fusion and sending unit, and verify the legality of their format and the validity of the target device address.

[0049] The communication protocol adaptation and encapsulation unit is used to encapsulate the verified tracker control commands into data frames of the corresponding protocol according to the network type and communication protocol accessed by the target NCU or TCU unit.

[0050] The instruction issuing unit is used to send the data frame to the target NCU or TCU unit through the corresponding physical communication link, driving the tracker to perform an angle adjustment action;

[0051] The execution feedback acquisition unit is used to receive execution status feedback data from the NCU or TCU unit in real time after the command is issued, and return it to the data analysis module to update the equipment status and form a control closed loop.

[0052] As a preferred embodiment of the present invention, the workflow of the execution feedback acquisition unit specifically includes:

[0053] After the instruction issuing unit successfully sends a data frame, it immediately starts a feedback waiting timer for the current instruction and marks the instruction as being in execution; it listens for and receives the execution status feedback data after being reverse-parsed by the communication protocol adaptation and encapsulation unit;

[0054] The received feedback data is matched and verified with the original command issued. If the deviation between the actual execution angle and the target angle is within the allowable error range and the status code is success, the command is determined to have been executed successfully.

[0055] If no feedback is received before the feedback waiting timer expires, or the feedback status code is failure, or the actual execution angle deviation exceeds the limit, it is determined that the instruction execution is abnormal.

[0056] For instructions that are determined to be executed successfully, the execution status mark is cleared, and the execution status feedback data containing the actual execution result is packaged and sent back in real time to the data preprocessing unit of the data acquisition module and the data analysis module for updating the real-time status database of the device.

[0057] For instructions that are determined to be abnormal, the instruction resending process is triggered according to the preset retry policy; if the error still occurs after retry, an alarm event containing the device address, the type of error, and a timestamp is generated and pushed to the AI ​​fault diagnosis unit and the operation and maintenance alarm system of the data analysis module.

[0058] Compared with the prior art, the present invention has the following advantages:

[0059] 1. By introducing AI-powered health diagnostics and predictive maintenance models, the traditional passive and planned operation and maintenance model is transformed into proactive predictive maintenance based on the real-time health status of equipment, thereby significantly reducing operation and maintenance costs and power generation losses. By constructing a revenue optimization model that integrates time-of-use pricing and equipment status, the power generation curve is dynamically adjusted to match high-price periods in the electricity market, while ensuring the safe operation of equipment, achieving a fundamental shift from pursuing power generation volume to pursuing power generation revenue.

[0060] 2. By designing an integrated proactive protection strategy unit for weather forecasting, early warning and rapid automatic avoidance of severe weather such as strong winds and heavy snow were achieved, significantly improving the operational safety of the equipment in complex environments. By establishing a closed-loop command execution system that includes full-process verification and feedback confirmation, reliable delivery and accurate execution of control commands were ensured, effectively solving the uncertainty problem in remote control and greatly enhancing the robustness of the overall system.

[0061] 3. By supporting a dual-mode deployment architecture (local-cloud) and deeply integrating BeiDou communication, the platform can simultaneously meet the diverse needs for data security, real-time performance, and network coverage in different scenarios, thus expanding its applicability. By adopting a fully domestically produced software technology stack, the core software system has achieved independent controllability, effectively safeguarding the network security and industrial chain resilience of energy infrastructure. Attached Figure Description

[0062] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0063] Figure 1 This is a block diagram of the platform described in Embodiment 1 of the present invention.

[0064] Figure 2 This is a flowchart of the data acquisition module of the platform described in Embodiment 1 of the present invention.

[0065] Figure 3 This is a flowchart of the data analysis module of the platform described in Embodiment 1 of the present invention.

[0066] Figure 4 This is a flowchart of the control strategy generation module of the platform described in Embodiment 1 of the present invention.

[0067] Figure 5 This is a flowchart of the remote control execution module of the platform described in Embodiment 1 of the present invention. Detailed Implementation

[0068] 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.

[0069] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0070] Example 1

[0071] like Figure 1 - Figure 5 As shown, the present invention provides an AI tracker intelligent platform, comprising:

[0072] S1. Data acquisition module, used to collect operational and environmental data from the NCU or TCU unit of the photovoltaic tracker; specifically including:

[0073] S11. Equipment Data Acquisition Unit: This unit uses multiple sensors deployed within the tilt control unit (NCU) or tracking control unit (TCU) of the photovoltaic tracker to perform high-frequency, real-time monitoring of the tracker's operating status. The specific acquisition process is as follows:

[0074] The system uses angle sensors or rotary encoders to accurately acquire real-time angle data of the tracker bracket, including azimuth and pitch angles; it collects the operating current and voltage of the drive motor through current and voltage sensors to determine the motor load and energy consumption level; it reads the drive status word in the internal register of the TCU or NCU to obtain the motor start / stop status, running direction, limit switch trigger status, and driver temperature information; and it captures the fault diagnosis codes generated by the tracker control system, covering fault types such as motor overload, communication interruption, position deviation exceeding limits, and power supply abnormality.

[0075] After analog-to-digital conversion and numerical filtering, the raw data is initially structured and encapsulated within the device's data acquisition unit according to a predefined data frame format, forming a standardized data packet containing device ID, timestamp, data type identifier, numerical value, unit, and verification field, providing a unified data interface for subsequent transmission and processing.

[0076] S12. Environmental data acquisition unit, which acquires external environmental parameters of the photovoltaic power station through two complementary paths:

[0077] Firstly, a professional meteorological monitoring station is deployed at the power station site. This meteorological station integrates a total radiation meter to measure the irradiance of the horizontal and inclined surfaces, deploys an ambient temperature sensor to obtain the atmospheric temperature, installs an ultrasonic anemometer to obtain real-time wind speed and direction data, sets up a rain and snow gauge to record precipitation intensity, and attaches a temperature sensor to the back panel of the photovoltaic module to directly measure the module's operating temperature.

[0078] Secondly, it accesses an authoritative meteorological data service platform via API to obtain macro-meteorological forecast data and historical climate statistics for the geographical coordinates of the power station location.

[0079] The environmental data acquisition unit aligns the above-mentioned on-site measured data with third-party service data along the timeline, and integrates geographical information data such as the latitude and longitude of the power station, altitude, surface reflectivity, and array layout orientation to form a complete environmental dataset, which is used for subsequent AI model analysis of irradiation resource availability and environmental stress factors.

[0080] S13. Protocol adaptation unit: Based on the actual deployment topology and communication infrastructure conditions of the photovoltaic power station, intelligently decides and dynamically switches the data transmission protocol stack. The specific switching mechanism is as follows:

[0081] Under the condition that a local area network is set up inside the power station and a Zigbee gateway or Lora gateway is available, the protocol adaptation unit will prioritize the activation of the Zigbee wireless mesh network protocol or the Lora protocol to communicate between the device data acquisition unit and the edge computing node, and use its low power consumption, low latency, high security and self-organizing network characteristics to achieve efficient data collection within the station.

[0082] When the tracker is located in a remote area or requires remote monitoring across regions, the protocol adaptation unit activates the 4G or 5G mobile network communication module, transmits encrypted data to the cloud platform through the operator's public network, and can automatically switch between 4G and 5G modes according to network signal strength and tariff policies.

[0083] In special application scenarios such as military and border defense where public network signal coverage is blind or where strict requirements are placed on communication autonomy, the protocol adaptation unit enables the BeiDou short message communication terminal to compress and package key operational data and environmental parameters according to the BeiDou short message capacity limit. It then uses the short message service of the BeiDou satellite navigation system to achieve emergency data reporting and remote command reception, ensuring communication reliability under extreme conditions.

[0084] The above three protocols can be enabled in parallel based on network quality monitoring results to achieve redundant backup and load balancing of data transmission paths.

[0085] S2. Data Analysis Module: This module receives operational and environmental data, uses AI models for fault diagnosis and status assessment, and generates a health assessment report. Specifically, it includes:

[0086] S21. Data preprocessing unit receives multi-source heterogeneous data streams from the data acquisition module. First, it performs data cleaning: it identifies and removes abnormal jump values ​​in the running data and environmental data using the 3σ criterion or the isolated forest algorithm. For missing values ​​caused by communication interruption or sensor failure, it fills them with linear interpolation or K-nearest neighbor interpolation based on historical similar working conditions. At the same time, it aligns the time series based on the timestamp information of each data source and the tracker control cycle to eliminate the timing deviation caused by network transmission delay.

[0087] After cleaning, the maximum-minimum normalization method is used to map physical quantities with different dimensions such as tracker angle, motor current and voltage, irradiance, and temperature to the [0,1] numerical range, thereby eliminating the interference of dimensional differences on subsequent model training.

[0088] Based on this, a feature engineering system is constructed using knowledge of photovoltaic tracker operation mechanism and fault diagnosis: current harmonic distortion rate is extracted from motor current and voltage data to identify driver abnormalities; the angle tracking deviation rate is calculated by comparing the target angle with the actual angle to evaluate tracking accuracy; the temperature change gradient is obtained by using the temporal difference between component temperature and ambient temperature to determine the risk of thermal runaway; and multi-dimensional features such as wind speed and direction change rate and irradiance fluctuation intensity are extracted. Finally, these features are integrated to form a structured analysis dataset containing device ID, timestamp, and dozens of derived features, providing standardized input for AI fault diagnosis.

[0089] S22. The AI ​​fault diagnosis unit receives structured analysis datasets and, through a hybrid model architecture that deeply integrates equipment physical mechanisms and data-driven algorithms, achieves accurate identification of potential faults and capture of abnormal patterns in the NCU and TCU units; specifically:

[0090] S221. At the model architecture design level, the pre-trained machine learning model called by this unit adopts a hierarchical heterogeneous structure: the bottom layer is a time-series encoder based on Long Short-Term Memory (LSTM) network, which is specifically used to process continuous time-series data such as current, voltage, and angle, and capture their long-range dependencies and dynamic evolution laws; the middle layer is a local feature extractor based on Convolutional Neural Network (CNN), which performs spatial pattern mining on derived features such as harmonic distortion rate and temperature gradient, and identifies the microscopic texture of early signs of faults; the top layer is a fusion decision layer based on the attention mechanism, which dynamically allocates the weight coefficients of different feature sources. For example, when the wind speed changes suddenly, it automatically increases the attention score of mechanical structure-related features, and strengthens the contribution of electrical parameter features when the irradiance fluctuates drastically.

[0091] The model training dataset covers real-world fault cases accumulated over the past three years, including twelve pre-defined fault labels such as motor stall, gearbox tooth breakage, encoder signal loss, communication protocol packet loss, power module ripple exceeding limits, and controller memory overflow. Each fault category has no fewer than 500 samples, and the dataset also includes more than 100,000 hours of normal operating baseline data. Through a semi-supervised learning strategy, the model not only grasps the data distribution patterns but also incorporates physical mechanism constraints such as tracker transmission ratio, motor torque coefficient, and PID control parameters.

[0092] S222. In the sliding window feature extraction stage, the model performs multi-scale window scanning on the input structured analysis dataset: a short window with a length of 50 sampling points, corresponding to a duration of 10 seconds, and a step size of 1 sampling point, is set to capture transient impact features of faults, such as the current spike at the moment of motor startup; a long window with a length of 600 sampling points, corresponding to a duration of 2 minutes, and a step size of 10 sampling points is set to identify fault evolution trend features, such as the drift pattern of angle deviation accumulating over time.

[0093] Each window not only extracts basic statistics such as mean, variance, kurtosis, and skewness, but also calculates complexity indicators such as autocorrelation coefficient, sample entropy, and permutation entropy. Furthermore, it obtains the energy spectrum distribution characteristics through wavelet transform, and finally maps each window into a 128-dimensional time-series feature vector.

[0094] For multivariate time series, a strategy of independent channel encoding followed by concatenation is adopted. That is, 128-dimensional features are extracted for each variable such as current, voltage, and angle, and then concatenated by channel to form a 384-dimensional joint feature representation, ensuring that the coupling relationship between different physical quantities is preserved.

[0095] S223. In the pattern matching and probability calculation stage, the top classification head of the model adopts a fully connected layer + Softmax activation function structure to map the 384-dimensional joint features to a 12-dimensional fault category space and output the normalized confidence probability of each preset fault category.

[0096] To improve the sensitivity of the judgment criteria, a dynamic threshold adjustment mechanism is introduced: when the equipment health status score is higher than 80, the fault judgment threshold is automatically raised to 0.85 to reduce false alarms; when the health status score is lower than 60, the threshold is lowered to 0.65 to improve the fault detection rate.

[0097] The model synchronously calculates the Mahalanobis distance between the feature vector and the normal baseline distribution. When this distance exceeds the 95th percentile of the normal samples in the training set and the probabilities of all fault categories are below the threshold, the system initiates the abnormal operation mode judgment process.

[0098] For abnormal patterns, the isolated forest anomaly score and the local outlier factor (LOF) are calculated separately. If both indicators deviate significantly from the normal range, the current running status is marked as an unknown anomaly, and a data recording and expert review mechanism is triggered.

[0099] S224. In the inference execution and result output stages, the AI ​​fault diagnosis unit processes real-time data streams using streaming computing. Each time a sliding window of time-series data is received, a forward inference of the model is triggered, with a single inference latency controlled within 50 milliseconds to ensure the real-time performance of fault detection. The output is encapsulated as a structured JSON object, containing the following fields: fault type (e.g., F001 represents motor stall); confidence; severity level (1-5); anomaly score; first detected time; and window data hash for traceability.

[0100] If the same fault is detected in three consecutive windows and the confidence level continues to rise, the system will automatically upgrade the alarm level of the fault and shorten the subsequent sliding window step to 1 sampling point to achieve accurate tracking. For occasional low-confidence alarms, they will be included in the statistical counter. When the frequency exceeds the preset threshold within a unit of time, a formal alarm will be triggered to effectively suppress noise interference.

[0101] S23. The status assessment unit, through the construction of a multi-dimensional quantitative assessment system and a dynamic baseline comparison mechanism, achieves accurate measurement of the real-time health of the equipment and forward-looking prediction of performance degradation trends; specifically:

[0102] S231. In the equipment impact weight library construction and query stage, the pre-built weight library within this unit is not a static experience table, but a knowledge graph dynamically generated based on the Failure Mode and Effects Analysis (FMEA) method. For the NCU unit, the weight library is subdivided into 8 types of fault modes, such as main control chip failure, memory bit flipping, CAN bus congestion, and power supply ripple exceeding limits; for the TCU unit, it covers 10 types of fault modes, such as motor drive module overload, encoder signal distortion, limit switch failure, and reducer wear.

[0103] The impact coefficient of each type of fault mode is a weighted composite of three sub-factors: the severity factor S is divided into 1-10 levels based on the power generation loss rate caused by the fault. For example, if the motor stall directly causes power generation interruption, the S value is set to 9; if the encoder has a slight deviation that only affects the tracking accuracy, the S value is 3; the occurrence frequency factor O is adjusted based on the fault probability density of historical statistics. The O value of high-frequency faults is close to 1, and the O value of low-frequency faults is close to 0; the detectability factor D reflects the length of the fault latency period. The longer the latency period, the higher the D value.

[0104] The final influence coefficient α = S × O × D / 100 is normalized to the interval [0, 1].

[0105] After receiving the fault type code output by the AI ​​fault diagnosis unit, the state assessment unit instantly retrieves the corresponding α value from the weight library using a hash index, and then performs a weighted correction based on the confidence probability β to obtain the actual impact weight ω = α × β. 1 / 2 This ensures that the impact of low-confidence faults on overall health is moderately attenuated.

[0106] S232. In the historical performance baseline generation and retrieval phase, the baseline data is not a simple historical average, but a dynamic expectation model constructed based on operating condition clustering and similar day matching. This unit divides the equipment's historical operating data into 4×5×4×3=240 operating condition cells according to four dimensions: season, irradiance range, wind speed level, and working period. Each cell stores the probability distribution characteristics of key parameters such as motor current, angle tracking error, and response time under that operating condition: mean μ, standard deviation σ, and skewness γ.

[0107] Upon receiving current operating data, the unit first extracts real-time operating condition tags, such as "Spring - Irradiance 800W / m²". 2 "-Wind speed 3m / s-Morning period" is used to locate the corresponding operating condition cell. The μ and σ values ​​of that cell are used as dynamic baselines instead of using the historical average values ​​for the entire period, thereby eliminating the interference of seasonality and operating condition fluctuations on the evaluation results.

[0108] The baseline data is updated incrementally daily through an online learning mechanism: the effective data of the day is assigned to the corresponding cells according to the operating condition label, and μ and σ are updated using an exponentially weighted moving average (EWMA). The learning rate λ is set to 0.05 to ensure that the baseline slowly tracks the natural aging trend of the equipment while remaining sensitive to short-term performance fluctuations.

[0109] S233. In the multi-dimensional calculation of performance deviation rate, the unit not only calculates the single overall deviation, but also decomposes it into three sub-dimensions: electrical performance deviation, mechanical performance deviation, and control performance deviation.

[0110] Electrical performance deviation rate δ elec The value is obtained by calculating the relative increase of the real-time current harmonic distortion rate (THD) to the baseline (THD0), i.e., δ. elec =(THD-THD0) / (THD0+0.01), add a small constant to the denominator to prevent division by zero;

[0111] Mechanical performance deviation rate δ mech Constructed based on the root mean square error (RMSE) of angle tracking, δ mech =RMSE / θ allow , where θ allow The allowable threshold for tracking accuracy is typically ±2°;

[0112] Control performance deviation rate δ ctrl By comparing the actual response time τ with the baseline response time τ0, δ ctrl =(τ-τ0) / τ0.

[0113] The three types of deviation rates are weighted and combined to form the overall performance deviation rate Δ = 0.3 × δ elec +0.5×δ mech +0.2×δ ctrl The weighting distribution reflects the dominant contribution of mechanical tracking accuracy to power generation.

[0114] If a certain type of deviation rate exceeds 0.5, the unit will automatically trigger a special in-depth analysis of that type of performance, calling higher frequency sampling data to locate the root cause of the deviation.

[0115] S234. In the calculation of the health index decline model, the model adopts an exponential health decay function H(t) = H0·exp(-∫0 t ·λ(t)dt), where: H0 is the initial health value of the device, set to 100; ∫0 t Let λ(t) represent the definite integral from the initial time 0 to the current time t, that is, the continuous summation of the instantaneous decay rate λ(t) over the time interval [0, t]; λ(t) is the instantaneous decay rate, which is dynamically determined by the fault impact weight ω and the performance deviation rate Δ: λ(t) = λ0 + ∑ω i·k1+Δ·k2, where λ0 is the natural aging rate, usually taken as 0.001 / day, and k1 and k2 are proportionality coefficients, set to 0.5 and 0.3 respectively.

[0116] The differential equation is solved in discrete time, with the health value updated once every evaluation period Δt, set to 15 minutes: H(t+Δt)=H(t)·exp(-λ(t)·Δt).

[0117] The model also introduces a health recovery mechanism, whereby the fault weight term ω in λ(t) is adjusted after the fault is repaired or the anomaly disappears. i Once the health value is reset to zero, it enters a slow recovery phase, but the upper limit of recovery is set at 95% of the level before the failure, reflecting the engineering reality that the performance of the equipment cannot be fully restored after repair.

[0118] S235. In the health status scoring and dynamic threshold process, after the calculated H(t) value is mapped to the 0-100 scoring range, the unit further divides it into five health status levels: 90-100 points is the healthy level, the equipment can operate normally and the maintenance cycle can be extended; 70-89 points is the good level, it is recommended to strengthen monitoring and plan preventive maintenance; 50-69 points is the attention level, the inspection cycle needs to be shortened and spare parts need to be prepared; 30-49 points is the poor level, the machine should be shut down for maintenance in the near future; 0-29 points is the danger level, the machine must be shut down immediately and key components must be replaced.

[0119] The thresholds at each level are dynamically adjusted based on the equipment's operating time: for new equipment that has been in operation for less than one year, the threshold is increased by 5 points to implement stricter quality control standards; for older equipment that has been in operation for more than five years, the threshold is decreased by 3 points to accommodate normal aging phenomena.

[0120] In addition, the unit monitors the variance of the health score fluctuation over 30 consecutive evaluation cycles. If the variance exceeds 100 and the mean is below 60, an unstable status flag is automatically triggered, indicating that the equipment has intermittent soft faults.

[0121] S236. In the performance degradation trend prediction curve generation stage, the unit adopts a time-series extrapolation algorithm based on Gaussian process regression (GPR), using the health score sequence {H(t-7d),...,H(t)} of the past 7 days as training samples to construct a non-parametric probabilistic model of the evolution of health values ​​over time. The kernel function of this model is a composite form of a squared exponential kernel and a periodic kernel to capture the long-term monotonic trend and short-term cyclical fluctuations of health degradation, such as day-night temperature cycles and weekday / weekend maintenance differences.

[0122] The GPR model outputs a mean prediction curve for the health score over the next 30 days. pred (t) and the 95% confidence interval [μ] pred -1.96σ pred , μ pred+1.96σ pred ], where: μ pred The mean prediction curve representing the health score indicates the best estimate of health status at a future point in time, i.e., the most likely health score; σ pred The standard deviation representing health prediction quantifies the degree of uncertainty in the prediction results, σ. pred A larger value indicates a wider range of possible fluctuations in health status at that moment, and a lower predictive reliability.

[0123] When the prediction curve indicates that the health value will fall below the danger threshold of 30 points within the next 72 hours, the unit automatically generates an emergency warning and reverse-engineers the main causes of the decline by analyzing ω. i The biggest contributor, as clearly indicated in the trend report, is that "the health value is expected to drop to a dangerous level on [Month] [Day], primarily due to performance degradation of the motor drive module; immediate preparation of replacement parts is recommended." The trend report also includes a degradation acceleration indicator, a = d²H / dt². If a remains negative and its absolute value increases, it indicates that the equipment is in an accelerated deterioration phase, requiring an upgrade to the maintenance response level.

[0124] S24. The report generation unit summarizes and visualizes all the analysis results generated in S22 and S23. First, it generates a fault warning list by sorting the faults in descending order of confidence level. The list clearly lists the potential fault type, the time of first detection, the confidence level probability, the description of the abnormal operation mode, and the possible root cause analysis. Second, in the health status overview section, it displays the real-time health status score in the form of a dashboard and overlays the performance degradation trend prediction curve, while also marking the environmental stress factor level under the current operating conditions.

[0125] Finally, based on health scores, fault weights, and trend prediction results, the analytic hierarchy process (AHP) is used to calculate the maintenance priority score for each tracker, generating a health assessment report that includes specific maintenance recommendations, estimated processing time, and priority ranking. This report is output to the control strategy generation module in JSON structured format, providing a decision-making basis for generating tracker control instructions that balance economy and reliability.

[0126] S3. Control strategy generation module, used to generate tracker control instructions based on the health assessment report and time-of-use electricity price data; specifically including:

[0127] S31. Policy Input Unit, serving as the data hub for controlling policy generation, receives two types of heterogeneous input streams in parallel and performs spatiotemporal reference unification processing:

[0128] Firstly, a message queue mechanism is used to continuously subscribe to health assessment reports from the data analysis module. These reports are in JSON structured format and include fields such as device ID, timestamp, health status score, performance degradation trend curve, fault warning list, and maintenance priority suggestions.

[0129] Secondly, it accesses the time-of-use electricity price data stream released by the provincial power trading center in real time through a WebSocket long connection. This data stream follows the DL / T 645 protocol and includes the electricity price forecast, real-time electricity price, peak-valley-flat indicator and market clearing price for every 15 minutes in the next 24 to 168 hours.

[0130] The strategy input unit has a built-in high-precision clock synchronization module. Based on the Network Time Protocol (NTP), it aligns the local timestamp of the health assessment report with the UTC standard time of the time-of-use electricity price data to eliminate cross-system time delay deviation. Then, the two types of data are resampled and linearly interpolated at a unified time granularity. Finally, it is encapsulated into a fusion strategy input dataset that includes equipment health status and electricity price economics, providing a standardized spatiotemporal aligned data source for subsequent strategy calculation.

[0131] S32. Active protection strategy unit, used to generate active risk avoidance instructions, including strong wind protection strategy and heavy snow protection strategy, based on real-time or predicted meteorological data; specifically:

[0132] S321. In the multi-source meteorological data fusion and forecasting integration stage, this unit continuously receives two types of meteorological information through parallel data channels:

[0133] One is the high-frequency measured data uploaded in real time by the on-site meteorological monitoring station of the photovoltaic power station via the Modbus RTU protocol, including the three-dimensional wind speed vector output by the ultrasonic anemometer: average wind speed, peak gust, and wind direction angle; the instantaneous snowfall intensity and accumulated snow depth collected by the rain and snow meter; and the snow cover status indication transmitted by the module backsheet temperature sensor, which determines snow cover if the temperature is below -2℃ for 30 minutes.

[0134] Secondly, it subscribes to national-level meteorological numerical forecast products via a WebSocket long connection to obtain hourly updated wind speed forecast time series curves, cumulative snowfall forecasts, and temperature change trends for the next 72 hours. This unit has a built-in data quality arbitration module that automatically switches to meteorological server data and triggers a downgrade operation flag when on-site meteorological station data is lost due to sensor failure or communication interruption.

[0135] When both sets of data are available, the Kalman filter algorithm is used for optimal estimation fusion. The meteorological station data is used to correct the systematic bias of the forecast model, and the forecast data is used to smooth the random noise of the measured data, so as to generate a more confident comprehensive meteorological situation estimate.

[0136] S322. In the dynamic protection threshold coupling calculation process, the wind protection threshold and the snow protection threshold are not fixed empirical values, but are dynamically calculated based on the real-time attitude and structural finite element analysis (FEA) results of the tracker.

[0137] The high wind threshold calculation engine incorporates a 3D structural model of the tracker, including the geometric parameters and material properties of components such as support beams, rotary reducers, columns, and foundations. It reads the current tracking angles in real time: azimuth θ and pitch φ, and calculates the bending moment M(θ,φ,v)=0.5·ρ·C generated by wind load at the critical cross-section of the structure. d (θ,φ)·A·v²·L, where: ρ is the air density, C d Here, A is the angle-dependent drag coefficient, A is the frontal area, v is the wind speed, and L is the lever arm length. When M approaches 70% of the structure's yield strength, the wind speed value corresponding to the inverse solution is used as the dynamic threshold v. threshold For example, when the tracker is at a large east-west tilt angle φ=±60°, the wind drag torque increases by 2.3 times compared to the horizontal state, and the strong wind threshold is automatically lowered from the normal 15m / s to 12.3m / s; when the tracker retracts its propellers to a wind-avoiding attitude of φ=0°, the threshold can be relaxed to 18m / s.

[0138] Calculating the heavy snow protection threshold is more complex, requiring consideration of snowfall amount S and snow density related to temperature.

[0139] ρ snow Component tilt angle φ panel And the safety factor FS of the tracker's load-bearing structure, calculate the equivalent snow load P. snow =S·ρ snow ·g·cos(φ panel ), where: g is the acceleration due to gravity; when P snow The protection mechanism is triggered when the load exceeds 60% of the structure's ultimate bearing capacity. This time threshold is set during blizzard conditions, specifically when the snow density is 150 kg / m³. 3 The following figure shows the cumulative snowfall of 8cm, while in wet snow weather, the snow density is 300kg / m³. 3 The lower threshold is reduced to 4cm to achieve adaptive adjustment of the threshold; among which, the 60% threshold is a specific manifestation of the safety factor FS, FS≈1 / 0.6, which decouples physical calculation from engineering safety margin judgment.

[0140] S323. In the protection trigger determination and instruction generation stage, the unit adopts a two-layer determination mechanism of immediate triggering when the actual exceedance exceeds the standard and early triggering when the predicted exceedance exceeds the standard.

[0141] In the real-time decision branch, the merged current wind speed v current With dynamic threshold v threshold Comparison, if v current >v threshold And if it continues for more than 3 sampling periods, it is determined that the conditions for strong wind protection are met; the real-time snowfall intensity S current With threshold S threshold Comparison, if S current >S thresholdFurthermore, the component temperature is below 0℃, which is considered to meet the conditions for heavy snow protection.

[0142] In the prediction decision branch, the rolling window method is used to scan the wind speed prediction sequence for the next 6 hours {v pred If the predicted wind speed exceeds the threshold for two consecutive hours and the confidence level is higher than 80%, the preventive strong wind protection will be triggered 30 minutes in advance to avoid structural damage due to command execution delay during the rapid wind speed rise phase.

[0143] Once the protection is triggered, the unit immediately retrieves the safety strategy library. This library pre-calculates the optimal safety angle based on the tracker model, the wind and snow pressure zoning level of the installation location, and the structural simulation results: The high wind protection strategy library stores two options: 0° horizontal paddle retraction, reducing the windward area, and 180° leeward paddle retraction, utilizing the back of the component to block the wind, which is automatically selected according to the wind direction angle; The heavy snow protection strategy library stores two options: 70° large tilt angle, utilizing gravity skiing, and 90° vertical paddle retraction, completely preventing snow accumulation, which is selected according to the snowfall type.

[0144] The generated enforcement command includes fields such as: command type, target safety angle, angular velocity limit, execution priority, trigger meteorological condition snapshot, and release criterion expression, forming a complete protection action description message.

[0145] S324. During the continuous monitoring and intelligent deactivation of the protection status, the unit does not simply go into sleep mode during protection execution. Instead, it activates a high-frequency monitoring sub-thread to continuously collect meteorological data at a 1-second interval, and calculates the 15-minute moving averages of wind speed and snowfall in real time to filter out instantaneous disturbances. The deactivation decision adopts a triple insurance mechanism of threshold backoff + duration + prediction confirmation.

[0146] The requirements for releasing the gale protection are that the real-time 15-minute average wind speed is below 80% of the dynamic threshold for 30 consecutive minutes, i.e., a safety margin is reserved, and the maximum predicted wind speed for the next 2 hours is below the threshold to prevent frequent start-stop due to repeated gusts; the requirements for releasing the heavy snow protection are that snowfall has completely stopped and the temperature has risen to above 2°C and remained above 2°C for 2 hours, while confirming through image recognition or weighing sensors that the snow load on the component surface has dropped to below 30% of the safety threshold.

[0147] Once the release conditions are met, the unit generates a release command and switches the state machine to the protection exit transition state. In this state, the tracker is first slowly adjusted to the power generation starting angle at a speed of 1° / s. After monitoring the motor current and angle feedback and finding no abnormalities, the control is officially transferred to the power generation revenue optimization unit, achieving a seamless and smooth transition from the safety mode to the economic mode, and avoiding secondary damage to sub-healthy equipment caused by the state switching impact.

[0148] S325. In the abnormal handling and degraded operation phase, the unit has a built-in fault circuit interruption mechanism: if the tracker angle feedback shows that it fails to reach the target safe angle within the specified time during the protection execution process, or the motor current continues to exceed the standard, it is determined to be a drive system failure, and an alarm for protection execution failure is immediately reported and the tracker is locked at the current position; if the meteorological station data is missing for a long time and the confidence of the meteorological service data is low, the unit switches to conservative mode and actively lowers the strong wind threshold by 20% and the heavy snow threshold by 30% to prevent unknown risks with a higher safety margin.

[0149] S33. The power generation revenue optimization unit, under the premise of no immediate safety threat to the equipment, achieves the globally optimal decision on power generation revenue within the dispatch cycle by constructing a mixed-integer optimization model that integrates equipment health constraints, electricity price economics, and the physical characteristics of photovoltaic power generation; specifically:

[0150] S331. In the real-time monitoring and unit state machine management phase of forced instruction execution, this unit and the S32 active protection strategy unit achieve microsecond-level state synchronization through a shared memory area. Specifically:

[0151] When generating the forced execution instruction, the S32 active protection strategy unit atomically sets the "force protect flag" to 0xFF in the memory-mapped file and writes the protection type code and trigger timestamp. The S33 power generation revenue optimization unit has a built-in high-priority monitoring thread that polls this flag every 10 milliseconds. Once the flag is detected, the current revenue optimization calculation process is immediately interrupted, the instruction output buffer is cleared, and the unit's state machine is switched to the SLEEP PROTECT state to prevent any conflict between power generation optimization instructions and protection instructions. After the S32 unit releases protection and clears the flag, the S33 unit automatically wakes up after a 30-second delay, enters the OPTIMIZE ACTIVE state, and restarts the revenue optimization loop. This mechanism ensures real-time state transitions and thread safety through lock-free programming and condition variable synchronization.

[0152] S332. In the equipment health status constraint extraction and dynamic mapping stage, the unit parses the real-time health status score H and the slope k of the performance degradation trend curve from the health assessment report of the strategy input unit, and generates two types of constraint boundaries through a piecewise linear mapping function.

[0153] The first category is the allowable tracking angle range [θ] min ,θ max The mapping rule is as follows: when H ≥ 80 points, the device is in a healthy state and is allowed to track up to the theoretically optimal angle θ. opt Within ±5°, i.e. θ min =θ opt -5°, θ max =θopt +5°; When 60≤H<80 minutes, the health status is good but there is a potential risk. The angle range is widened to ±10° to balance power generation and mechanical load. When H<60 minutes, the equipment enters a sub-healthy state. The angle range is compressed to ±15° and tracking is only enabled when the irradiance G>600W / m². Under low irradiance conditions, a fixed tilt angle is maintained to avoid ineffective adjustment.

[0154] The second category is the maximum angular velocity limit ω. max The mapping rule is ω max =5° / s×(H / 100) 0.5 That is, for every 20% decrease in health score, the angular velocity limit decreases by approximately 10%, for example, when H=81, ω max =4.5° / s, H=49 when ω max =3.5° / s, reducing the impact on weak mechanical components by decreasing the intensity of motion. Furthermore, for equipment with a performance degradation slope k>0.5 / day, an additional angular acceleration limit α is applied. max =1° / s², to prevent frequent steering from causing accelerated wear of the reducer.

[0155] S333. In the revenue maximization model construction and objective function design phase, the model uses a future scheduling cycle T as the optimization time domain, with T assuming a default of 4 hours, covering both peak and off-peak electricity prices. A nonlinear programming problem with N = T / Δt = 16 decision variables is constructed using a discretization step size of Δt = 15 minutes. The objective function employs differential revenue calculation accurate to a single tracker.

[0156] Max{Σ i=1 N [P] i (θ i G i , T i )·π i - C maint (H i ) - C penalty (Δθ i ) ]};

[0157] Where: P i (θ i G i , T i Let P be the power generation during the i-th time period, calculated by the photovoltaic physical model: i = η·A·G i ·[1-β·(T i -25)]·cos(θ i -θ sun_i )·SF(H i), η is the component efficiency, A is the component area, β is the temperature coefficient, θ sun_i is the solar altitude angle, SF(H i ) is the health state impact factor. When H < 70, SF = 0.95 to reflect the conversion efficiency loss of sub-healthy equipment; π i is the time-of-use electricity price in the i-th period, obtained from the strategy input unit; C maint (H i ) is the maintenance cost penalty term, designed as a piecewise function: it is 0 when H > 75, 10×(75 - H) yuan per period when 60 < H ≤ 75, and 50 + 20×(60 - H) yuan per period when H ≤ 60, reflecting the hidden failure risk cost of sub-healthy equipment; C penalty (Δθ i ) is the angle adjustment penalty term, Δθ i =|θ i -θ i-1 |, calculated as C penalty =0.01×(Δθ i / ω max ) 2 , punishing frequent and drastic adjustments to extend the mechanical life.

[0158] The model constraint conditions include: the decision variable θ i must satisfy the healthy constraint angle range [θ min , θ max ; the angle change amount Δθ i in adjacent periods needs to satisfy the angular velocity limit |Δθ i |≤ω max ·Δt; during the night period when the irradiance G i <100 W / m², force θ i =θ night to reduce wind load, such as θ night being the 0° horizontal blade folding; when the emergency fault flag bit F emerg = 1, force θ i =θ safe , θ safe being the safe blade folding angle.

[0159] S334. In the rolling optimization solution and efficient algorithm implementation link, the model solution adopts a hybrid optimization strategy: in the offline stage, use historical data to train a policy approximation network based on the long short-term memory network LSTM to learn from (H, G i ,π i ) to θ iThe mapping relationship enables rapid online pre-decision making; during the online phase, a re-optimization is initiated every 15 minutes, and the sequential quadratic programming (SQP) algorithm is used to fine-tune the pre-decision results. The convergence accuracy is set to 1e-3, and the maximum number of iterations is 50, ensuring that the solution is completed within 30 seconds.

[0160] To address the collaborative optimization problem of multiple trackers in a large-scale power plant, a Lagrange relaxation method is introduced to decouple the joint constraints. Each tracker independently solves its subproblem, and then the global power generation target is coordinated using a subgradient method, achieving distributed parallel computing. The optimization computation time for 100 trackers can be reduced to less than 5 minutes. The optimization result output is the optimal time series {θ} containing 16 target angles. i *, i=1...N}, along with the estimated expected power generation revenue R for each period. i *=P i (θ i *)·π i This is used for subsequent profit tracking and strategy evaluation.

[0161] S335. In the instruction conversion, security verification, and output stages, the generated angle time series must undergo multiple security checks before it can be converted into executable instructions:

[0162] Use the angle feasibility checker to examine each θ i *Whether it is within the current mechanical limit range and does not cause shadow occlusion with adjacent arrays, calculate the minimum spacing through a three-dimensional geometric model;

[0163] By using a rate smoothing filter and Kalman filtering to denoise the original angle sequence, a continuous and smooth angular velocity curve is generated, avoiding step shocks.

[0164] Through dynamic health status verification, if the health score H drops by more than 5 points before the instruction is executed, the issuance will be stopped and the optimization calculation will be retried.

[0165] The verified instructions are encapsulated into CANopen or Modbus TCP messages. Each instruction contains a 32-byte data segment: bytes 0-3 are UTC timestamps, bytes 4-7 are the target angle, byte 8 is the angular velocity level code, byte 9 is the maintenance priority, and bytes 10-15 are the checksum and digital signature to ensure the integrity of the instruction.

[0166] The final instruction stream is output to the S34 instruction fusion and dispatch unit, awaiting final arbitration.

[0167] S34. The instruction fusion and distribution unit, as the final decision-making gate of the control strategy, performs hard real-time priority arbitration and conflict resolution. Its arbitration rules are based on the principle of safety first and adopt priority-coded preemptive scheduling.

[0168] When unit S32 outputs a forced execution command, regardless of whether unit S33 has already generated an angle adjustment command, the arbitrator immediately interrupts the power generation optimization process, unconditionally transfers the execution right to the protection command, and records the status of the interrupted optimization command for recovery.

[0169] The arbitrator has a built-in conflict detection submodule. When there is no active protection command, it performs a feasibility check on the angle adjustment command output by S33: checks whether the target angle exceeds the allowable range of the current health status, verifies whether the angular velocity violates the mechanical limit, and verifies whether the command timestamp matches the actual response delay of the tracker. If the command requires a 10° adjustment to be completed within 5 seconds, but the maximum speed of the driver is only 2° / s, it is determined to be a conflict command and an alarm is triggered.

[0170] The verified commands are addressed and encapsulated according to the RS-485 bus address of the target tracker NCU or TCU, generating a complete message frame conforming to the Modbus RTU or NCU private protocol, including the slave address, function code, register start address, target angle code value, and CRC checksum. When commands are sent through the IoT gateway, a dual-route redundancy mechanism is enabled: the primary path is the 4G / 5G public network, and the backup path is the Zigbee local mesh network or BeiDou short message service, ensuring reliable delivery of control commands under extreme communication conditions. Finally, after being parsed by the remote control execution module, the tracker control commands drive the servo driver of the NCU / TCU unit to complete the angle closed-loop adjustment, realizing an end-to-end closed loop from strategy decision-making to physical action.

[0171] S4. Remote control execution module, used to send tracker control commands to the corresponding NCU or TCU unit to adjust the tracker angle; specifically including:

[0172] S41. The instruction receiving and verification unit, as the entry point for control instructions, implements a two-layer verification mechanism to ensure the integrity and logical rationality of the instructions; specifically:

[0173] At the integrity verification level, the unit first parses the raw data packets received from the instruction fusion and distribution unit to verify whether they conform to the predefined message frame structure specifications: the message header must contain a fixed 0xAA55 synchronization word modulus, followed by a 2-byte protocol version number, a 4-byte target device ID, a 4-byte instruction sequence number, and a 1-byte instruction type code; the checksum verification uses the CRC-16 / CCITT-FALSE algorithm to perform cyclic redundancy calculation on the message payload data, and compares the calculation result with the 2-byte checksum field attached to the end of the message. If the two are inconsistent, it is determined that the transmission is corrupted and is directly discarded; the sequence number verification checks whether the current instruction sequence number is strictly equal to the sequence number of the last successfully executed instruction of the target NCU or TCU unit + 1, to prevent repeated execution of instructions or timing errors caused by network retransmission or out-of-order delivery. If the sequence number jumps beyond the allowed window, an anti-replay attack alarm is triggered and the instruction source is requested to retransmit the synchronization frame.

[0174] At the logical verification level, the unit is based on the allowable tracking angle range [θ] mapped in the target device's most recent health assessment report. min ,θ max The unit performs boundary checks on the target angle field in the instruction. If the target angle exceeds this range, it is determined to be an invalid instruction and an angle out-of-bounds error code is returned. Simultaneously, the unit maintains a timestamp-based instruction issuance timing window, verifying the reasonableness of the instruction issuance timing in conjunction with arbitration rules. For example, if any non-protection type angle adjustment instruction is detected within the 300-second mandatory protection window after the S32 active protection strategy unit triggers strong wind protection, regardless of its content, it is determined to be a logical error violating the arbitration rules, rejected, and the violation log is recorded. For instructions that pass the double-layer verification, the unit appends a verification pass identifier code 0x7E7E to its header and forwards it to the S42 unit.

[0175] S42. The communication protocol adaptation and encapsulation unit acts as a translator for cross-network heterogeneous communication. It maintains a dynamically updatable protocol mapping library. This library stores metadata such as the network type, physical layer interface, transport layer protocol, application layer protocol, and maximum transmission unit size registered by each NCU or TCU unit, using the target device ID as the key. Upon receiving a valid general instruction, the unit first queries the protocol mapping library to obtain the complete communication profile of the target device.

[0176] For the Zigbee protocol, the target angle floating-point value in the general instructions is converted into IEEE 754 single-precision format, encapsulated into the Zigbee APS application support sublayer data frame, the transmission mode of the frame control field is set to 0b11, i.e. reliable transmission + authentication, the cluster ID points to 0x0007, i.e. tracker control cluster, and the 16-bit network layer frame sequence number and message integrity check code MIC are calculated according to the Zigbee Pro specification, and finally a data frame that conforms to the requirements of the Zigbee 3.0 protocol stack is generated.

[0177] For 4G or 5G protocols, the general instructions are first encapsulated in JSON format, and then published to the topic via PUBLISH messages of the MQTT protocol. The QoS level is set to 2, and TLS 1.3 encryption and two-way certificate authentication are enabled at the TCP layer to ensure the security of public network transmission.

[0178] For the BeiDou short message protocol, the instructions need to be compressed to the extreme according to the 70 Chinese character capacity limit of the RDSS service: the device ID is mapped to a 2-byte short address, the target angle is quantized to a 2-byte fixed-point number, the instruction type is encoded as a 4-bit half-byte, and after being encapsulated in a custom compact binary format, it is embedded into the user data segment of the BeiDou short message, and the service type identifier 0x01 and the message sequence number are filled in the message header, and finally a data frame conforming to the BeiDou civilian short message communication protocol is generated.

[0179] After the protocol encapsulation is completed, the unit appends a protocol type identifier to the end of the data frame so that the S43 unit can select the physical link.

[0180] S43. The instruction issuing unit selects the corresponding physical communication link to implement differentiated transmission based on the protocol identifier of the data frame.

[0181] For Zigbee protocol frames, the unit connects to the Zigbee gateway cluster manager at the power plant site via a socket. This cluster consists of three Zigbee gateways that are hot-standby for each other, covering different frequency bands to avoid interference. The unit selects the optimal gateway based on the routing table information of the target device. If the target NCU / TCU and the gateway are in the same mesh network cluster, the data frame is sent directly via unicast with MAC layer acknowledgment enabled. If cross-cluster transmission is required, multi-hop forwarding is performed using the mesh routing protocol between gateways. In terms of transmission strategy, emergency protection commands are sent using a dual-channel redundant transmission of broadcast and unicast to ensure a high delivery rate in low signal-to-noise ratio environments.

[0182] For 4G or 5G protocol frames, the unit sends MQTT messages to the MQTT Broker on the cloud platform or local edge server through a dedicated network interface isolated by the Linux network namespace. The Broker routes the messages to the 4G / 5G communication module built into the target TCU / NCU through the operator's core network. To cope with public network jitter, the unit implements a timeout retransmission mechanism at the application layer, with an initial timeout of 5 seconds, increasing exponentially with a backoff strategy, and a maximum of 3 retries.

[0183] For BeiDou protocol frames, the unit connects to the BeiDou RDSS terminal in the main control room of the power station via a serial port, sends AT commands at a baud rate of 9600, writes short messages into the terminal's transmission queue, and the terminal sends the messages to the BeiDou GEO satellite via an L-band antenna. The satellite then forwards the messages to the ground center station, which finally sends them to the BeiDou user terminal of the target device through the BeiDou short message data channel.

[0184] After sending, the unit initiates an independent blocking wait to receive the SBDIX command response returned by the Beidou terminal, confirming that the message has successfully entered the satellite transmission queue. If no response is received within 3 minutes, it is determined that the transmission has failed and an alarm is triggered.

[0185] S44. The execution feedback acquisition unit constructs a closed-loop verification and exception handling mechanism for instruction execution. After a command is successfully sent, the unit immediately creates an entry in the memory hash table with the command sequence number as the key and a feedback status structure as the value. The structure includes the target device ID, sending timestamp, target angle, feedback wait timer, retry count counter, and execution status enumeration. The unit starts a feedback listening thread, receiving asynchronous status reports from the NCU / TCU through the reverse parsing function of unit S42: For Zigbee networks, it listens for APS confirmation frames from the gateway and cluster messages indicating angle execution completion or failure actively uploaded by the device; for 4G / 5G networks, it subscribes to MQTT topics to receive JSON-formatted feedback messages uploaded by the device; for BeiDou networks, it polls the RDSS terminal receive buffer and reads the short confirmation messages returned by the device using the AT+SBDRB command.

[0186] The parsed feedback data includes the following fields: device address, actual execution angle θ. actual Status codes: success, motor fault, angle overtravel, communication timeout; execution completion timestamp; and real-time values ​​of current motor current and driver temperature. The unit will then display θ. actual With respect to the target angle θ targetA deviation check is performed, with an allowable error range of ±0.5°. If the deviation is within the range and the status code is successful, the hash table entry status is updated to SUCCESS, and the actual angle and electrical parameters of the device are updated from the TCU / NCU real-time status database for the data analysis module to use in the next cycle's health assessment. If the feedback waiting timer times out, or the received status code is unsuccessful, or the angle deviation exceeds the limit, it is determined as an execution anomaly.

[0187] The exception handling sub-process first checks whether the number of retries is less than the preset limit. If it is not exceeded, the retry interval is calculated according to the exponential backoff strategy, the S43 unit is called again to issue the instruction, and the retry counter is incremented by 1. If the exception still occurs after the retry or the retry limit has been reached, the status is set to FAILED, and a structured alarm event is generated: the alarm ID is generated using UUID, the device address is accurate to the physical MAC address of the NCU / TCU, the exception type is encoded and mapped to the specific fault: motor stall, encoder failure, communication interruption, timestamp accurate to milliseconds, and a snapshot of the original instruction and the last feedback data is attached.

[0188] The alarm event is asynchronously pushed to the AI ​​fault diagnosis unit of the data analysis module via the message queue, triggering online diagnosis as a new fault sample. At the same time, it is pushed to the operation and maintenance alarm system, driving the work order system to automatically create an emergency repair work order.

[0189] For commands that are determined to be executed successfully, the feedback data is cleaned by the data preprocessing unit and then fed back into the equipment's real-time status database. This data is used to update status variables such as tracker angle and motor load rate, forming a control closed loop. This ensures that the data analysis module always generates health assessment and optimization strategies for the next cycle based on the latest actual equipment operating data.

[0190] Example 2

[0191] To verify the technical effectiveness and advancement of this invention, the following simulation experiments were designed and compared with existing typical technical solutions.

[0192] 1. Experimental Environment and Configuration

[0193] The simulation covers a large-scale photovoltaic power plant located in complex mountainous terrain, with an installed capacity of 100MW, employing bifacial modules and single-axis tracking brackets, and comprising approximately 20,000 NCU control units. The simulation operates for a full year, encompassing seasonal changes, typical severe weather, and time-of-use electricity price fluctuations in the electricity market.

[0194] Comparison of options:

[0195] Option A: Deploy the AI ​​tracker intelligent platform described in this invention, adopting a private local deployment mode for power plants, and enable AI health diagnosis, proactive protection, and revenue optimization functions.

[0196] Option B: Adopt a mainstream general photovoltaic monitoring system on the market, which has basic SCADA data monitoring, alarm and group control functions, and relies on planned maintenance.

[0197] Option C: Use the supporting software provided by a tracking bracket manufacturer to mainly achieve status monitoring and simple high wind protection, without AI diagnosis and fine-grained revenue optimization.

[0198] 2. Experimental Design and Comparison of Key Indicators

[0199] To address the core innovations of this invention, the following key experiments were conducted for comparison:

[0200] Experiment 1: Comparison of Operation and Maintenance Intelligence Levels, i.e., Predictive Maintenance vs. Scheduled / Reactive Maintenance

[0201] Comparison indicators Solution A (This invention) Option B (General Platform) Option C (Simplified Platform) Mean Time to Find Faults <2 hours 24-72 hours ≥48 hours Fault prediction accuracy 92% 0% 0% Number of unplanned downtimes 3 times / year 11 times / year 15 times / year Annual maintenance and inspection frequency 4 times 12 times 6 times Simulated annual power generation loss (due to faults) Approximately 0.8% Approximately 3.5% Approximately 4.9%

[0202] Experimental Conclusion 1: The AI ​​health diagnosis function of this invention transforms the operation and maintenance mode from passive response to proactive prediction, significantly shortens the fault detection time, greatly reduces the number of unplanned shutdowns and the resulting power generation losses, and effectively reduces the frequency of unnecessary on-site inspections.

[0203] Experiment 2: Comparison of Power Generation Revenue Optimization Capabilities

[0204] Comparison indicators Solution A (This invention) Option B (General Platform) Option C (Simplified Platform) Control command issuance delay <1 second 3-5 seconds 1-3 seconds Tracking angle control accuracy ±0.1° ±0.5° ±0.3° Does it support time-of-use pricing optimization? yes Partial support no Simulated annual power generation revenue improvement The baseline performance represents a 5.8% improvement over Option B and an 8.2% improvement over Option C. Low, relying solely on tracking itself, without electricity price optimization. Low performance, simplistic strategy, lack of optimization

[0205] Conclusion 2: This invention benefits from deep integration and optimization at the hardware level, achieving ultra-low latency and high-precision control. More importantly, by incorporating time-of-use electricity price data into a revenue maximization model, it can proactively adjust the power generation curve to match high-price periods, thereby obtaining significantly higher power generation revenue under the same illumination conditions.

[0206] Experiment 3: Comparison of System Safety and Reliability, i.e., Active Protection and Closed-Loop Control

[0207] Comparison Scenes Solution A (This invention) Option B (General Platform) Option C (Simplified Platform) High wind protection trigger speed Real-time trigger, execution time <10 seconds Delayed trigger, 30-60 seconds Triggers relatively quickly, <15 seconds Snow protection strategy have none none Instruction execution feedback closed loop Complete closed loop Open-loop or weak feedback Partial feedback Equipment damage rate under severe weather conditions in simulation 0% 0.5% 0.2%

[0208] Experimental Conclusion 3: The active protection strategy unit and execution feedback acquisition unit constructed in this invention form a multi-layered security protection and reliable execution closed loop. Not only is the response speed fast, but the strategy is also more intelligent, and feedback ensures that instructions are executed correctly, greatly reducing equipment safety risks under extreme weather conditions.

[0209] Experiment 4: Adaptability and Autonomous Controllability in Complex Scenarios

[0210] Network disconnection scenario: Simulate a 4-hour public network outage. Solution A can seamlessly switch to BeiDou short message service for critical command and status data transmission, ensuring basic monitoring and control; Solution B is completely interrupted; Solution C allows local control to be maintained, but remote monitoring is disabled.

[0211] Data security and self-control: Solution A uses a fully domestically produced software stack, with core data physically isolated in a local deployment mode. Solutions B and C rely on foreign commercial databases or operating systems, posing potential supply chain and information security risks.

[0212] 3. Conclusion: This simulation experiment fully verifies the AI ​​tracker intelligent platform provided by this invention. It has outstanding effects that existing technologies cannot match in terms of improving the power generation revenue of photovoltaic power plants, reducing operation and maintenance costs, ensuring equipment safety, enhancing system reliability, and achieving independent control of technology, which fully demonstrates its innovation and commercial application value.

[0213] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects:

[0214] This has enabled an intelligent transformation of the operation and maintenance (O&M) model and a refined improvement in power generation revenue: By introducing AI-powered health diagnosis and predictive maintenance models, the traditional passive and planned O&M model is transformed into proactive predictive maintenance based on the real-time health status of equipment, thereby significantly reducing O&M costs and power generation losses. Furthermore, by constructing a revenue optimization model that integrates time-of-use pricing and equipment status, the power generation curve is dynamically adjusted to match peak electricity market periods while ensuring safe equipment operation, achieving a fundamental shift from pursuing power generation volume to pursuing power generation revenue.

[0215] An active safety protection and high-reliability control system based on prediction and closed-loop design was constructed: By designing an active protection strategy unit integrating meteorological forecasting, early warning and rapid automatic avoidance of severe weather such as strong winds and heavy snow were achieved, significantly improving the operational safety of equipment in complex environments. By establishing a closed-loop command execution system that includes full-process verification and feedback confirmation, reliable delivery and accurate execution of control commands were ensured, effectively solving the uncertainty problem in remote control and greatly enhancing the robustness of the overall system.

[0216] It provides a flexible, adaptable, and independently controllable systematic solution: by supporting a dual-mode deployment architecture (local-cloud) and deeply integrating BeiDou communication, it can simultaneously meet the differentiated needs for data security, real-time performance, and network coverage in different scenarios, thus expanding the platform's applicability. By adopting a fully domestically produced software technology stack, it achieves independent control of the core software system, effectively safeguarding the network security and industrial chain resilience of energy infrastructure.

[0217] The embodiments and / or implementation methods described above are merely preferred embodiments and / or implementation methods for implementing the technology of the present invention, and are not intended to limit the implementation methods of the technology of the present invention in any way. Any person skilled in the art can make some modifications or alterations to other equivalent embodiments without departing from the scope of the technical means disclosed in the present invention, but these should still be regarded as the technology or embodiments that are substantially the same as the present invention.

[0218] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. An AI tracker intelligent platform, characterized in that, include: The data acquisition module is used to collect operational and environmental data from the NCU or TCU unit of the photovoltaic tracker; The data analysis module is used to receive the operational data and environmental data, perform fault diagnosis and status assessment through AI models, and generate a health assessment report. The control strategy generation module is used to generate tracker control instructions based on the health assessment report and time-of-use electricity price data; The remote control execution module is used to send the tracker control commands to the corresponding NCU or TCU unit to adjust the tracker angle.

2. The AI ​​tracker intelligent platform according to claim 1, characterized in that, The data acquisition module specifically includes: The equipment data acquisition unit is used to collect real-time operating status data and electrical parameters, including tracker angle, motor current and voltage, drive status and fault codes, through the sensors built into the NCU or TCU unit, and to perform preliminary structured packaging of the collected data. The environmental data acquisition unit is used to connect to the meteorological station deployed in the photovoltaic power station or access the authoritative meteorological data service interface to collect environmental data and geographic information data, including irradiance, ambient temperature, wind speed and direction, precipitation and snowfall, and component temperature. The protocol adaptation unit is used to dynamically select or enable communication protocols in parallel according to the power plant deployment mode and network conditions: in the local area network environment, the Zigbee protocol or LoRa protocol is enabled first; when remote public network communication is required, the 4G or 5G mobile network protocol is enabled; in scenarios where there is no public network signal or there are strict requirements for communication autonomy, the Beidou short message communication protocol is enabled.

3. The AI ​​tracker intelligent platform according to claim 1, characterized in that, The data analysis module specifically includes: The data preprocessing unit is used to clean, normalize, and extract features from the operational data and environmental data from the data acquisition module to form a structured analysis dataset. The AI ​​fault diagnosis unit is used to receive the structured analysis dataset and identify potential fault types and abnormal operating modes of the NCU or TCU unit through a pre-trained AI analysis model. The status assessment unit is used to assess the real-time health status and performance degradation trend of the equipment based on the potential fault types and abnormal operating modes, combined with the equipment's historical operating data. The report generation unit is used to summarize the potential fault types, abnormal operating modes, real-time health status and performance degradation trends, generate a health assessment report including a fault warning list, a health status overview and maintenance priority suggestions, and output it to the control strategy generation module.

4. The AI ​​tracker intelligent platform according to claim 3, characterized in that, The workflow of the AI ​​fault diagnosis unit specifically includes: The system receives a structured analysis dataset from the data preprocessing unit and calls a pre-trained machine learning model, which is trained using historical fault data and integrates equipment operation mechanisms with data-driven modes. The structured analysis dataset is input into the machine learning model. The machine learning model first performs sliding window feature extraction on the time series data to obtain time series feature vectors. Then, it performs pattern matching and probability calculation based on the time series feature vectors, outputs the confidence probability for each type of preset fault, and marks faults with probabilities exceeding a set threshold as potential fault types. For operational data that does not directly match the preset fault type but whose feature vector deviates significantly from the normal baseline, it is marked as an abnormal operation mode.

5. The AI ​​tracker intelligent platform according to claim 4, characterized in that, The workflow of the state assessment unit specifically includes: The system receives potential fault types, corresponding confidence probabilities, and abnormal operating modes from the AI ​​fault diagnosis unit. Based on the potential fault types and abnormal operating modes, it queries a pre-set equipment impact weight library to determine the quantitative impact coefficient of each fault or abnormality on the overall health of the equipment. The system retrieves the historical performance baseline of the device, compares the current operating data with the corresponding historical performance data under the same conditions, and calculates the performance deviation rate. Based on the quantified impact coefficient, confidence probability, and performance deviation rate, the data are input into a health index-based degradation model for calculation, outputting a quantified real-time health status score, and generating a performance degradation trend prediction curve based on time-series data analysis.

6. The AI ​​tracker intelligent platform according to claim 1, characterized in that, The control strategy generation module specifically includes: The strategy input unit is used to receive health assessment reports, synchronously access the time-of-use electricity price data stream released by the electricity market, and perform timestamp alignment and formatted encapsulation on both. The active protection strategy unit is used to generate active risk avoidance instructions, including strong wind protection strategy and heavy snow protection strategy, based on real-time or predicted meteorological data. The power generation revenue optimization unit is used to integrate the equipment status information and time-of-use electricity price data in the health assessment report, and calculate and generate tracker angle adjustment instructions for optimizing power generation revenue through the revenue maximization model. The command fusion and issuance unit is used to perform priority arbitration and conflict resolution on the active risk avoidance command and angle adjustment command, generate the final tracker control command, and output it to the remote control execution module.

7. The AI ​​tracker intelligent platform according to claim 6, characterized in that, The workflow of the active protection strategy unit specifically includes: It continuously receives formatted meteorological data from the strategy input unit, as well as short-term forecast meteorological data from the meteorological service terminal; The real-time wind speed data is compared with the gale protection threshold, and the real-time snowfall data or predicted snowfall data is compared with the heavy snow protection threshold. When the real-time wind speed exceeds the gale protection threshold, or the real-time snowfall or predicted snowfall exceeds the heavy snow protection threshold, it is determined that active protection needs to be triggered. Based on the type of protection triggered, the corresponding preset safety angle is retrieved from the preset safety policy library. The safety angle is determined based on the tracker's structural strength and local wind pressure and snow load model calculations; a forced execution command containing the target safety angle and with the highest execution priority is generated. The system monitors weather conditions in real time until wind speed or snowfall data falls below a safe threshold and remains below it for a preset duration. Then, it generates a command to release the protection status and returns control to the power generation revenue optimization unit.

8. The AI ​​tracker intelligent platform according to claim 7, characterized in that, The workflow of the power generation revenue optimization unit specifically includes: The system continuously monitors whether any forced execution commands are triggered. If not, it receives a health assessment report and real-time and predicted time-of-use electricity price data from the policy input unit. It extracts the real-time health status score and performance degradation trend from the health assessment report and determines the allowable tracking angle range and maximum angular velocity limit under the current device status based on a predefined mapping relationship. The allowable tracking angle range, maximum angular velocity limit, real-time irradiance data, and time-of-use electricity price data are used as boundary conditions and input parameters, and then input into the profit maximization model. The revenue maximization model aims to maximize the expected power generation revenue within a future scheduling cycle. Based on the photovoltaic power generation physical model and electricity price time series, it performs rolling optimization calculations to solve for the optimal tracker angle time series. The optimal tracker angle time series is converted into specific, timestamped angle adjustment instructions, and the angle adjustment instructions are output to the instruction fusion and distribution unit.

9. The AI ​​tracker intelligent platform according to claim 1, characterized in that, The remote control execution module specifically includes: The instruction receiving and verification unit is used to receive tracker control instructions from the instruction fusion and sending unit, and verify the legality of their format and the validity of the target device address. The communication protocol adaptation and encapsulation unit is used to encapsulate the verified tracker control commands into data frames of the corresponding protocol according to the network type and communication protocol accessed by the target NCU or TCU unit. The instruction issuing unit is used to send the data frame to the target NCU or TCU unit through the corresponding physical communication link, driving the tracker to perform an angle adjustment action; The execution feedback acquisition unit is used to receive execution status feedback data from the NCU or TCU unit in real time after the command is issued, and return it to the data analysis module to update the equipment status and form a control closed loop.

10. An AI tracker intelligent platform according to claim 9, characterized in that, The workflow of the execution feedback acquisition unit specifically includes: After the instruction issuing unit successfully sends a data frame, it immediately starts a feedback waiting timer for the current instruction and marks the instruction as being in execution; it listens for and receives the execution status feedback data after being reverse-parsed by the communication protocol adaptation and encapsulation unit; The received feedback data is matched and verified with the original command issued. If the deviation between the actual execution angle and the target angle is within the allowable error range and the status code is success, the command is determined to have been executed successfully. If no feedback is received before the feedback waiting timer expires, or the feedback status code is failure, or the actual execution angle deviation exceeds the limit, it is determined that the instruction execution is abnormal. For instructions that are determined to be executed successfully, the execution status mark is cleared, and the execution status feedback data containing the actual execution result is packaged and sent back in real time to the data preprocessing unit of the data acquisition module and the data analysis module for updating the real-time status database of the device. For instructions that are determined to be abnormal, the instruction resending process is triggered according to the preset retry policy; if the error still occurs after retry, an alarm event containing the device address, the type of error, and a timestamp is generated and pushed to the AI ​​fault diagnosis unit and the operation and maintenance alarm system of the data analysis module.

Images