A new energy power facility detection data management and analysis platform

Abstract

The application discloses a new energy power facility detection data management and analysis platform, comprising: a multi-source heterogeneous high-frequency synchronous induction system, each sensor node of which is equipped with a Beidou timing module, so that sampling error is controlled within 20ns by taking a second pulse as a sampling trigger source and adding a time stamp; an environment-driven dynamic reference flow form construction system for extracting low-dimensional embedding coordinates of environment characteristics through nonlinear flow form learning and constructing an ideal output hyper surface changing with the environment; and an endogenous performance deviation decoupling extraction system for calculating a residual matrix of real-time operation characteristics and ideal operation states and adopting an independent component analysis algorithm to decouple and extract an endogenous attenuation characteristic vector representing equipment intrinsic performance attenuation from the residual matrix. The application changes the monitoring reference from a fixed threshold to a dynamic baseline fluctuating with the environment, effectively separates environmental interference and equipment real performance attenuation, and improves diagnostic accuracy and preventive maintenance capability.

Description

Technical Field

[0001] This invention belongs to the field of power system informatization and data processing technology, specifically relating to a new energy power facility detection data management and analysis platform. Background Technology

[0002] With the profound transformation of the global energy structure, new energy power facilities, represented by photovoltaic power generation and wind farms, occupy a core position in the power system. To ensure the safe and stable operation of the power grid, digital management and intelligent analysis of widely distributed and environmentally complex new energy facilities have become an important research direction in the field of the energy internet. This process involves real-time collection and unified modeling of operational data from power generation equipment, transmission and transformation facilities, and environmental compensation devices, aiming to build a state monitoring system covering the entire lifecycle of power facilities, providing fundamental data support for the coordinated optimization and resource scheduling of the power system.

[0003] Among them, the New Energy Power Facility Monitoring Data Management and Analysis Platform, as a key carrier of digital operation and maintenance, is mainly responsible for integrating operational monitoring data from different geographical locations and heterogeneous equipment. By integrating multi-source sensor information and communication protocols, this platform can monitor key physical quantities of power facilities in real time, such as voltage, current, speed, and temperature, and use data analysis to assess the operational health status of the equipment. Its core objective is to establish equipment operation records and, through comprehensive comparison of historical and real-time data, achieve a quantitative evaluation of the operational efficiency of power facilities and preliminary identification of abnormal conditions.

[0004] Existing new energy power equipment monitoring and management platforms typically employ fixed threshold-based monitoring schemes, identifying faults by setting alarm upper or lower limits for key parameters. However, this approach faces significant technical bottlenecks in practical applications. Because the operating status of new energy facilities is heavily influenced by drastically fluctuating environmental variables such as light intensity, ambient temperature, and wind speed, traditional threshold methods struggle to effectively distinguish between normal operating fluctuations caused by environmental changes and performance degradation due to component aging, dust accumulation, or mechanical fatigue. This monitoring model lacks the necessary feature identification capabilities when dealing with the coupling relationship between environmental noise and equipment performance degradation, resulting in significant lag in detection results. Alarms are often triggered only after substantial equipment damage. Furthermore, the lack of dynamic modeling that deeply correlates environmental variables with operational data prevents the system from achieving true preventative maintenance, significantly increasing operation and maintenance costs and the risk of unplanned downtime. Summary of the Invention

[0005] The purpose of this invention is to provide a data management and analysis platform for the testing of new energy power facilities to solve the above-mentioned technical problems.

[0006] The technical solution of the present invention includes: A new energy power facility monitoring data management and analysis platform, comprising: The multi-source heterogeneous high-frequency synchronous sensing system is used to synchronously collect environmental, electrical, and structural characteristic parameters of new energy power facilities under a unified clock reference. Each sensor node in the multi-source heterogeneous high-frequency synchronous sensing system is equipped with a Beidou time synchronization module. The Beidou time synchronization module is used to output a second pulse signal as a sampling trigger source to control the sampling time error within 20 nanoseconds and to append standard time information as a timestamp to the sampling data. An environment-driven dynamic benchmark manifold construction system is used to receive environmental feature parameters, extract low-dimensional embedded coordinates of environmental features through a nonlinear manifold learning algorithm, and construct an ideal output hypersurface with low-dimensional embedded coordinates as input and ideal equipment output as output based on historical health operation data. The endogeneous performance deviation decoupling extraction system is used to receive the instantaneous ideal operating state point of electrical characteristic parameters and ideal output hypersurface output, calculate the residual matrix between real-time operating characteristics and ideal operating state, and use independent component analysis algorithm to decouple and extract at least one endogeneous attenuation feature vector with clear physical meaning from the residual matrix. The endogeneous attenuation feature vector is used to characterize the performance degradation of the equipment body due to dust accumulation, wear or aging.

[0007] Preferably, the multi-source heterogeneous high-frequency synchronous sensing system includes: The environmental sensing subsystem integrates a total radiation sensor, an ultrasonic three-dimensional anemometer, and a high-precision platinum resistance thermometer to acquire irradiance, wind speed, and temperature as environmental characteristic parameters. The power operation monitoring subsystem is used to collect current and voltage as electrical characteristic parameters at a sampling frequency of 12.8 kHz. The power operation monitoring subsystem includes a field-programmable gate array (FPGA), which is internally configured with a dual-port random access memory (DRAM). The DRAM is divided into two equal-length buffers and uses a ping-pong operation mechanism to ensure that while the analog-to-digital converter writes data to one buffer, the digital signal processor reads data from the other buffer for calculation, thus ensuring that the data acquisition process is continuous and uninterrupted. The structural feature sensing subsystem, deployed at key stress locations, includes vibration acceleration sensors, infrared thermal imaging sensors, and acoustic emission sensors, used to acquire vibration signals, temperature data, and acoustic emission signals as structural feature parameters.

[0008] Preferably, the environment-driven dynamic baseline manifold construction system further includes: The normalization and weighting module is used to perform min-max normalization on the received multidimensional environmental feature parameters, and dynamically weight the normalized environmental variables based on a pre-defined weight vector distinguished by season and geographical location type. The manifold learning module is used to reduce the weighted high-dimensional environmental features to three dimensions by constructing a neighborhood graph and calculating geodesic distances using the isometric mapping algorithm, thus obtaining low-dimensional embedded coordinates. The hypersurface generation module takes historical low-dimensional embedded coordinates during periods when the device is in a healthy state as input and the corresponding actual output of the device as output, and trains it using a random forest regression model to generate an ideal output hypersurface; and... The interpolation module is used to locate the corresponding instantaneous ideal operating state point on the ideal output hypersurface in real time for the current input low-dimensional embedded coordinates using a radial basis function interpolation algorithm.

[0009] Preferably, the endogeneity performance deviation decoupling extraction system further includes: The phase alignment module is used to perform nonlinear time axis alignment on the real-time running feature matrix and the ideal running state matrix composed of instantaneous ideal running state points using a dynamic time warping algorithm, so as to eliminate millisecond-level time phase deviations in the data transmission link. The residual calculation module is used to calculate the difference between the real-time running feature matrix after phase alignment and the ideal running state matrix, generating a residual matrix; and, The feature separation module is used to perform blind source separation on the residual matrix using a fast independent component analysis algorithm to obtain multiple statistically independent source signals. By analyzing the time-domain waveform and spectral energy distribution of each source signal, it classifies and labels them as dust accumulation attenuation features, bearing wear features, or capacitor aging features, and combines them to form an endogenous attenuation feature vector.

[0010] Preferably, it also includes a multi-physics coupled state comprehensive diagnostic system, which integrates the following: The dust accumulation evolution analysis module is used to establish a regression model of photocurrent decline rate and cleaning cycle based on the dust accumulation decay characteristics in the endogenous decay feature vector and the irradiance data collected by the environmental sensing subsystem, in order to estimate the current dust accumulation density and automatically generate cleaning suggestions based on the comparison of power generation loss prediction and cleaning cost. The electrical physical performance evaluation module is used to derive the equivalent junction temperature of power semiconductor devices based on the current and voltage data in the electrical characteristic parameters, the shell temperature measured by the infrared thermal imaging sensor, and the thermal resistance characteristic parameters provided in the device manual, using a thermal resistance model, and triggering an aging warning when the equivalent junction temperature exceeds a preset threshold. The structural dynamics diagnostic module is used to extract microcrack propagation characteristic indicators based on vibration signals and acoustic emission signals in structural characteristic parameters through empirical mode decomposition and Hilbert transform, and to comprehensively derive fatigue damage indicators by calculating the centroid frequency offset of the vibration spectrum.

[0011] Preferably, the multi-physics coupled state comprehensive diagnostic system also includes a fuzzy comprehensive evaluation module, which takes the dust density output by the dust accumulation evolution analysis module, the aging risk index output by the electrical physical performance evaluation module, and the fatigue damage index output by the structural dynamics diagnostic module as inputs, performs nonlinear fusion through a preset membership function and fuzzy inference rule base, and outputs a comprehensive health index between 0 and 1 after defuzzification to quantitatively assess the overall health status of the equipment.

[0012] Preferably, it also includes a whole lifecycle preventive maintenance decision-making system, which is used for: Receive the comprehensive health index and establish a stochastic degradation model describing the evolution of health status over time based on the Wiener process; The drift coefficient and diffusion coefficient of the degradation model are updated online based on the newly acquired health index observations using a Bayesian method. Predict the remaining lifespan of the equipment based on the updated degradation model; Set multi-level early warning thresholds associated with the comprehensive health index, and implement differentiated response strategies for different levels of early warning, including increasing sampling frequency, initiating maintenance cost assessment, and automatically generating maintenance instructions that include a list of repair parts and spare parts.

[0013] Preferably, the whole lifecycle preventive maintenance decision-making system activates a maintenance cost assessment mechanism when executing a Level 2 warning; the maintenance cost assessment mechanism includes: Obtain meteorological forecast data, time-of-use electricity price curves, and equipment efficiency reduction factors determined by the current comprehensive health index; For multiple candidate maintenance dates within the next M days, calculate the expected net maintenance benefit for each candidate maintenance date. The expected net maintenance benefit is equal to the cumulative power generation loss avoided during the waiting period until the maintenance is carried out on the candidate maintenance date minus the corresponding maintenance cost. Among them, the avoidable cumulative power generation loss is determined based on the predicted equivalent full-load hours, time-of-use electricity price curve, and the corresponding relationship between the equipment efficiency reduction coefficient and the health index decay in the meteorological forecast data. The candidate maintenance date that maximizes the expected net maintenance benefit is determined as the optimal maintenance window, and a recommendation for this optimal maintenance window is output.

[0014] Preferably, it also includes a data management center, which adopts a layered and decoupled architecture, including: A distributed time-series database is used to store high-frequency sampled environmental, electrical, and structural characteristic parameters, and to set the thermal data retention period. Relational databases are used to store static ledger information of devices and metadata generated by algorithm models; The publish-subscribe mechanism based on message queues is used to realize data exchange and loosely coupled communication between various system modules; Hardware encryption chips are configured at key transmission nodes to sign and encrypt transmitted data packets according to the national standard SM2 algorithm.

[0015] Preferably, the environment-driven dynamic benchmark manifold construction system also includes a self-correction mechanism. This mechanism is used to automatically trigger a self-supervised learning program when a systematic deviation that cannot be explained by the existing model is detected in the residual sequence output by the endogeneous performance deviation decoupling extraction system. The program identifies new operating conditions through cluster analysis and uses incremental learning techniques to fine-tune the model in the local area of ​​the ideal output hypersurface to adapt to the operating condition drift caused by environmental changes.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention achieves nanosecond-level synchronous acquisition of environmental, electrical, and structural characteristic parameters under a unified clock reference through a multi-source heterogeneous high-frequency synchronous sensing system, providing a high-fidelity and highly consistent data foundation for subsequent accurate analysis and significantly improving the reliability and effectiveness of multi-source data fusion processing.

[0017] 2. This invention, through the synergistic effect of an environment-driven dynamic benchmark manifold construction system and an endogenous performance deviation decoupling extraction system, transforms the monitoring benchmark from a fixed threshold into a dynamic baseline that fluctuates with the environment. This effectively eliminates the interference of environmental factors on the equipment's operating status, and can accurately identify and quantify the performance degradation of the equipment body caused by dust accumulation, wear, or aging, significantly improving the accuracy of equipment health diagnosis and early warning capabilities.

[0018] 3. This invention integrates a multi-physics field coupled state comprehensive diagnostic system with a full life cycle preventive maintenance decision system, realizing a complete closed loop from decoupled identification of equipment performance degradation, multi-dimensional health status comprehensive assessment to remaining life prediction and optimal maintenance window decision-making. It transforms the operation and maintenance mode from passive post-event maintenance to proactive predictive maintenance, effectively reducing the risk of unplanned downtime and overall operation and maintenance costs. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the new energy power facility testing data management and analysis platform of the present invention; Figure 2 This is a flowchart of the construction of environment-driven dynamic benchmark manifold and the decoupling and extraction of endogeneous performance deviations in this invention; Figure 3 This is a flowchart illustrating the multi-physics field coupling state comprehensive diagnostic system of the present invention; Figure 4This is a flowchart illustrating the whole lifecycle preventive maintenance decision-making system of this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between edge computing and cloud platform in this invention. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of specific embodiments based on the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0021] Reference Figures 1 to 5 As shown, the new energy power facility testing data management and analysis platform proposed in this invention consists of multiple decoupled and deeply collaborative functional units at the system logic level. The core design concept of this platform is to achieve high-precision preventive maintenance by actively modeling environmental factors and physically separating external interference from the performance degradation of the equipment itself.

[0022] The multi-source heterogeneous high-frequency synchronous sensing system serves as the data base for the entire platform, synchronously collecting environmental, electrical, and structural characteristic parameters of new energy power facilities under a unified clock reference.

[0023] The multi-source heterogeneous high-frequency synchronous sensing system is deployed within a cluster of new energy power facilities. It consists of a distributed sensor array, with each sensor node equipped with an independent BeiDou time synchronization module. This module receives the Coordinated Universal Time (UTC) signal from the BeiDou-3 satellite system and outputs a second pulse signal and standard time information. The microcontroller within the node uses this second pulse signal as the sampling trigger source and initiates the analog-to-digital converter for synchronous sampling via a hardware interrupt. At the same time, the standard time information is appended to the sampled data as a timestamp. All nodes use the same hardware triggering mechanism to ensure that the sampling time error is controlled within 20 nanoseconds, thereby achieving absolute alignment of multi-source data in the time dimension.

[0024] First, the environmental sensing subsystem is used to acquire environmental characteristic parameters and integrates a total radiation sensor, a direct radiometer, an ultrasonic three-dimensional anemometer, a high-precision platinum resistance thermometer, an environmental barometer, and a salt spray concentration sensor.

[0025] The total radiation sensor works in conjunction with the direct radiometer to monitor the incident irradiance on the surface of the photovoltaic module in real time. The analog signal output by the sensor is converted into a digital signal by a 16-bit analog-to-digital converter and then transmitted to the system data bus via an industrial fieldbus such as CAN bus or Modbus RTU protocol.

[0026] The ultrasonic three-dimensional anemometer uses the ultrasonic time-of-flight method to measure wind speed and direction. Its internal digital signal processor directly outputs the digital value of the wind speed vector, which is also uploaded through the bus interface.

[0027] The sampling trigger signals of all environmental sensors come from the second pulse of the Beidou timing module, ensuring that environmental data is synchronized with electrical and structural data.

[0028] Secondly, the power operation monitoring subsystem undertakes the task of high-frequency acquisition of electrical physical quantities. This subsystem is connected to the current and voltage sampling module through fiber optic transformers. The sampling frequency is strictly set to 12.8 kHz. This frequency is generated by high-precision temperature-compensated crystal oscillator after frequency multiplication and division by phase-locked loop. The timer outputs a trigger signal to drive the analog-to-digital converter to sample at fixed intervals.

[0029] The sampled data is first written to the dual-port random access memory inside the field-programmable gate array (FPGA). This memory is divided into two equal-length buffers and uses a ping-pong operation mechanism: while the first buffer is being written to by the analog-to-digital converter (ADC), the contents of the second buffer can be read by the digital signal processor (DSP) via a high-speed parallel bus and used for calculations such as fast Fourier transform (FFT) or wavelet analysis. The read and write pointers of the two buffers are automatically switched by the state machine inside the FPGA, thereby ensuring that the data acquisition process is continuous and uninterrupted, and that the original data is not lost due to calculation delays, thus ensuring the integrity and continuity of the data.

[0030] Next, the structural feature sensing subsystem focuses on monitoring the mechanical condition of the physical entities of the equipment. Vibration acceleration sensors, infrared thermal imaging sensors, and acoustic emission sensors are deployed at key stress points such as the fan main shaft bearing housing, transformer winding housing, and inverter power module heat sink.

[0031] The charge signal output by the vibration acceleration sensor is converted into a voltage signal by a charge amplifier, and then sampled by a high-precision analog-to-digital converter after passing through an anti-aliasing filter to obtain the frequency domain characteristics of mechanical vibration.

[0032] The weak signal output by the acoustic emission sensor is amplified by a preamplifier and filtered out for low-frequency mechanical noise by a bandpass filter. It is then acquired by a high-speed data acquisition card at a megahertz sampling rate. This is specifically designed to detect ultrasonic signals generated by metal fatigue, enabling early warning of the propagation of microcracks inside materials.

[0033] The sampling triggers of all sensors are also connected to the Beidou second pulse synchronization signal to ensure that the mechanical status data is strictly aligned with the environmental and electrical data.

[0034] The three subsystems mentioned above together constitute a multi-source heterogeneous high-frequency synchronous sensing system. Based on the hard synchronization pulse provided by the Beidou timing module, and using diverse sensors and corresponding signal conditioning circuits, it achieves comprehensive high-fidelity perception of the operating status of new energy power facilities. Furthermore, the high-dimensional real-time data stream generated by this system carries a unified timestamp and will be received by the subsequent environment-driven dynamic benchmark manifold construction system. This will be used to remove environmental interference and establish a dynamic baseline for equipment performance, thereby laying a solid data foundation for accurate diagnosis and preventive maintenance.

[0035] The environment-driven dynamic baseline manifold construction system receives real-time streaming data from the environmental sensing subsystem. This data includes multi-dimensional environmental characteristic parameters such as irradiance, wind speed, and temperature. The system normalizes these multi-dimensional environmental variables. Because the dimensions of each parameter are very different, the min-max normalization method is used to map all inputs to a numerical range of 0 to 1.

[0036] The specific approach is as follows: For each environmental variable, the global minimum and maximum values ​​are extracted from the historical database. These two values ​​are updated quarterly to reflect seasonal variations. The normalization formula is: ; Where x is the current sampled value, x min x is the minimum value of this variable in the historical database. max Let x' be the maximum value of the variable in the historical database, and x' be the normalized dimensionless value. This formula is a standard method in the field of data preprocessing, which eliminates the influence of dimensions through linear transformation, making all inputs of the same order of magnitude.

[0037] After normalization is completed, the adaptive weight adjustment function is called. This function dynamically allocates the weight coefficients of each environmental variable on the equipment output based on prior knowledge of different seasons and geographical locations.

[0038] The system has a built-in weight knowledge base that stores the weight vectors of each environmental variable by season and geographic location. The weight vectors are pre-defined and updated periodically in the following way: Historical data on the equipment's health status over the past two years were selected. For each combination of season and geographical location, a multiple linear regression analysis was performed, with normalized environmental variables as input and actual equipment output as output. The obtained regression coefficients were normalized and used as the weight vector for that combination.

[0039] When the system is running, it determines the season based on the current date, determines the geographical location type based on the power station coordinates, and reads the corresponding weight vector w from the knowledge base. i Then, the normalized environment variables are weighted: ; Where w i Let be the weight coefficient of the i-th environmental variable, satisfying Let x be the i-th environment variable after normalization. weighted,i These are the weighted feature values, which highlight the impact of key environmental factors on equipment performance through weighting.

[0040] The weighted environmental feature vector is used as input to enter the nonlinear manifold learning stage. The system uses an equidistant mapping algorithm to eliminate the strong collinearity between environmental variables. The algorithm first constructs a neighborhood graph in a high-dimensional space and connects sample points through the K-nearest neighbor method, where the K value is taken as an empirical value of 15 and the distance metric is Euclidean distance. Then, the shortest path distance between each sample point is calculated as an approximation of the geodesic distance. In this embodiment, Dijkstra's algorithm is used to solve for the shortest path.

[0041] Next, a multidimensional scaling method is applied to the geodesic distance matrix to maintain this distance relationship in the low-dimensional eigenspace, thereby obtaining the low-dimensional embedded coordinates of the environmental manifold. In this embodiment, the dimension after dimensionality reduction is set to 3 to balance computational complexity and information retention, effectively extract the intrinsic structure of the environmental data, and provide clean input features for subsequent modeling.

[0042] After constructing the environmental manifold, an ideal output hypersurface is established. This hypersurface is generated by training a high-performance computing model using massive amounts of historical operating data and represents the theoretical performance boundary of the equipment under fully healthy conditions as the environment changes.

[0043] The training process selects data from periods of good health in the equipment's historical operation. The criteria for judging the health status are that there are no fault records for 30 consecutive days during the initial operation of the equipment or after thorough maintenance, and all performance indicators are better than the factory standard.

[0044] Using the low-dimensional embedded coordinates of the environmental manifold as input and the actual output parameters of the equipment, such as theoretical active power, theoretical reactive power, theoretical bus voltage, and theoretical converter temperature rise, as output, a random forest regression model is used for supervised learning. The number of trees in the random forest is set to 100, the maximum depth is 10, and the other parameters use default values.

[0045] Once the model is trained, a multidimensional mapping surface from the environmental manifold to the output parameters is formed. This surface is actually a regression model stored in memory, which can predict the ideal output value based on the input environmental manifold coordinates.

[0046] When the sensing system receives a set of current environmental characteristic parameters, it first normalizes, weights, and projects them onto the environmental manifold space to obtain the coordinates of the current environment on the low-dimensional manifold. Then, it performs radial basis function interpolation to locate an instantaneous ideal operating state point in real time on the ideal output hypersurface. The radial basis function interpolation uses a Gaussian kernel function. ; Where r represents the Euclidean distance between the point to be interpolated and a certain historical manifold sample point, and σ is the kernel width parameter, which is determined to be 0.3 times the average distance between all sample points through cross-validation in this embodiment. The interpolation calculation is centered on all historical manifold sample points, and the interpolation weight of each sample point is obtained by solving a system of linear equations.

[0047] Specifically, suppose there are N historical sample points, and their corresponding environmental manifold coordinates are p. j The output value is y j (Can be a multidimensional vector), for the current environment manifold coordinates p, the interpolation formula is: ; Where λ j The weighting coefficients are to be determined. They are obtained by solving the linear equation system Φλ=y, where Φ is an N×N matrix with elements... λ is the weight column vector, and y is the output value matrix (N rows, each row corresponding to the output vector of a sample point). Since N may be very large, in actual calculations, a portion of sample points, such as 1000, are randomly selected for interpolation to ensure real-time performance.

[0048] Through the above interpolation, the theoretical output value corresponding to the current environment is calculated. This state point is not a single value, but a multi-dimensional vector that includes theoretical active power, theoretical reactive power, theoretical bus voltage, and theoretical converter temperature rise.

[0049] Through the above process, the environment-driven dynamic benchmark manifold construction system transforms the monitoring benchmark from a fixed threshold into a dynamic baseline that fluctuates with the environment, accurately reflecting the ideal performance that the equipment should have under the current environmental conditions. This provides a reliable reference benchmark for the subsequent endogeneous performance deviation decoupling and extraction system, enabling the system to effectively isolate environmental factors and focus on the accurate diagnosis of equipment performance degradation.

[0050] The endogeneity performance deviation decoupling and extraction system receives real-time operational data streams from the power operation monitoring subsystem over 20 consecutive sampling periods at a sampling frequency of 12.8 kHz. Therefore, 20 periods correspond to a data window of approximately 1.56 milliseconds. These data are organized into a dynamic real-time operational feature matrix R, where rows correspond to the 20 sampling times and columns correspond to various electrical operating parameters. Assuming there are m electrical operating parameters, including active current, reactive current, and phase voltage, the dimension of matrix R is 20×m, thus comprehensively describing the dynamic behavior of the equipment within a short time window.

[0051] Meanwhile, the system receives the instantaneous ideal operating state point sequence output by the environment-driven dynamic reference manifold construction system. This sequence also contains ideal values ​​at 20 time points, with each time point corresponding to a multi-dimensional vector, forming the ideal operating state matrix I, whose dimension is the same as R.

[0052] Due to millisecond-level delays caused by network congestion or protocol conversion in the data transmission link, there is a time phase deviation between the real-time running feature matrix R and the ideal running state matrix I. To solve this problem, a dynamic time warping algorithm is used to phase align the row sequences of the two matrices.

[0053] Dynamic time warping is a classic time series similarity measurement method. Its core idea is to find the optimal matching path between two series by non-linearly stretching or compressing the time axis. Let r be the i-th row vector of the real-time running feature matrix R. i The j-th row vector of the ideal operating state matrix I is i j .

[0054] First, construct a 20×20 cumulative distance matrix D, where each element D(i,j) represents the minimum cumulative distance from the starting point (1,1) to the current point (i,j). The local distance d(i,j) is defined as the vector r i with i j The Euclidean distance between them is calculated using the following formula: ; The cumulative distance matrix D is calculated using the following recursive formula: D(1,1)=d(1,1); For i = 1 to 20, j = 1 to 20, and (i,j) ≠ (1,1): ; When i-1 or j-1 is less than 1, the corresponding D value is infinity to ensure that the path starts from (1,1).

[0055] After calculating the entire D matrix, the system backtracks from D(20,20) to find the optimal curved path. The backtracking rule is: starting from the current point (i,j), select the predecessor point that minimizes the cumulative distance, that is, compare D(i-1,j), D(i,j-1), and D(i-1,j-1), select the point corresponding to the minimum value as the previous step, and repeat this process until returning to (1,1).

[0056] This path consists of a series of coordinate pairs (p1, q1), (p2, q2), ..., (p...). L ,q L The expression consists of ) where L is the path length, and p1=1, q1=1, p L =20, q L =20.

[0057] Based on this optimal path, the row sequence of the ideal operating state matrix I is rearranged to obtain an aligned ideal matrix I' that is synchronized with the real-time matrix R on the time axis. Specifically, for each row index i of the real-time operating feature matrix, all coordinate pairs p=i are found in the path. These coordinate pairs may correspond to more than one q value. If a certain i corresponds to multiple q values, the average of the ideal row vectors corresponding to these q values ​​is taken as the i-th row of I'. If a certain i does not appear in the path, it means that there is no corresponding ideal point for that real-time point. In this case, a linear interpolation method is used to generate the ideal value for that row based on adjacent aligned points. In this way, an aligned ideal matrix I' with the same number of rows as R is obtained.

[0058] Based on phase alignment, the residual matrix Δ between the real-time running feature matrix and the aligned ideal matrix is ​​calculated, and its elements are the differences between the corresponding row vectors: Δ=R-I'; Where Δ is a 20×m matrix, and each row has δ t =r t -i' t This represents the deviation vector of each electrical parameter at time t. The residual matrix contains all performance deviations caused by non-environmental factors, such as sensor measurement noise, external electromagnetic interference, and internal physical attenuation of the equipment.

[0059] To extract features with clear physical meaning from the residual matrix Δ, the Independent Component Analysis (ICA) algorithm is used. ICA is a statistical method based on blind source separation theory. It assumes that the observed mixed signal is a linear combination of multiple statistically independent source signals. In this scenario, Δ is regarded as the sampled values ​​of m observation channels at 20 time points. The observation value of each channel is a linear mixture of 20 source signals. By solving an unmixing matrix, the mixed signal can be decomposed into mutually independent feature components.

[0060] The specific implementation employs a fast independent component analysis algorithm, which maximizes the non-Gaussianity of the output signal through fixed-point iteration. The system first preprocesses Δ: subtracting the mean of each channel's data to make its mean zero yields a centered matrix. Then, whitening is performed, and principal component analysis transforms the data into a matrix where each dimension is uncorrelated and has a variance of 1. The whitened data is denoted as Z, a 20×m matrix.

[0061] Subsequently, the system initializes an m-dimensional unmixing vector w (column vector) and updates it according to the following iterative formula: ; in, It is a nonlinear function. Its derivative, This represents the expectation, which is actually replaced by the sample mean in the actual calculation. That is, the average is taken over 20 time points, and the iteration continues until w converges, that is, the Euclidean distance between two consecutive times is less than a preset threshold, such as 10. -6 For each converged unmixed vector w, an independent component s=w can be extracted. T Z and s are 1×20 row vectors.

[0062] The system is set to separate 3 components, so three independent components need to be extracted sequentially. After extracting the first component, the influence of the extracted component is removed by orthogonalization, and the above process is repeated to extract the next component.

[0063] Orthogonalization is performed using the Gram-Schmidt method: Let the extracted unmixed vectors form a matrix W. Then, the new unmixed vector w needs to be projected onto a subspace orthogonal to W, i.e., w = w - WW. T w is then normalized, resulting in three independent components s1, s2, and s3, each a time series of length 20.

[0064] To assign physical meaning to these components, the time-domain waveform and spectrum of each component are analyzed. The specific method is as follows: For each independent component s, its power spectral density is first calculated. Using the Welch method, the 20-point sequence is segmented and averaged to obtain the frequency-power curve. Then, the spectrum of the component is compared with the typical fault characteristic frequencies of various components of the equipment. For example, current attenuation caused by dust accumulation in photovoltaic modules mainly exhibits a low-frequency trend, with its power spectral energy concentrated below 0.1Hz; wear of the wind turbine main bearing will cause energy attenuation near its natural frequency, which can be obtained through equipment parameters or historical data; aging of inverter capacitors will lead to an increase in output voltage ripple, with the ripple frequency typically near the switching frequency, such as several kilohertz.

[0065] By analyzing the spectral energy distribution of each component, each component is classified into the corresponding physical phenomenon. In addition, combined with the time domain trend, if the component shows a slow decreasing trend and is unrelated to irradiance, it is determined to be dust accumulation attenuation; if the component's energy continues to decrease in a specific frequency band, it is determined to be bearing wear; if the component's energy increases in the high-frequency band, it is determined to be capacitor aging.

[0066] Based on the above analysis, the three independent components are labeled as dust accumulation attenuation feature f1, bearing wear feature f2, and capacitor aging feature f3, respectively, and combined and encapsulated into a multidimensional endogenous attenuation feature vector. , where each feature f k The value is taken as the average energy value of the component or the energy value of a specific frequency band. This feature vector serves as the core input of the subsequent multi-physics coupling state integrated diagnostic system.

[0067] This decoupling extraction mechanism effectively eliminates false alarms caused by sudden environmental changes such as cloud cover and gusts of wind from a physical perspective. This is because environmental factors have been stripped away by the environment-driven dynamic benchmark manifold construction system, and the remaining residuals only contain non-environmental factors related to the equipment itself and the measurement system, which greatly improves the system's signal-to-noise ratio and diagnostic accuracy.

[0068] Through the above process, the endogeneity performance deviation decoupling extraction system transforms the original monitoring data into feature vectors with clear physical meaning, providing accurate input for the multi-physics coupled state comprehensive diagnostic system, thereby enabling early identification and quantitative assessment of equipment performance degradation.

[0069] The multi-physics coupled state comprehensive diagnostic system integrates multiple specialized analysis modules. These modules utilize environmental, electrical, and structural data collected by the aforementioned multi-source heterogeneous high-frequency synchronous sensing system to conduct in-depth analysis of different physical fields and failure modes. Specifically, it includes the following analysis modules: a dust accumulation evolution analysis module assesses the degree of dust accumulation on photovoltaic modules based on irradiance data collected by the environmental sensing subsystem and the current gain attenuation component in the endogenous attenuation characteristic vector; an electrical physical performance evaluation module derives the power semiconductor junction temperature using current and voltage data collected by the power operation monitoring subsystem and the casing temperature measured by an infrared thermal imaging sensor; and a structural dynamics diagnostic module assesses mechanical structure fatigue damage based on vibration and acoustic emission signals collected by the structural feature sensing subsystem. The analysis results from each module are ultimately integrated to form a comprehensive evaluation of the overall health status of the equipment.

[0070] The dust accumulation evolution analysis module is specifically designed for photovoltaic module applications. This module first establishes a regression model between the photocurrent decay rate and the cleaning cycle using historical operating data. Specifically, it collects photocurrent data before and after each cleaning cycle, calculates the average daily decay rate within each cleaning cycle, uses the number of days in the cleaning cycle as the independent variable and the daily decay rate as the dependent variable, and employs a univariate linear regression to obtain the following relationship: d = a·T + b; Where d is the average daily photocurrent decrease rate, expressed as a percentage per day, T is the number of days in the cleaning cycle, and a and b are regression coefficients. The coefficients are solved using the least squares method.

[0071] When the system is running in real time, the dust accumulation decay feature f1 in the endogenous decay feature vector reflects the current relative current loss due to dust accumulation. Dividing f1 by the daily decay rate d under the corresponding cleaning cycle obtained from the regression model, the equivalent number of days of dust accumulation since the last cleaning can be estimated, and then the current dust density ρ can be calculated according to the empirical formula: ρ = k·f1; Where k is the conversion coefficient, which is obtained through on-site calibration. For example, it is taken as 0.5 g / m² per percentage decrease. This coefficient is determined by the statistical relationship between the measured dust density and the corresponding f1 in historical data.

[0072] After obtaining the dust density, the dust evolution analysis module predicts the power generation loss over a future period. Based on the power decay versus dust density relationship curve provided by the photovoltaic module manufacturer, a function L(ρ) is fitted, representing the percentage of power loss per unit area due to dust accumulation under standard irradiance. Combining the predicted future daily average irradiance and the rated power of the photovoltaic modules, the total power generation loss E caused by dust accumulation over the next N days is calculated. loss : ; Where P rated H is the rated power of the power station. t Let denoted as the predicted peak sunshine hours on day t, and Δρ as the daily growth rate of dust accumulation density, which can be obtained from historical data.

[0073] Cleaning cost C cleam Including labor costs and equipment rental fees, data is obtained from the operation and maintenance knowledge base. When the expected increase in power generation revenue within the next 30 days exceeds the cleaning cost, a cleaning recommendation is automatically generated. ; The electrical physics performance evaluation module focuses on the reliability evaluation of power semiconductor devices. This module uses the classic thermal resistance model to derive the internal junction temperature: ; Where T jT represents the equivalent junction temperature of a power semiconductor, in degrees Celsius. c P represents the surface temperature of the device casing as measured by an infrared thermal imaging sensor, in degrees Celsius. loss Power loss calculated based on real-time current and voltage samples, in watts; R th(j-c) This is a characteristic parameter of thermal resistance from the semiconductor junction to the casing, expressed in degrees Celsius per watt. The formula is derived from Ohm's law of thermal path in heat transfer.

[0074] Power loss P loss The calculation method is as follows: For insulated-gate bipolar transistors in inverters, a simplified formula is used: ; Where I is the collector current, in amperes; V ce(sat) The saturation pressure drop is expressed in volts (E). sw The energy for switching is expressed in joules (f). sw The switching frequency is expressed in Hertz. These parameters can be obtained from the device datasheet. Thermal resistance R... th(j-c) Similarly, the information can be obtained from the device datasheet or calibrated experimentally.

[0075] When the calculated equivalent junction temperature exceeds 125 degrees Celsius, it is determined that the power device has a serious aging risk or a heat dissipation system failure, and a high-level aging warning is immediately triggered.

[0076] The structural dynamics diagnostic module uses acoustic emission sensors and vibration acceleration sensors to assess the mechanical structure status. For acoustic emission signals, empirical mode decomposition is first performed to decompose the signal into several intrinsic mode functions. Hilbert transform is performed on each intrinsic mode function to obtain the instantaneous frequency and instantaneous amplitude. The average slope value of the instantaneous frequency (FAE) is extracted as a characteristic index of microcrack propagation.

[0077] For the vibration signal acquired by the vibration accelerometer, spectral analysis is performed to calculate the power spectral density S(f). Then, the centroid frequency of the vibration spectrum is calculated. ; And calculate its offset Δf relative to the normal state. threshold This offset reflects the stiffness changes and fatigue accumulation of rotating components or supporting structures.

[0078] The fatigue damage index D is derived from the average slope of the instantaneous frequency and the frequency offset of the center of gravity. ; Where α and β are weighting coefficients, α + β = 1, and in this embodiment, α = 0.4 and β = 0.6; F AE,threshold and Δf thresholdThe threshold is an empirical threshold obtained through statistical analysis of historical fault data. When D>1, the fatigue level is considered to be excessive.

[0079] The multi-physics coupled state comprehensive diagnostic system ultimately adopts the fuzzy comprehensive evaluation method, which converts dust density ρ and aging risk index (based on whether the junction temperature exceeds the limit) into a risk value R. thermal The aging risk index R is nonlinearly fused with the fatigue damage index D. First, the membership functions of each input variable are defined. Taking dust density as an example, a trapezoidal membership function is used to divide the data into three fuzzy sets: low, medium, and high. Parameters ρ1, ρ2, and ρ3 are set based on field experience. thermal Similar membership functions are defined for the fatigue damage index D.

[0080] Establish a fuzzy inference rule base, using if-then rules, for example: if ρ is high and R thermal If ρ is high and D is high, then the overall health index is low; if ρ is medium and R is high... thermal If ρ is low and R is high, then the overall health index is medium. thermal If both D and 'low' are true, then the overall health index is high. A total of 27 rules cover all combinations, and the output health index is also divided into three fuzzy sets: low, medium, and high.

[0081] Using the Mamdani inference method, for each input combination, the activation degree of each rule is calculated (taking the minimum value of the membership degrees of each premise), resulting in the truncated membership function of the output fuzzy set. Finally, the centroid method is used to defuzzify and obtain the accurate comprehensive health index H. ; Where μ out (h) is the aggregated output membership function. H takes values ​​from 0 to 1. 1 represents the initial optimal state of the equipment when it leaves the factory, and 0 represents that the performance has degraded to the failure threshold and must be stopped and repaired immediately.

[0082] Through the above multi-physics field coupling diagnostic process, the system can comprehensively assess the health status of equipment from multiple dimensions such as dust accumulation, electrothermal stress, and mechanical fatigue, providing accurate and comprehensive health quantification indicators for the subsequent whole life cycle preventive maintenance decision system.

[0083] The whole life cycle preventive maintenance decision system receives the comprehensive health index output by the multi-physics field coupled state comprehensive diagnosis system, and formulates dynamic optimization strategies and hierarchical early warning response mechanisms based on the index.

[0084] First, a Wiener process model is established to describe the evolution of health status over time. The Wiener process is a Brownian motion with a drift term, suitable for simulating stochastic degradation processes. The comprehensive health index is considered as a stochastic process, and its continuous-time expression is as follows: ;

[0085] Where H(t) is the health index at time t, μ is the drift coefficient reflecting the average degradation rate, σ is the diffusion coefficient reflecting the random fluctuation amplitude, and dW(t) is the standard Brownian motion increment. This formula is the definition of the Wiener process.

[0086] In actual discrete observations, if the sampling interval is Δt, then the discretized model is: ; in It is a standard normal random variable.

[0087] The system uses a Bayesian method to update model parameters μ and σ² online. It assumes that the prior distribution of the parameters is a normal-inverse gamma distribution, that is, given μ, σ² follows a normal distribution and σ² follows an inverse gamma distribution. Whenever a new health index observation is obtained, the increment is calculated based on the discretized model. The parameters follow a normal distribution with mean μΔt and variance σ²Δt. By combining the prior data to obtain the posterior distribution of the parameters, the estimated values ​​of μ and σ² are updated. This update process allows the model to adapt to changes in the rate of equipment degradation, improving prediction accuracy.

[0088] Based on the updated model, remaining lifespan can be predicted. Remaining lifespan is defined as the health index first decreasing from its current value H0 to the failure threshold H. th The time. When μ < 0, the expected remaining lifetime is: ;

[0089] The failure threshold is set according to the importance of the equipment; for example, 0.4 is set for critical equipment.

[0090] The system is configured with a three-level early warning response mechanism, and the thresholds and handling strategies for each level are as follows: Level 1 warning: When the overall health index drops to 0.8, the device enters a state of mild concern, is added to the key observation list, and the sampling frequency of all associated sensors is increased to twice the original frequency, for example, from 12.8 kHz to 25.6 kHz, in order to obtain more refined high-frequency data.

[0091] Level 2 Warning: When the health index drops to 0.6, the equipment enters a sub-healthy state, triggering a maintenance cost assessment. This assessment quantifies the net benefits of different maintenance timings. The net maintenance benefit equals the avoided power generation loss minus the maintenance cost. The avoided power generation loss is calculated based on future weather forecasts, electricity price curves, and equipment efficiency reduction. Let the current day be day 0. Considering maintenance dates d within the next M days, the loss avoided by maintenance on day d is: ; Where P rated For rated power, Ht Let η(t) be the predicted equivalent full-load hours for day t, η(t) be the efficiency reduction factor corresponding to the health index (which can be taken as η=H(t)), price(t) be the time-of-use electricity price, and C be the maintenance cost. repair Obtained from the operations and maintenance knowledge base, the system calculates the net benefit for all d values. Select the date with the largest net profit as the optimal maintenance window. If all net profits are not positive, it is recommended not to carry out maintenance for the time being.

[0092] Level 3 Warning: When the health index drops to 0.4, the equipment faces an extremely high risk of failure and requires immediate intervention. Maintenance instructions are automatically generated, including details of the parts to be inspected, such as the wind turbine main shaft bearing or the inverter power module, a spare parts list with model, quantity, and supplier, and standard operating procedures. Simultaneously, a geographic information system is used to plan the shortest path, and Dijkstra's algorithm is used to calculate the optimal route from the current point to the target equipment, ensuring that maintenance personnel can arrive quickly.

[0093] Through the Wiener process modeling and Bayesian parameter updates described above, as well as three-level early warning and maintenance decision support, the system transforms health indices into specific operation and maintenance action instructions, thereby achieving closed-loop preventive maintenance.

[0094] To support the efficient and stable operation of the aforementioned complex algorithms, the platform has built a dedicated data management center based on a distributed architecture. This center adopts a layered and decoupled design in the storage layer to meet the differentiated needs of different data types for read / write performance and storage costs.

[0095] For millisecond-level high-frequency sampling data, such as 12.8 kHz current and voltage waveforms and real-time values ​​from various sensors, a distributed time-series database is used for storage. In this embodiment, the open-source time-series database InfluxDB is selected and a cluster deployment mode is adopted. The data is stored in time-sharded chunks, and the retention policy is set to retain the most recent 30 days for hot data. The time-series database has been specially compressed and optimized for time series data, which can handle millions of data write requests per second and supports efficient aggregation queries based on time windows, which can quickly extract statistical features within any time period.

[0096] Static equipment ledger information, such as equipment model, installation location, commissioning date, historical maintenance records, and metadata generated by the algorithm model, is stored in a traditional relational database. In this embodiment, PostgreSQL is selected, and a master-slave replication architecture is adopted to ensure data consistency and high availability. Relational databases support complex join queries, making it easy for maintenance personnel to retrieve equipment files. This hierarchical storage strategy not only ensures the stringent data throughput requirements of real-time diagnostics, but also takes into account the economy and flexibility of long-term data management.

[0097] The data management center also provides a self-describing data access interface. Any upper-layer algorithm module can obtain the required data slices through a publish-subscribe model. The self-describing interface adopts a standardized data format and semantic tags, specifically using Apache Avro as the serialization format, and defines a clear schema for each data topic, including field names, data types, and physical meaning descriptions. Newly connected algorithm modules can automatically discover and understand available data topics by reading the definitions in the schema registry, without the need for pre-configuration of complex mapping relationships.

[0098] The publish-subscribe model is based on message queues. In this embodiment, Apache Kafka is selected as the message middleware. The data producer publishes the processed data stream to specific topics, such as the raw sampling data topic, the health index topic, and the early warning event topic. The subscriber dynamically subscribes to relevant topics according to its own needs. Load balancing and breakpoint resume are achieved through Kafka consumer groups, realizing loose coupling between modules and efficient data flow.

[0099] In terms of security, the data management center is equipped with hardware encryption chips that conform to the national standard algorithm SM2 on all key transmission nodes. The chip is connected to the main control CPU through the SPI interface and has a built-in SM2 public key cryptographic algorithm engine. Whether it is wireless transmission from edge sensors to the site gateway or public network transmission from the site to the cloud central platform, all data packets are signed and encrypted by the chip before being sent.

[0100] The specific process is as follows: the sender uses the receiver's public key to encrypt the data packet and generates a digital signature using its own private key. After the receiver verifies the signature, it decrypts the data using its own private key. Hardware encryption ensures that critical infrastructure information such as power operation data is not illegally intercepted, tampered with, or forged during transmission. This end-to-end security mechanism builds a trusted data transmission environment for the entire platform, meeting the highest requirements of the power industry for data security and privacy protection.

[0101] Through the aforementioned layered storage design, self-describing interface and publish-subscribe mechanism, as well as hardware-level encrypted transmission, the data management center provides reliable, efficient, and secure data services to the upper-layer diagnostic and decision-making modules. This ensures that the entire platform can still operate stably in new energy operation and maintenance scenarios with high concurrency and extremely high real-time requirements, providing a solid data foundation for subsequent complex algorithms such as environmental stripping, deviation decoupling, health diagnosis, and maintenance decision-making.

[0102] As a further refinement of this embodiment, the environment-driven dynamic benchmark manifold construction system integrates a self-correction mechanism to cope with the operating condition drift that may occur during long-term operation. The system continuously monitors the endogeneous performance deviation and decouples and extracts the residual sequence of the system output. This residual sequence is the sampled value of a 20-dimensional vector in a continuous time. When it is detected that the mean of the residual over 30 consecutive calendar days exhibits a systematic deviation that cannot be explained by the existing physical field coupling model, a self-supervised learning program is automatically triggered.

[0103] Systematic bias is determined using a dual-threshold strategy. First, a one-sample t-test is performed, comparing the residuals at each time step to zero. The test statistic is... ,in Let s be the mean of the residuals over 30 days, s be the standard deviation, n=30, and the significance level be set at α=0.05. If the deviation is significant, it is considered to be significantly off from zero. To prevent frequent corrections caused by small but statistically significant deviations, an engineering threshold is introduced: the absolute value of the residual mean is required to exceed 0.05 for 30 consecutive days.

[0104] The self-supervised learning process is triggered only when both conditions are met simultaneously.

[0105] The self-supervised learning program first re-examines the relationship between environmental inputs and equipment outputs in the most recent quarter using cluster analysis techniques to identify whether there are new operating conditions caused by changes in the surrounding environment, such as new building obstruction, vegetation growth, and surface changes. In this embodiment, the DBSCAN density clustering algorithm is used, with the low-dimensional embedded coordinates of the environmental manifold as input, the neighborhood radius Eps is set to 0.2, and the minimum number of samples MinPts is set to 10. In the clustering results, if a new cluster is formed and the number of samples in the cluster exceeds 5% of the total number of samples, it is determined to be a new operating condition.

[0106] Once a new operating condition is identified, incremental learning techniques are used to fine-tune the model for the corresponding local region while retaining the original knowledge of the ideal output hypersurface. Specifically, the sample points of the new operating condition are added to the historical sample set, but the radial basis function interpolation model is retrained only for sample points that are close to the new operating condition. The distance threshold is set to twice the kernel width σ, and the interpolation weights of these local regions are resolved. The weights of sample points far from the new operating condition remain unchanged. Through this local update, while ensuring real-time performance, the dynamic benchmark adapts to the new operating environment, thereby ensuring that the dynamic benchmark maintains extremely high accuracy and adaptability throughout its operating cycle of several years.

[0107] The multi-physics coupled state integrated diagnostic system adopts an edge computing and cloud collaboration deployment mode when processing large-scale site group data. The edge computing gateway deployed at each site uses an industrial-grade embedded computer equipped with a quad-core ARM processor and 8GB of memory, running a lightweight Linux system.

[0108] An endogeneous performance deviation decoupling extraction algorithm module is deployed inside the gateway. This module calculates the deviation residual in real time and performs independent component analysis on the high-frequency sampled data according to the endogeneous performance deviation decoupling extraction method described above, generating a lightweight endogeneous attenuation feature vector. The feature vector contains only three feature components and their corresponding timestamps, and the data volume is much smaller than the original waveform, thereby achieving efficient compression.

[0109] The gateway uploads the feature vector to the cloud center through an encrypted tunnel established based on the Transport Layer Security (TLS) protocol. If the site communication is interrupted, the gateway will temporarily store the feature vector on the local solid-state drive and record the breakpoint timestamp. The transmission will be automatically resumed after communication is restored to ensure that the data is not lost.

[0110] The cloud platform is deployed on a high-performance computing cluster, running cross-site deep evolutionary analysis algorithms, global lifetime prediction models, and multi-dimensional decision-making logic, and performing global optimization by integrating data from all sites.

[0111] The full lifecycle preventive maintenance decision-making system deeply integrates with the spare parts inventory management system when generating maintenance recommendations. When the decision-making system determines that a component needs to be replaced, it automatically calls the RESTful API interface provided by the inventory system to query the inventory quantity, storage location, and batch information of the corresponding model component in JSON format. The inventory system is built on an enterprise resource planning system, and the interface uses HTTPS protocol and API key for security authentication. If the inventory is found to be lower than the preset safety threshold, a purchase requisition is automatically generated and pushed to the supply chain management module via email or WeChat, ensuring that maintenance work is not delayed due to material shortages.

[0112] Meanwhile, the system establishes digital twin simulation models under different operation and maintenance strategies. These models are built based on the degradation mechanism and operating data of the equipment. The degradation mechanism adopts the Wiener process model mentioned above, and the drift coefficient and diffusion coefficient are estimated using historical health index sequences.

[0113] The simulation model is based on the Wiener process and simulates the evolution path of the health index under different maintenance strategies starting from the current health state in a virtual space. The maintenance strategies include immediate maintenance, maintenance delayed for 30 days, and no maintenance. Each simulation generates a health index trajectory and calculates the power generation loss based on the relationship between the health index and efficiency reduction. At the same time, the maintenance cost is obtained from the inventory system.

[0114] The system simulates ten thousand times using the Monte Carlo method to obtain the expected net present value of each strategy. The system compares the expected net present values ​​of each strategy, selects the optimal strategy, and adjusts the current decision parameters, such as the three-level early warning threshold and the economic evaluation coefficient, in reverse according to the threshold parameters corresponding to the optimal strategy, so that the platform management level continues to evolve over time.

[0115] New energy power facilities are often located in remote areas with harsh environments. To address this, the multi-source heterogeneous high-frequency synchronous sensing system has been expanded to include real-time monitoring of ambient air pressure, ultraviolet intensity, and salt spray concentration.

[0116] During the decoupling and extraction of endogeneous performance deviations, a metal corrosion rate compensation factor based on a specific geographical location is introduced based on the salt spray concentration data collected by the sensor. This factor is pre-calibrated by referring to the corrosion rate level of the corresponding region in the national standard GB / T 15957 or based on the field plate test data.

[0117] For island wind farms deployed along the southeast coast, the corrosion impact weight in the structural health assessment model is automatically increased. Specifically, salt spray concentration is used as an important correction term for fatigue damage indicators, and the correction formula is as follows: ; Where D is the original fatigue damage index, and C salt The normalized salt spray concentration and γ, the corrosion influence coefficient, were determined experimentally. For photovoltaic modules deployed in the northwestern desert, the focus shifted to the analysis of polymer material aging caused by ultraviolet radiation. The aging rate of the insulation material was corrected using ultraviolet intensity data, and the correction method was similar to that for salt spray.

[0118] Through the aforementioned self-correction mechanism, edge cloud collaborative architecture, inventory linkage and digital twin optimization, and environmental adaptability expansion, the platform of this invention has demonstrated outstanding robustness, scalability, and intelligence in long-term operation, providing solid technical support for subsequent emergency strategies and full lifecycle management.

[0119] To enhance the intuitive experience and interactive feel of operations and maintenance personnel, the platform integrates a digital twin view system based on a 3D graphics engine. Operations and maintenance personnel can access the system through a web browser or mobile browser to view 3D modeled images of the entire site and individual devices in real time. The 3D models are loaded in glTF format and rendered using the Three.js engine.

[0120] The surface color of each component of the device changes dynamically according to the real-time value of the comprehensive health index. A health index greater than or equal to 0.8 is displayed in green, between 0.6 and 0.8 is displayed in yellow, and below 0.6 is displayed in red. Color updates are implemented by pushing health index data to the front end in real time via WebSocket.

[0121] When any dimension of the endogeneous decay characteristic vector of a component exceeds the preset safety boundary, the system highlights the potential fault location in the view and pops up a floating window to display the relevant physical quantity fluctuation curves in real time, such as vibration amplitude and temperature change trends. Maintenance personnel can retrieve the historical maintenance frequency and replacement records of the part with one click. All data is obtained through the RESTful API provided by the backend data management center, thereby shortening the on-site troubleshooting and decision-making time.

[0122] The platform's underlying software architecture adopts a strict layered and decoupled design to ensure system maintainability and scalability. The hardware abstraction layer shields the underlying protocol differences of sensors from different manufacturers and models through a unified device driver interface, converts the acquired raw signals into a standardized data packet format and uploads them to the data bus. In the specific implementation, protocol adapters such as Modbus and CANopen are used to map the data into JSON format, which is then published by Kafka producers.

[0123] The middle algorithm logic layer is the core of the system. It is deployed using containerization technology. Each algorithm module, such as environment-driven dynamic baseline manifold construction and endogeneity deviation decoupling, runs in an independent Docker container and is managed by Kubernetes orchestration. It can automatically scale horizontally according to the computing load.

[0124] The top-level application presentation layer is designed for managers at different levels, providing multi-dimensional data dashboards ranging from details of individual devices to macro-level indicators of the site cluster, including real-time monitoring, historical trends, and early warning events. The front end is developed using the Vue.js framework.

[0125] Data exchange between different levels is achieved through the high-performance asynchronous message queue Kafka, ensuring system stability and response speed under high concurrency, while also achieving loose coupling between modules.

[0126] In terms of data storage strategy, the platform implements a fine-grained data hot and cold classification mechanism to balance performance and cost. High-frequency raw operating data within the last 30 days is defined as hot data, which is frequently accessed and is crucial for real-time diagnostics. It is stored in a solid-state disk array and stored using the InfluxDB time-series database. The data retention policy is set to automatically delete the data after 30 days.

[0127] Historical data older than 30 days but less than 1 year is defined as warm data. It is compressed using the LZ4 lossless compression algorithm and stored on a large-capacity hard disk array. An index is created to record the file path and time range. Historical archive data older than 1 year is defined as cold data. The system automatically extracts its statistical feature summaries, such as mean, variance, and extreme values. The complete original message is then transferred to a low-cost cloud cold storage system, such as AWS S3 Glacier or Alibaba Cloud OSS archive storage. At the same time, metadata and summary information are retained in the local database for fast retrieval.

[0128] This multi-level storage strategy significantly optimizes storage costs while ensuring long-term data backtracking capabilities.

[0129] The platform also includes a collaborative optimization control interface that links with the power plant control center, enabling bidirectional data exchange based on the standard industrial communication protocol IEC61850 or Modbus TCP.

[0130] When the preventive maintenance decision system identifies an inverter or converter unit as being in an extremely high-risk state, i.e., the comprehensive health index drops below 0.4 and the predicted remaining lifespan is less than 72 hours, and maintenance personnel are unable to reach the site for repairs in time due to severe weather conditions such as typhoons or blizzards, the system automatically activates the emergency strategy.

[0131] The emergency strategy sends a derated operation command to the power plant control center through the collaborative optimization control interface. The command follows the MMS service of IEC61850 or the write register request of Modbus TCP, and specifies the target equipment identification, the derated range (e.g., limiting the output power to 60% of the rated value), and the duration.

[0132] The system monitors and controls the response to confirm the execution of commands. By actively limiting the output power of potentially hazardous equipment, it can significantly reduce the junction temperature fluctuation range of its internal power semiconductors, slow down the accumulation of mechanical fatigue, effectively delay the performance degradation process, keep the equipment in a sub-healthy state, avoid catastrophic hardware damage or safety accidents, and buy time for subsequent manual maintenance.

[0133] Through the aforementioned digital twin visualization, layered decoupling architecture, hot and cold data classification, and collaborative control emergency response, the platform of this invention constructs a complete closed loop in improving operation and maintenance efficiency, ensuring stable system operation, and responding to extreme scenarios. This further consolidates the entire technology chain from data collection to intelligent decision-making, providing a solid guarantee for achieving lean management of the entire life cycle of new energy power facilities.

[0134] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0135] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A data management and analysis platform for new energy power facility testing, characterized in that, include: The multi-source heterogeneous high-frequency synchronous sensing system is used to synchronously collect environmental, electrical, and structural characteristic parameters of new energy power facilities under a unified clock reference. Each sensor node in the multi-source heterogeneous high-frequency synchronous sensing system is equipped with a Beidou time synchronization module. The Beidou time synchronization module is used to output a second pulse signal as a sampling trigger source to control the sampling time error within 20 nanoseconds and to append standard time information as a timestamp to the sampling data. An environment-driven dynamic benchmark manifold construction system is used to receive environmental feature parameters, extract low-dimensional embedded coordinates of environmental features through a nonlinear manifold learning algorithm, and construct an ideal output hypersurface with low-dimensional embedded coordinates as input and ideal equipment output as output based on historical health operation data. The endogeneous performance deviation decoupling extraction system is used to receive the instantaneous ideal operating state point of electrical characteristic parameters and ideal output hypersurface output, calculate the residual matrix between real-time operating characteristics and ideal operating state, and use independent component analysis algorithm to decouple and extract at least one endogeneous attenuation feature vector with clear physical meaning from the residual matrix. The endogeneous attenuation feature vector is used to characterize the performance degradation of the equipment body due to dust accumulation, wear or aging.

2. The new energy power facility testing data management and analysis platform according to claim 1, characterized in that, The multi-source heterogeneous high-frequency synchronous sensing system includes: The environmental sensing subsystem integrates a total radiation sensor, an ultrasonic three-dimensional anemometer, and a high-precision platinum resistance thermometer to acquire irradiance, wind speed, and temperature as environmental characteristic parameters. The power operation monitoring subsystem is used to collect current and voltage as electrical characteristic parameters at a sampling frequency of 12.8 kHz. The power operation monitoring subsystem includes a field-programmable gate array (FPGA), which is internally configured with a dual-port random access memory (DRAM). The DRAM is divided into two equal-length buffers and uses a ping-pong operation mechanism to ensure that while the analog-to-digital converter writes data to one buffer, the digital signal processor reads data from the other buffer for calculation, thus ensuring that the data acquisition process is continuous and uninterrupted. The structural feature sensing subsystem, deployed at key stress locations, includes vibration acceleration sensors, infrared thermal imaging sensors, and acoustic emission sensors, used to acquire vibration signals, temperature data, and acoustic emission signals as structural feature parameters.

3. The new energy power facility testing data management and analysis platform according to claim 1, characterized in that, The environment-driven dynamic baseline manifold construction system further includes: The normalization and weighting module is used to perform min-max normalization on the received multidimensional environmental feature parameters, and dynamically weight the normalized environmental variables based on a pre-defined weight vector distinguished by season and geographical location type. The manifold learning module is used to reduce the weighted high-dimensional environmental features to three dimensions by constructing a neighborhood graph and calculating geodesic distances using the isometric mapping algorithm, thus obtaining low-dimensional embedded coordinates. The hypersurface generation module takes historical low-dimensional embedded coordinates during periods when the device is in a healthy state as input and the corresponding actual output of the device as output, and trains it using a random forest regression model to generate an ideal output hypersurface; and... The interpolation module is used to locate the corresponding instantaneous ideal operating state point on the ideal output hypersurface in real time for the current input low-dimensional embedded coordinates using a radial basis function interpolation algorithm.

4. The new energy power facility testing data management and analysis platform according to claim 1, characterized in that, The endogeneity performance bias decoupling extraction system further includes: The phase alignment module is used to perform nonlinear time axis alignment on the real-time running feature matrix and the ideal running state matrix composed of instantaneous ideal running state points using a dynamic time warping algorithm, so as to eliminate millisecond-level time phase deviations in the data transmission link. The residual calculation module is used to calculate the difference between the real-time running feature matrix after phase alignment and the ideal running state matrix, generating a residual matrix; and, The feature separation module is used to perform blind source separation on the residual matrix using a fast independent component analysis algorithm to obtain multiple statistically independent source signals. By analyzing the time-domain waveform and spectral energy distribution of each source signal, it classifies and labels them as dust accumulation attenuation features, bearing wear features, or capacitor aging features, and combines them to form an endogenous attenuation feature vector.

5. The new energy power facility testing data management and analysis platform according to claim 1, characterized in that, It also includes a multi-physics coupled state comprehensive diagnostic system, which integrates the following: The dust accumulation evolution analysis module is used to establish a regression model of photocurrent decline rate and cleaning cycle based on the dust accumulation decay characteristics in the endogenous decay feature vector and the irradiance data collected by the environmental sensing subsystem, in order to estimate the current dust accumulation density and automatically generate cleaning suggestions based on the comparison of power generation loss prediction and cleaning cost. The electrical physical performance evaluation module is used to derive the equivalent junction temperature of power semiconductor devices based on the current and voltage data in the electrical characteristic parameters, the shell temperature measured by the infrared thermal imaging sensor, and the thermal resistance characteristic parameters provided in the device manual, using a thermal resistance model, and triggering an aging warning when the equivalent junction temperature exceeds a preset threshold. The structural dynamics diagnostic module is used to extract microcrack propagation characteristic indicators based on vibration signals and acoustic emission signals in structural characteristic parameters through empirical mode decomposition and Hilbert transform, and to comprehensively derive fatigue damage indicators by calculating the centroid frequency offset of the vibration spectrum.

6. The new energy power facility testing data management and analysis platform according to claim 5, characterized in that, The multi-physics coupled state comprehensive diagnostic system also includes a fuzzy comprehensive evaluation module, which takes the dust density output by the dust evolution analysis module, the aging risk index output by the electrical physical performance evaluation module, and the fatigue damage index output by the structural dynamics diagnosis module as inputs. It performs nonlinear fusion through a preset membership function and fuzzy inference rule base, and outputs a comprehensive health index between 0 and 1 after defuzzification to quantitatively evaluate the overall health status of the equipment.

7. The new energy power facility testing data management and analysis platform according to claim 6, characterized in that, It also includes a full lifecycle preventative maintenance decision-making system, which is used for: Receive the comprehensive health index and establish a stochastic degradation model describing the evolution of health status over time based on the Wiener process; The drift coefficient and diffusion coefficient of the degradation model are updated online based on the newly acquired health index observations using a Bayesian method. Predict the remaining lifespan of the equipment based on the updated degradation model; Set multi-level early warning thresholds associated with the comprehensive health index, and implement differentiated response strategies for different levels of early warning, including increasing sampling frequency, initiating maintenance cost assessment, and automatically generating maintenance instructions that include a list of repair parts and spare parts.

8. The new energy power facility testing data management and analysis platform according to claim 7, characterized in that, When the whole life cycle preventive maintenance decision-making system executes a Level 2 warning, it activates a maintenance cost assessment mechanism; the maintenance cost assessment mechanism includes: Obtain meteorological forecast data, time-of-use electricity price curves, and equipment efficiency reduction factors determined by the current comprehensive health index; For multiple candidate maintenance dates within the next M days, calculate the expected net maintenance benefit for each candidate maintenance date. The expected net maintenance benefit is equal to the cumulative power generation loss avoided during the waiting period until the maintenance is carried out on the candidate maintenance date minus the corresponding maintenance cost. Among them, the avoidable cumulative power generation loss is determined based on the predicted equivalent full-load hours, time-of-use electricity price curve, and the corresponding relationship between the equipment efficiency reduction coefficient and the health index decay in the meteorological forecast data. The candidate maintenance date that maximizes the expected net maintenance benefit is determined as the optimal maintenance window, and a recommendation for this optimal maintenance window is output.

9. The new energy power facility testing data management and analysis platform according to claim 1, characterized in that, It also includes a data management center, which adopts a layered and decoupled architecture, including: A distributed time-series database is used to store high-frequency sampled environmental, electrical, and structural characteristic parameters, and to set the thermal data retention period. Relational databases are used to store static ledger information of devices and metadata generated by algorithm models; The publish-subscribe mechanism based on message queues is used to realize data exchange and loosely coupled communication between various system modules; Hardware encryption chips are configured at key transmission nodes to sign and encrypt transmitted data packets according to the national standard SM2 algorithm.

10. The new energy power facility testing data management and analysis platform according to claim 1, characterized in that, The environment-driven dynamic benchmark manifold construction system also includes a self-correction mechanism. This mechanism is used to automatically trigger a self-supervised learning program when a systematic deviation that cannot be explained by the existing model is detected in the residual sequence of the decoupled extraction system for endogeneous performance deviation. The program identifies new operating conditions through cluster analysis and uses incremental learning techniques to fine-tune the model in the local area of ​​the ideal output hypersurface to adapt to the operating condition drift caused by environmental changes.

Images