A photovoltaic inverter abnormal operation fault diagnosis system

Abstract

This invention relates to the field of intelligent operation and maintenance technology for new energy power generation and power electronic equipment, specifically a fault diagnosis system for photovoltaic inverter operation anomalies. The system includes: an edge data acquisition and dynamic reconstruction unit that maps discrete time-series data into high-dimensional phase space trajectory data; a topological cross-section feature extraction unit that generates a discrete set of two-dimensional intersection coordinates; a topological fingerprint generation unit that generates a topological fingerprint matrix; a cloud-based topological inference unit that generates diagnostic status results; and an anomaly tracing closed-loop control unit that monitors the diagnostic status results in real time and generates a backtracking lock command when the diagnostic status result is abnormal or suspected of being faulty, triggering the edge controller to stop overwriting the circular buffer and lock the corresponding high-frequency original waveform data. This invention enables resource-constrained microcontrollers to complete feature extraction within microseconds, ensuring the real-time performance of edge-side processing.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for new energy power generation and power electronic equipment, specifically a fault diagnosis system for abnormal operation of photovoltaic inverters. Background Technology

[0002] With the transformation of the global energy structure, the deployment scale and equipment complexity of photovoltaic power generation systems have increased significantly. As the core energy conversion unit of photovoltaic power generation systems, the operating status of photovoltaic inverters directly affects the safety and stability of the power grid.

[0003] Currently, the operation monitoring and fault diagnosis of photovoltaic inverters mainly rely on regular maintenance by technicians and threshold alarms based on time-domain signals. Technicians typically use equipment such as oscilloscopes to collect voltage and current waveforms of the inverters and determine the operating status of the equipment through manual analysis or simple time-series modeling, recording various parameters of the equipment operation. However, traditional time-domain waveform analysis methods are difficult to effectively reveal early, weak fault signs in nonlinear dynamic systems. In high-frequency sampling environments, edge controllers face severe bottlenecks in central processing unit interrupt handling and are limited by hardware computing power, making it difficult to perform complex high-dimensional matrix operations locally. At the same time, the communication bandwidth in industrial sites is often insufficient to support the real-time full upload of massive amounts of raw waveform data. In the existing architecture, due to the physical delay between cloud inference and edge response, the key raw data at the moment of a fault often gets overwritten by subsequent sampling before the traceback command is issued, resulting in the loss of evidence at the fault site and making it difficult to conduct in-depth anomaly tracing and cause analysis. Therefore, how to achieve accurate identification of early, weak faults in resource-constrained edge environments and narrowband communication environments, and ensure the complete traceability of high-frequency raw data at abnormal moments, has become an urgent problem to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a fault diagnosis system for abnormal operation of photovoltaic inverters. Specifically, the technical solution of this invention includes:

[0005] A fault diagnosis system for abnormal operation of a photovoltaic inverter includes:

[0006] The edge data acquisition and dynamics reconstruction unit is used to acquire discrete time series data output by the sensor in real time in the circular buffer of the edge controller through the direct memory access mechanism, and to map the discrete time series data into high-dimensional phase space trajectory data by introducing a preset time delay parameter and embedding dimension based on the Takens embedding principle.

[0007] The topological section feature extraction unit is used to receive the high-dimensional phase space trajectory data, calculate the two-dimensional intersection coordinates of the high-dimensional phase space trajectory data and the preset virtual Poincaré section, and generate a discrete set of two-dimensional intersection coordinates.

[0008] The topological fingerprint generation unit is used to divide the coordinate space of the discrete two-dimensional intersection coordinate set into a grid of preset resolution, count the number of intersections falling into each grid of preset resolution, convert the number of intersections into a density heatmap matrix, and generate a topological fingerprint matrix.

[0009] The cloud-based topology inference unit is used to receive the topology fingerprint matrix, classify and determine the density heatmap matrix using a pre-trained lightweight convolutional neural network, and generate diagnostic status results.

[0010] An anomaly tracing closed-loop control unit is used to monitor the diagnostic status results in real time, and generate a backtracking lock command when the diagnostic status results are abnormal or suspected faults, triggering the edge controller to stop the overwrite operation on the circular buffer and lock the high-frequency original waveform data at the corresponding time.

[0011] Optionally, the edge data acquisition and dynamics reconstruction unit is specifically used for:

[0012] The time delay parameter is set to a preset proportion of the system's main frequency cycle to ensure that the geometric unfolding of the reconstructed trajectory is maximized.

[0013] By applying the time delay parameter and the embedding dimension, the one-dimensional discrete time series data is reconstructed into a sequence of state vectors in a high-dimensional state space.

[0014] Optionally, the topological section feature extraction unit is specifically used for:

[0015] The embedding dimension is set to 3 to construct a three-dimensional state space;

[0016] A plane in the three-dimensional state space with a fixed state variable in one dimension is selected as the preset virtual Poincaré section;

[0017] The spatial position of the preset virtual Poincaré section is set to bisect the limit loop trajectory of the system during normal operation;

[0018] The intersection point of the high-dimensional phase space trajectory data when it passes through the preset virtual Poincaré section is calculated using a linear interpolation algorithm.

[0019] Optionally, the topology fingerprint generation unit is specifically used for:

[0020] The size of the preset resolution grid is set based on the balance between diagnostic accuracy and bandwidth limitations;

[0021] Generate the density heatmap matrix, where each element value represents the frequency density of the high-dimensional phase space trajectory data passing through the corresponding grid region;

[0022] The topological fingerprint matrix is ​​sent to the cloud or the main control unit via the communication module.

[0023] Optionally, the cloud-based topology inference unit is specifically used for:

[0024] If the density heatmap matrix shows tight single-point clustering, the diagnostic state result is determined to be a normal steady state;

[0025] If the density heatmap matrix shows point cloud diffusion or blurred edges, the diagnostic status result is determined to be early aging or parameter drift.

[0026] If the density heatmap matrix shows point cloud splitting or new cluster centers appear, the diagnostic status result is determined to be a structural fault.

[0027] Optionally, the anomaly tracing closed-loop control unit is specifically used for:

[0028] When the diagnostic status result is normal, maintain the low bandwidth characteristic transmission mode;

[0029] When the backtracking lock instruction is generated, the backtracking lock instruction is sent to the edge controller;

[0030] The edge controller performs a backtracking operation based on the timestamp in the backtracking lock instruction;

[0031] The high-frequency raw waveform data temporarily stored in the circular buffer, corresponding to the time before and after the abnormal moment, is persistently transferred.

[0032] Optionally, the circular buffer is configured to have a preset time depth, which is greater than the maximum round-trip communication latency required for data feature uploading, cloud inference, and backtracking lock command issuance.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This invention effectively reveals early, subtle fault signs in nonlinear systems through dynamic reconstruction and topological section feature extraction. The system utilizes phase space reconstruction technology to transform time-domain cluttered signals into geometric attractors with specific topological structures, and uses virtual Poincaré sections to remove redundant information, retaining only the dynamic features reflecting the system's periodicity and stability. This makes the system highly sensitive to subtle faults such as voltage distortion and parameter drift, significantly improving identification accuracy.

[0035] 2. This invention realizes a paradigm shift from massive data transmission to lightweight feature transmission, significantly reducing communication bandwidth usage; by transforming high-dimensional trajectory intersections into a density heatmap matrix, it achieves an extremely high data reduction ratio, effectively alleviating network congestion problems in industrial settings; this probability density-based feature carrier not only breaks through narrowband limitations but also uses statistical properties to filter out occasional electromagnetic interference, ensuring high-frequency, high-reliability real-time monitoring even in low-bandwidth environments;

[0036] 3. This invention ingeniously solves the engineering challenge of insufficient computing power in edge devices, enabling high-dimensional analysis with low resource consumption; it adopts a direct memory access mechanism to independently collect data, avoiding the computing power bottleneck of the processor under high-frequency interrupts; at the same time, it uses geometric cutting and linear interpolation algorithms to replace the solution of complex differential equations, greatly reducing the time complexity of the algorithm; this enables resource-constrained microcontrollers to complete feature extraction in microseconds, ensuring the real-time performance of edge processing.

[0037] 4. This invention establishes an intelligent strategy for monitoring normal characteristics and tracing anomalies on demand, ensuring the integrity and traceability of original fault data. By configuring a circular buffer with a depth greater than the system's round-trip communication delay and establishing a reverse locking mechanism for anomaly triggering, the risk of data being overwritten in a loop is eliminated. This design ensures that engineers can obtain uncompressed, high-fidelity original waveform data during the critical window of fault occurrence, providing irreplaceable data support for subsequent fault reproduction and responsibility identification. Attached Figure Description

[0038] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0039] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0041] Example 1:

[0042] Please see Figure 1 A fault diagnosis system for abnormal operation of a photovoltaic inverter, comprising:

[0043] The edge data acquisition and dynamics reconstruction unit is used to acquire discrete time series data output by the sensor in real time in the circular buffer of the edge controller through the direct memory access mechanism, and to map the discrete time series data into high-dimensional phase space trajectory data by introducing a preset time delay parameter and embedding dimension based on the Takens embedding principle.

[0044] The topological section feature extraction unit is used to receive the high-dimensional phase space trajectory data, calculate the two-dimensional intersection coordinates of the high-dimensional phase space trajectory data and the preset virtual Poincaré section, and generate a discrete set of two-dimensional intersection coordinates.

[0045] The topological fingerprint generation unit is used to divide the coordinate space of the discrete two-dimensional intersection coordinate set into a grid of preset resolution, count the number of intersections falling into each grid of preset resolution, convert the number of intersections into a density heatmap matrix, and generate a topological fingerprint matrix.

[0046] The cloud-based topology inference unit is used to receive the topology fingerprint matrix, classify and determine the density heatmap matrix using a pre-trained lightweight convolutional neural network, and generate diagnostic status results.

[0047] An anomaly tracing closed-loop control unit is used to monitor the diagnostic status results in real time, and generate a backtracking lock command when the diagnostic status results are abnormal or suspected faults, triggering the edge controller to stop the overwrite operation on the circular buffer and lock the high-frequency original waveform data at the corresponding time.

[0048] In this embodiment, a fault diagnosis system for photovoltaic inverter operation abnormalities is proposed. Its design concept is to construct an asymmetric closed-loop diagnostic architecture that spans resource-constrained edge computing and high-computing-power cloud computing. The edge data acquisition and dynamics reconstruction unit, serving as the interface between the physical world and digital space, resides in the inverter's DSP or ARM controller. To avoid the CPU's computational bottleneck when handling high-frequency interrupts, this unit is configured with a direct memory access channel, independent of the CPU core, to cyclically write the discrete-time sequence data (after ADC conversion) of the voltage or current analog signals output by the Hall sensor into a pre-allocated memory ring buffer. The task performed by this unit is the up-dimensional reconstruction of dynamic characteristics, aiming to solve the technical pain point that traditional time-domain waveform analysis cannot effectively reveal early weak fault symptoms in nonlinear systems. Based on Takens' embedding theorem in nonlinear dynamics, this unit introduces a preset time delay parameter and embedding dimension to mathematically map the scalar data sequence originally laid flat on a one-dimensional time axis into a vector sequence in a multi-dimensional state space, generating high-dimensional phase space trajectory data. High-dimensional phase space trajectory data refers to a type of... The set of vectors in the Euclidean space that describes the evolution path of the system state serves to transform signals that appear chaotic or noise-affected in the time domain into geometric attractors with specific topological structures. Its origin is the recombination of the original one-dimensional time-series signal after multiple delays.

[0049] The topological section feature extraction unit takes over the aforementioned high-dimensional data and is designed to solve the engineering problem of insufficient computing power at the edge to support high-dimensional matrix operations. This unit does not analyze the entire trajectory, but instead uses a geometric cutting strategy to set a preset virtual Poincaré section in the reconstructed high-dimensional space. When the high-dimensional trajectory evolves in the state space and passes through this virtual section, the unit captures its penetration point, thereby generating a set of discrete two-dimensional intersection coordinates. This processing is essentially a dimensionality reduction filter, which removes redundant information in the time dimension and linear correlations in continuous trajectories, retaining only the dynamic characteristics that reflect the periodicity, stability, and nonlinear structure of the system.

[0050] A topological fingerprint generation unit intervenes to perform spatial discretization and density mapping of features. This unit divides the continuous two-dimensional coordinate space where the intersection points are located into a discrete grid system, counts the number of intersection points falling into each tiny grid region, and then transforms it into a density heatmap matrix to generate a topological fingerprint matrix. The topological fingerprint matrix refers to a highly compressed two-dimensional data structure containing spatial distribution probability information. The value of each element represents the frequency density of the system state trajectory in that specific phase space region. Its role is to replace the original massive data transmission with a lightweight feature carrier, realizing a paradigm shift from data transmission to feature transmission.

[0051] The cloud-based topology inference unit utilizes a pre-trained lightweight convolutional neural network deployed on the server to perform image-based pattern recognition on the received matrix. This unit transforms the power electronics fault diagnosis problem into a computer vision classification problem, and directly generates diagnostic status results by recognizing the texture and clustering patterns of the heat map.

[0052] The closed-loop control of the system is completed by the anomaly tracing closed-loop control unit. This unit continuously scans the inference results in the cloud. Once it captures an anomaly or suspected fault judgment signal, it immediately generates a backtracking lock command and routes it backward to the edge. The edge controller responds to this high-priority command, utilizes the historical data temporary storage feature of the ring buffer, immediately suspends the current write pointer, stops the overwrite operation, and locks the high-frequency raw waveform data at the corresponding time according to the timestamp.

[0053] This invention establishes a topological characteristic of extremely low bandwidth transmission under normal conditions and an intelligent strategy of on-demand backtracking of high-fidelity original data under abnormal conditions. This design effectively overcomes the dual physical constraints of insufficient computing power of edge devices and narrow communication bandwidth in industrial fields. While reducing the daily data transmission volume by two orders of magnitude, it ensures high-precision identification of early minor faults and the integrity and traceability of original waveform evidence after a fault occurs.

[0054] Example 2:

[0055] The edge data acquisition and dynamic reconstruction unit is specifically used to: set the time delay parameter as a preset proportion of the system's main frequency period to ensure that the geometric expansion of the reconstructed trajectory is maximized; and apply the time delay parameter and the embedding dimension to reconstruct the one-dimensional discrete time series data into a state vector sequence in a high-dimensional state space.

[0056] In this embodiment, clear engineering standards are established for the key parameter configuration of the edge data acquisition and dynamic reconstruction unit. To maximize the removal of autocorrelation from the nonlinear signal in the reconstructed phase space trajectory, thereby clearly expressing the system's dynamic characteristics, the unit follows the principle of minimizing mutual information when setting the time delay parameter. In actual engineering configurations, this parameter is strictly locked to a specific proportion of the system's main frequency period, typically set to a value... The edge controller is equipped with a real-time frequency monitoring module to periodically acquire the current fundamental frequency of the power grid. The calculation logic of the time delay parameter is as follows: divide the current sampling frequency by four times the fundamental frequency of the power grid, and round the result to the nearest integer to obtain the integer value of the corresponding number of sampling points. When the change amplitude of the fundamental frequency of the power grid exceeds a preset threshold, the system automatically updates the time delay parameter using the above logic to ensure that the reconstructed trajectory always remains in an orthogonal unfolding state.

[0057] The time delay parameter here specifically refers to the number of sampling points used to construct the time interval between components when constructing the delayed coordinate vector. Its function is to determine whether the geometry of the attractor trajectory in the reconstructed phase space is fully expanded. If this parameter is set too small, the trajectory will collapse near the diagonal, forming a flat structure; if it is set too large, the trajectory structure will become folded and disordered due to excessive decorrelation. By setting it to a value that corresponds to the main frequency period... By utilizing the principle of sinusoidal orthogonality, the geometric expansion of the reconstructed trajectory in phase space is maximized, thereby exposing the nonlinear characteristics of the system to the fullest extent.

[0058] Based on this parameter baseline, the unit performs vectorization operations for each time point. Data Combined with embedding dimension and delay time Construct a vector This transforms the originally flat one-dimensional discrete time series data point by point into a state vector sequence in a high-dimensional state space. By executing the above parameter setting and reconstruction logic, feature aliasing or distortion caused by arbitrary parameter selection is effectively avoided, ensuring that subsequent feature extraction steps can be performed based on high-quality, high-discrimination dynamic trajectories, and significantly improving the system's sensitivity to weak fault signs such as voltage distortion and harmonic injection.

[0059] Example 3:

[0060] The topological section feature extraction unit is specifically used to: set the embedding dimension as... A three-dimensional state space is constructed; a plane in the three-dimensional state space with a fixed state variable of one dimension is selected as the preset virtual Poincaré section; the spatial position of the preset virtual Poincaré section is set to intercept the limit loop trajectory when the system is running normally; and the intersection point of the high-dimensional phase space trajectory data when it passes through the preset virtual Poincaré section is calculated using a linear interpolation algorithm.

[0061] This embodiment details the dimensionality reduction algorithm logic executed by the topological section feature extraction unit; the embedding dimension is explicitly set in the algorithm configuration of this unit. This constructs a geometrically analyzable three-dimensional Euclidean state space. Within this spatial framework, to accurately capture the periodic fluctuation characteristics of the system, the unit selects a specific dimension of the state vector and fixes it to a numerical value. The plane is defined as a preset virtual Poincaré section; the spatial position of this section is not arbitrary, but is determined through system steady-state analysis to ensure that its spatial coordinates can laterally truncate the closed limit loop trajectory formed by the system under normal inverter operation.

[0062] In real-time operation, when continuous high-dimensional phase space trajectory data flows through the space, the unit monitors the positional relationship of the trajectory points relative to the cross-sectional plane. Once it detects that two consecutive sampling points are distributed on both sides of the cross-section and exhibit a unidirectional crossing feature, such as when the state variable value increases from small to large, the unit initiates a linear interpolation algorithm. Based on the coordinates of these two points and the cross-sectional equation, it quickly resolves the precise intersection coordinates of the trajectory line segment and the cross-sectional plane. By employing the aforementioned three-dimensional spatial cross-sectional mapping and linear interpolation techniques, this embodiment cleverly reduces the dimensionality of complex differential equation dynamic analysis to simple geometric intersection operations. This processing method greatly reduces the time complexity of the algorithm and its dependence on floating-point arithmetic units, enabling resource-constrained edge microcontrollers to complete feature extraction within microseconds, while effectively preserving the topological structure information of the system state, providing a high signal-to-noise ratio input for subsequent pattern recognition.

[0063] Example 4:

[0064] The topology fingerprint generation unit is specifically used for: setting the size of the preset resolution grid based on the balance between diagnostic accuracy and bandwidth limitations; obtaining the amplitude extreme values ​​of the high-dimensional phase space trajectory data within the current time window, and using the amplitude extreme values ​​to perform spatial normalization processing on the two-dimensional intersection point coordinate set, mapping all intersection point coordinates to the unit coordinate system; generating the density heatmap matrix, where each element value represents the frequency density of the high-dimensional phase space trajectory data passing through the corresponding grid region; extracting the start and end timestamps of the original discrete time series data corresponding to the density heatmap matrix, and encapsulating the start timestamp, the end timestamp, and the density heatmap matrix into a topology fingerprint data packet; and sending the topology fingerprint data packet to the cloud or main control unit through the communication module.

[0065] In this embodiment, the topological fingerprint generation unit implements the data discretization, quantization, and dimensionality reduction mapping process; the configuration strategy of this unit is based on a trade-off between diagnostic resolution requirements and on-site narrowband communication limitations, setting a preset resolution grid size, with a typical implementation value of [value missing]. or The matrix grid; the preset resolution grid is a tool for spatial quantization of continuous phase space cross-sectional coordinates. Its function is to map floating-point coordinates with infinite precision into a finite integer index region. In the processing flow, this unit traverses all cross-sectional intersections within a certain time window, calculates the grid index to which each intersection belongs, and accumulates the number of points falling into each grid region. By performing spatial normalization operations before grid mapping, the correlation between system operating power and topological characteristics is decoupled, ensuring that the geometric shape of the reconstructed trajectory is not distorted due to amplitude scaling under photovoltaic inverter power fluctuation conditions.

[0066] This unit normalizes statistical values, converting them into frequency density values ​​to fill the generated density heatmap matrix. The specific logic of this transformation process is as follows: The number of intersections falling into a specific grid region within the current statistical time window is obtained, and this number is divided by the sum of the number of intersections in all grid regions within that time window to obtain the normalized frequency density value for that specific grid region. This process ensures that the matrix element values ​​are strictly distributed between zero and one, eliminating the influence of differences in the total data volume caused by different sampling durations on the feature matrix values. In this matrix, each element value no longer carries specific voltage or current physical quantities, but rather represents the probabilistic statistical characteristics of the system state trajectory traversing the corresponding phase space region. This unit drives the communication module to send this topological fingerprint matrix, containing only a few hundred bytes, to the cloud. Through the above gridding and density mapping processing, this embodiment achieves a high information content of up to [amount missing] from the original high-frequency sampling data through feature extraction. to The dimensionality reduction greatly reduces the data load of IoT communication; at the same time, this probability density-based heatmap representation has natural low-pass filtering characteristics, which can effectively filter out occasional electromagnetic interference noise and retain only the statistically significant system dynamic structural features.

[0067] Example 5:

[0068] The cloud-based topology inference unit is specifically used for: if the density heatmap matrix shows tight single-point clustering, determining the diagnostic state result as normal steady state; if the density heatmap matrix shows point cloud diffusion or blurred edges, determining the diagnostic state result as early aging or parameter drift; if the density heatmap matrix shows point cloud splitting or the appearance of new cluster centers, determining the diagnostic state result as structural failure.

[0069] In this embodiment, the cloud-based topology inference unit adopts a set of morphological decision criteria based on the physical meaning of nonlinear dynamics. After receiving the density heatmap matrix uploaded from the edge, the unit analyzes its topological distribution morphology: if the heatmap shows that high-density areas are concentrated in a tight single-point cluster, this physically maps to the system trajectory running stably and repeatedly on a predetermined limiting cycle in phase space, and the unit determines that the system is in a normal steady state; if the heatmap shows that the point cloud spreads outward, the area increases, or the edge contours are blurred, this physically means that the internal parameters of the system have drifted, causing the stability of the attraction domain of the limiting cycle to decrease but not yet breaking the closure, and the unit determines the result as early aging or parameter drift.

[0070] If the heatmap shows significant point cloud splitting, multi-peak distribution, or the emergence of new cluster centers outside the main cluster, this physically indicates that the dynamic structure of the system has undergone topological abrupt change or bifurcation, and the unit determines the result as a structural fault. By establishing this deterministic mapping relationship between topological form and physical state, this embodiment not only achieves fully automated intelligent diagnosis, but also endows the artificial intelligence algorithm with physical interpretability. It can accurately distinguish between external sensor noise and internal system faults, thereby significantly reducing the false alarm rate and achieving highly robust fault identification.

[0071] Example 6:

[0072] The anomaly tracing closed-loop control unit is specifically used for: maintaining a low-bandwidth feature transmission mode when the diagnostic status result is normal; when the backtracking lock instruction is generated, parsing the start and end timestamps in the topological fingerprint data packets determined to be abnormal, encapsulating them in the backtracking lock instruction, and sending the backtracking lock instruction to the edge controller; the edge controller calculating the memory offset address of the circular buffer based on the start and end timestamps in the backtracking lock instruction and performing a backtracking operation; and persistently transferring the high-frequency raw waveform data temporarily stored in the circular buffer corresponding to the time before and after the abnormality.

[0073] In this embodiment, the anomaly tracing closed-loop control unit defines the system's dynamic response and data governance mechanisms. During system operation, when the diagnostic status result is determined to be normal, the unit instructs the system to maintain a low-bandwidth feature transmission mode, only periodically uploading a lightweight topology fingerprint matrix to minimize communication resource consumption. When the cloud inference result triggers an anomaly alarm and generates a backtracking lock instruction, the unit immediately sends this instruction to the edge controller through the downlink control channel. After receiving the instruction, the edge controller starts the interrupt service routine, parses the anomaly occurrence timestamp carried in the instruction, calculates the memory offset address in the circular buffer at that moment, and performs a pointer backtracking operation in the past time direction.

[0074] Once the system locates the memory block within a preset range before and after the time of the anomaly, it immediately triggers direct memory access or memory copy logic to transfer the high-frequency raw waveform data temporarily stored in the volatile buffer to a non-volatile storage area for locking and protection, preventing it from being overwritten by subsequent real-time data streams. Through this data governance strategy of normal discarding and anomaly locking, this embodiment completely solves the contradiction of the inability to fully store and transmit massive amounts of high-frequency data in industrial sites. It ensures that during the critical time window of the fault, engineers can obtain raw waveform data with the highest fidelity, which has not been processed by any compression algorithm, thus providing irreplaceable black-box-level data support for subsequent fault reproduction, responsibility identification, and in-depth cause analysis.

[0075] Example 7:

[0076] The circular buffer is configured to have a preset time depth, which is greater than the maximum round-trip communication latency required for data feature uploading, cloud inference, and backtracking lock command issuance.

[0077] In this embodiment, the physical capacity configuration of the circular buffer is subject to strict timing constraints. To ensure the logical effectiveness of the anomaly tracing mechanism and prevent critical historical fault data from being overwritten by new sampled data before the cloud-based backtracking command reaches the edge, the circular buffer is configured with a specific preset time depth. The preset time depth refers to the length of time that the buffer memory space can continuously store historical sampled data. In hardware implementation, the circular buffer is usually mapped to a large-capacity SRAM or DDR memory external to the edge controller to overcome the on-chip RAM capacity limit. Its value is set based on the communication link analysis under the worst-case operating conditions and must be strictly greater than the maximum round-trip communication latency of the entire closed-loop diagnostic link. This latency consists of three parts: the network transmission time for edge feature extraction and uploading, the inference calculation time of the cloud model, and the network transmission time for issuing the backtracking command.

[0078] By setting this buffer depth threshold based on latency analysis, this embodiment eliminates the risk of data write-off or overwriting at the hardware architecture level, ensuring that no matter how the network environment fluctuates or how the cloud load changes, as long as it is within the maximum design latency range, the system can always reliably capture the original field data at the moment the fault occurs, thereby ensuring the reliability and data integrity of the diagnostic system.

[0079] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fault diagnosis system for abnormal operation of a photovoltaic inverter, characterized in that, include: The edge data acquisition and dynamics reconstruction unit is used to acquire discrete time series data output by the sensor in real time in the circular buffer of the edge controller through the direct memory access mechanism, and to map the discrete time series data into high-dimensional phase space trajectory data by introducing a preset time delay parameter and embedding dimension based on the Takens embedding principle. The topological section feature extraction unit is used to receive the high-dimensional phase space trajectory data, calculate the two-dimensional intersection coordinates of the high-dimensional phase space trajectory data and the preset virtual Poincaré section, and generate a discrete set of two-dimensional intersection coordinates. The topological fingerprint generation unit is used to divide the coordinate space of the discrete two-dimensional intersection coordinate set into a grid of preset resolution, count the number of intersections falling into each grid of preset resolution, convert the number of intersections into a density heatmap matrix, and generate a topological fingerprint matrix. The cloud-based topology inference unit is used to receive the topology fingerprint matrix, classify and determine the density heatmap matrix using a pre-trained lightweight convolutional neural network, and generate diagnostic status results. An anomaly tracing closed-loop control unit is used to monitor the diagnostic status results in real time, and generate a backtracking lock command when the diagnostic status results are abnormal or suspected faults, triggering the edge controller to stop the overwrite operation on the circular buffer and lock the high-frequency original waveform data at the corresponding time. The topological fingerprint generation unit is specifically used for: The size of the preset resolution grid is set based on the balance between diagnostic accuracy and bandwidth limitations; Generate the density heatmap matrix, where each element value represents the frequency density of the high-dimensional phase space trajectory data passing through the corresponding grid region; The topology fingerprint matrix is ​​sent to the cloud or main control unit via the communication module; The cloud-based topology inference unit is specifically used for: If the density heatmap matrix shows tight single-point clustering, the diagnostic state result is determined to be a normal steady state; If the density heatmap matrix shows point cloud diffusion or blurred edges, the diagnostic status result is determined to be early aging or parameter drift. If the density heatmap matrix shows point cloud splitting or new cluster centers appear, the diagnostic status result is determined to be a structural fault.

2. The photovoltaic inverter operation abnormality fault diagnosis system according to claim 1, characterized in that, The edge data acquisition and dynamics reconstruction unit is specifically used for: The time delay parameter is set to a preset proportion of the system's main frequency cycle to ensure that the geometric unfolding of the reconstructed trajectory is maximized. By applying the time delay parameter and the embedding dimension, the one-dimensional discrete time series data is reconstructed into a sequence of state vectors in a high-dimensional state space.

3. The photovoltaic inverter operation abnormality fault diagnosis system according to claim 1, characterized in that, The topological section feature extraction unit is specifically used for: The embedding dimension is set to 3 to construct a three-dimensional state space; A plane in the three-dimensional state space with a fixed state variable in one dimension is selected as the preset virtual Poincaré section; The spatial position of the preset virtual Poincaré section is set to bisect the limit loop trajectory of the system during normal operation; The intersection point of the high-dimensional phase space trajectory data when it passes through the preset virtual Poincaré section is calculated using a linear interpolation algorithm.

4. The photovoltaic inverter operation abnormality fault diagnosis system according to claim 1, characterized in that, The anomaly tracing closed-loop control unit is specifically used for: When the diagnostic status result is normal, maintain the low bandwidth characteristic transmission mode; When the backtracking lock instruction is generated, the backtracking lock instruction is sent to the edge controller; The edge controller performs a backtracking operation based on the timestamp in the backtracking lock instruction; The high-frequency raw waveform data temporarily stored in the circular buffer, corresponding to the time before and after the abnormal moment, is persistently transferred.

5. The photovoltaic inverter operation abnormality fault diagnosis system according to claim 1, characterized in that, The circular buffer is configured to have a preset time depth, which is greater than the maximum round-trip communication latency required for data feature uploading, cloud inference, and backtracking lock command issuance.

Images