Photovoltaic module failure prediction and health management system based on multi-source internet of things data

Abstract

The present application relates to photovoltaic power generation equipment fault diagnosis and internet of things health management technical field, specifically to photovoltaic module fault prediction and health management system based on multi-source internet of things data, including: heterogeneous data acquisition module, for obtaining low-frequency environmental observation data and high-frequency state sampling data;Derivative gating evaluation module;Feature mask control module;Cross-domain covariance calculation module;Fault discrimination output module;The present application takes low-frequency environmental change as the control condition of high-frequency fault diagnosis, completes the joint judgment of heterogeneous data without time axis interpolation, and realizes the differentiation of environmental disturbance and real fault, reduces the false alarm rate and storage and computing resource occupation.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation equipment fault diagnosis and Internet of Things health management technology, specifically a photovoltaic module fault prediction and health management system based on multi-source Internet of Things data. Background Technology

[0002] Currently, photovoltaic module fault prediction and health management are usually based on single electrical sampling data or by forcibly aligning environmental data with status data in time. When external environmental conditions such as irradiance and temperature change rapidly, existing solutions struggle to effectively distinguish between normal fluctuations caused by environmental disturbances and real faults caused by modules and sampling links. Each abnormal trigger may misjudge a large number of high-frequency fluctuations caused by environmental changes as faults, reducing the accuracy of fault diagnosis and increasing the storage and computing resource consumption on the edge side. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a photovoltaic module fault prediction and health management system based on multi-source Internet of Things data. Specifically, the technical solution of this invention includes:

[0004] The heterogeneous data acquisition module is used to acquire low-frequency environmental observation data and high-frequency state sampling data from edge node devices. The sampling rate of the low-frequency environmental observation data is lower than that of the high-frequency state sampling data.

[0005] The derivative-gated evaluation module is used to calculate the rate of environmental change of low-frequency environmental observation data and compare the rate of environmental change with a preset stationarity threshold.

[0006] The feature mask control module is used to generate a feature mask signal and inject it into the high-frequency state sampling domain that processes high-frequency state sampling data when the rate of change of the environment is greater than the stationary threshold. Based on the feature mask signal, the fault judgment logic based on the preset high-frequency transient threshold is suspended, and the accompanying time window is started.

[0007] The cross-domain covariance calculation module is used to perform integral operations on high-frequency state sampling data within the accompanying time window to obtain high-frequency integral values, calculate the state change of the high-frequency integral values ​​relative to the preset historical stationary integral values, construct cross-domain feature vectors by combining the fluctuation range of low-frequency environmental observation data within the accompanying time window, and calculate the cross-domain covariance matrix based on the cross-domain feature vectors.

[0008] The fault identification output module is used to compare the cross-domain covariance matrix with the preset healthy reference manifold representing the historical normal operation state. When the cross-domain covariance matrix deviates from the healthy reference manifold, a system fault alarm command is generated, and when the cross-domain covariance matrix does not deviate from the healthy reference manifold, an environmental disturbance normal label is generated.

[0009] The fault identification output module is also used to monitor the instantaneous fluctuation rate of high-frequency state sampling data when the rate of environmental change is less than or equal to the stability threshold. When a sudden change in data is detected, a suspected hardware fault interrupt signal is generated; if no sudden change in data is detected, an environmental disturbance normal label is output.

[0010] Preferably, the heterogeneous data acquisition module is used to acquire low-frequency environmental physical quantities of the edge node device and use the low-frequency environmental physical quantities as low-frequency environmental observation data; acquire high-frequency electrical physical quantities of the edge node device and use the high-frequency electrical physical quantities as high-frequency state sampling data; wherein, the low-frequency environmental observation data and the high-frequency state sampling data maintain the original asynchronous sampling state and are not interpolated and aligned on the time axis.

[0011] Preferably, the edge node device is a photovoltaic module device, and the low-frequency environmental physical quantities include irradiance data and temperature data, while the high-frequency electrical physical quantities include transient current data and transient voltage data.

[0012] Preferably, the derivative-gated evaluation module is used to extract the time series of low-frequency environmental observation data, perform approximate differentiation processing on the time series using a first-order difference algorithm, obtain the difference result, and use the difference result as the environmental change rate.

[0013] Preferably, the feature mask control module is used to trigger an internal hardware interrupt signal when the rate of change of the environment is greater than the stability threshold, and uses the internal hardware interrupt signal as the feature mask signal; by changing the direct memory access transmission target address corresponding to the high-frequency state sampling data through the feature mask signal, the feature mask signal is injected into the high-frequency state sampling domain.

[0014] Preferably, the cross-domain covariance calculation module is used to perform moving average filtering on the high-frequency state sampling data to obtain denoised high-frequency data; within the time range of the accompanying time window, the denoised high-frequency data is integrated to obtain the high-frequency integral value; the difference between the high-frequency integral value and the preset historical stationary integral value is calculated, and the difference is used as the state change quantity.

[0015] Preferably, the cross-domain covariance calculation module is used to calculate the fluctuation range of low-frequency environmental observation data within the accompanying time window, and to use the fluctuation range as a low-frequency change quantity; extract the state change quantity and the low-frequency change quantity to construct a cross-domain feature vector; and calculate the cross-domain covariance matrix based on the cross-domain feature vector.

[0016] Preferably, the fault identification output module is used to extract the eigenvectors of the cross-domain covariance matrix, calculate the Mahalanobis distance between the eigenvectors and the healthy baseline manifold, and compare the Mahalanobis distance with a preset manifold boundary threshold; when the Mahalanobis distance is greater than the manifold boundary threshold, it is determined that the cross-domain covariance matrix deviates from the healthy baseline manifold; when the Mahalanobis distance is less than or equal to the manifold boundary threshold, it is determined that the cross-domain covariance matrix does not deviate from the healthy baseline manifold.

[0017] Preferably, the fault identification output module is used to calculate the instantaneous volatility of the high-frequency state sampling data and compare the instantaneous volatility with a preset high-frequency mutation threshold; when the instantaneous volatility is greater than the high-frequency mutation threshold, it is determined that there is a data mutation in the high-frequency state sampling data; when the instantaneous volatility is less than or equal to the high-frequency mutation threshold, it is determined that there is no data mutation in the high-frequency state sampling data.

[0018] Preferably, the system further includes a baseline manifold update module, which is used to obtain the cross-domain covariance matrix corresponding to all environmental disturbance normal labels generated during the historical normal operation cycle; to form a historical covariance set from the cross-domain covariance matrices corresponding to all environmental disturbance normal labels; to recalculate the baseline manifold parameters based on the historical covariance set; and to adaptively update the healthy baseline manifold according to the baseline manifold parameters.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] This invention provides an edge-side diagnostic scheme that can distinguish between environmental disturbances and real faults by maintaining the original asynchronous sampling state of low-frequency environmental observation data and high-frequency state sampling data in edge nodes, using derivative gating and feature masking control of the high-frequency fault judgment logic with the low-frequency environmental change rate, and combining high-frequency integral values, state change quantities, low-frequency change quantities, cross-domain covariance matrices and health benchmark manifolds for joint discrimination within the accompanying time window. Compared with traditional methods based on single electrical sampling data or forced time alignment of environmental data and state data, this invention can reduce false alarms of high-frequency fluctuations caused by environmental abrupt changes, improve fault diagnosis accuracy, and reduce storage and computational overhead at the edge. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application:

[0022] Figure 1 This is a schematic diagram of a photovoltaic module fault prediction and health management system based on multi-source Internet of Things data provided in an embodiment of this application. Detailed Implementation

[0023] 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.

[0024] A photovoltaic module fault prediction and health management system based on multi-source IoT data is applied to edge node devices, including:

[0025] The heterogeneous data acquisition module is used to acquire low-frequency environmental observation data and high-frequency state sampling data from edge node devices, wherein the sampling rate of low-frequency environmental observation data is lower than the sampling rate of high-frequency state sampling data.

[0026] The derivative-gated evaluation module is used to calculate the rate of environmental change of low-frequency environmental observation data and compare the rate of environmental change with a preset stationarity threshold.

[0027] The feature mask control module is used to generate a feature mask signal and inject it into the high-frequency state sampling domain that processes high-frequency state sampling data if the rate of change of the environment is greater than the stability threshold. Based on the feature mask signal, the fault judgment logic based on the preset high-frequency transient threshold is suspended, and the accompanying time window is started.

[0028] The cross-domain covariance calculation module is used to perform integral operations on high-frequency state sampling data within the accompanying time window to obtain high-frequency integral values, calculate the state change of the high-frequency integral values ​​relative to the preset historical stationary integral values, construct cross-domain feature vectors by combining the fluctuation range of low-frequency environmental observation data within the accompanying time window, and calculate the cross-domain covariance matrix based on the cross-domain feature vectors.

[0029] The fault identification output module is used to compare the cross-domain covariance matrix with the preset healthy reference manifold representing the historical normal operation status. If the cross-domain covariance matrix deviates from the healthy reference manifold, a system fault alarm command is generated. If the cross-domain covariance matrix does not deviate from the healthy reference manifold, an environmental disturbance normal label is generated.

[0030] The fault identification output module is also used to monitor the instantaneous fluctuation rate of high-frequency state sampling data if the rate of change of the environment is less than or equal to the stability threshold. If a sudden change in data is detected, a suspected hardware fault interruption signal is generated; if no sudden change in data is detected, an environmental disturbance normal label is output.

[0031] This embodiment provides a photovoltaic module fault prediction and health management mechanism based on multi-source IoT data, such as Figure 1 As shown; specifically, this mechanism is deployed in the edge node devices on the photovoltaic power station array side, and the edge nodes can be microcontrollers with analog-to-digital conversion interfaces, serial bus interfaces, timers and direct memory access controllers;

[0032] The entire system operates continuously in the same industrial scenario: during normal power generation during the day, the edge nodes continuously access the environmental observation stream and electrical status stream of the same group of photovoltaic modules, and identify the source of anomalies at different stages such as rapid cloud cover, local wiring loosening, and sampling circuit drift;

[0033] In this embodiment, the sampling frequency range of low-frequency environmental observation data is defined as 0.1Hz to 1Hz, and the sampling frequency range of high-frequency state sampling data is defined as 1kHz to 10kHz. The sampling frequency of high-frequency state sampling data is at least 1000 times that of low-frequency environmental observation data to ensure effective decoupling of slow environmental variables and fast electrical variables on the time scale during cross-domain analysis.

[0034] Specifically, the edge node is equipped with a low-frequency environment observation domain and a high-frequency state sampling domain; the former receives environmental data at a first preset sampling frequency, and the latter receives state data at a second preset sampling frequency that is greater than the first preset sampling frequency; the two types of data maintain their original clock rhythms and are not forced to be interpolated and aligned.

[0035] After the derivative gating evaluation module forms a time series of low-frequency environmental observation data, it calculates the rate of change between two adjacent sampling points and compares it with the stationary threshold. The method for setting the stationary threshold is as follows: obtain the set of environmental change rates for multiple consecutive low-frequency sampling periods under the stable operating conditions of photovoltaic modules with no historical faults and cloud shading area less than the preset shading area threshold, extract the statistical upper limit value with a probability distribution confidence of 95% in the set, and add the preset system environmental tolerance margin to establish the sum as the stationary threshold.

[0036] If the rate of change of the environment is greater than the stability threshold, it indicates that the external environment is changing rapidly. At this time, the feature mask control module generates a feature mask signal and sends a control command to the high-frequency state sampling domain, which temporarily suspends the original conventional fault judgment logic for high-frequency transient spikes and starts a period of accompanying time window. This period of accompanying time window can be understood as an observation period defined around this environmental change, such as 3 or 5 seconds after the triggering time of the environmental change event.

[0037] The specific duration of the accompanying time window is determined by a comprehensive calculation based on the thermal inertia time constant of the photovoltaic modules under the jurisdiction of the edge node and the response transition time of the electrical measurement front end inside the edge node, ensuring that the length of the time window can fully cover the entire cycle of the transient response of the internal electrical state of the photovoltaic induced by the sudden change of external environmental parameters.

[0038] During the observation period, the cross-domain covariance calculation module does not directly rely on the instantaneous spike alarm, but instead performs integration processing on the high-frequency state sampling data to obtain the high-frequency integral value. Then, it compares the integral value with the preset historical stationary integral value to obtain the state change. After that, it combines the low-frequency environmental observation data within the accompanying time window to extract cross-domain features and form a cross-domain covariance matrix.

[0039] The fault identification output module then compares the matrix with the pre-established health baseline manifold: if there is a deviation, it indicates that the causal relationship between high-frequency state changes and environmental changes has been disrupted, and outputs a system fault alarm command; if there is no deviation, it indicates that the high-frequency fluctuation can be explained by environmental disturbances, and outputs an environmental disturbance normal label.

[0040] On the other hand, if the rate of change of the environment is less than or equal to the stability threshold, it is determined that the external environment is in a relatively stable stage. At this time, mask suppression is not enabled, and the instantaneous fluctuation characteristics of the high-frequency state sampling data are directly monitored. If a data mutation with a rate of change greater than the preset mutation threshold is detected in a stable environment, a suspected hardware fault interruption signal is generated to indicate problems such as loose wiring terminals, sensor drift, distortion of the sampling front end, or failure of local components.

[0041] In one specific embodiment, assuming that the low-frequency environmental observation data are 100, 103, and 150 at three times, with a sampling interval of 10 seconds, the change rates of the first two segments are 0.3 and 4.7, respectively; if the stability threshold is set to 1.0, the second segment is determined to be a drastic environmental change.

[0042] Meanwhile, several transient current values ​​within the time window are selected in the high-frequency domain, such as 8.0, 8.4, 8.2, and 8.3. After integration, a cumulative value is obtained, such as 33.0. The historical stable integral reference value is 32.5, so the state change is 0.5. Then, the environmental change within the low-frequency window is taken, such as from 103 to 150, so the low-frequency change is 47. At this time, a cross-domain feature pair can be formed, such as [0.5, 47]. If the feature pair falls near the healthy reference manifold, the environmental disturbance normal label is output. If an abnormal relationship such as [-1.2, 47] occurs, where the environment increases and the state decreases in the opposite direction, it is considered a deviation, and a system fault alarm command is output.

[0043] As an anomaly handling mechanism, if low-frequency environmental data is temporarily missing, the system maintains regular fault monitoring in the high-frequency domain, but simultaneously records the unavailable environmental data status in the local log to avoid misjudging missing conditions as equipment failures; if the rate of environmental change is exactly equal to the stationary threshold, it is classified into the stationary branch for processing to reduce frequent state switching caused by boundary jitter; if the high-frequency sampling packet within the accompanying time window is incomplete, the current window is invalidated and marked as an unreliable observation window, and no final fault conclusion is output based on this; if the health baseline manifold has not yet been initialized, only suspected events are output, and no deterministic alarms are output.

[0044] During the afternoon operation of a photovoltaic power station, a target module was monitored at the same edge node; at 13:20, the irradiance dropped rapidly from 620 to 410, and the corresponding rate of change exceeded the threshold; at the same time, the current also dropped suddenly.

[0045] Instead of immediately triggering an alarm, the system injects a feature mask into the high-frequency domain, suspends the original instantaneous threshold judgment, and calculates the cross-domain covariance within the next 3 seconds. After comparison, it is found that the current drop is consistent with the sudden drop in irradiance, so the system outputs a normal environmental disturbance label. At 13:47, the environment remained stable, but the high-frequency voltage suddenly showed abnormal glitches and continued to jump. The system directly generated a suspected hardware fault interrupt signal, which was used by the station's maintenance host to check the sampling circuit and connector status first.

[0046] The purpose of this step is to use low-frequency environmental changes as a control condition for high-frequency fault diagnosis rather than simply splicing data dimensions, and to complete the joint judgment of heterogeneous data without time axis interpolation. This enables the low-computing-power edge side to distinguish between environmental disturbances and real faults, and reduces the false alarm rate and the occupation of storage and computing resources.

[0047] Furthermore, the heterogeneous data acquisition module acquires low-frequency environmental observation data and high-frequency state sampling data from the edge node device. Specifically, it is used to: acquire low-frequency environmental physical quantities of the edge node device and use them as low-frequency environmental observation data; acquire high-frequency electrical physical quantities of the edge node device and use them as high-frequency state sampling data; wherein, the low-frequency environmental observation data and the high-frequency state sampling data maintain their original asynchronous sampling state and are not interpolated or aligned on the time axis.

[0048] This embodiment provides a heterogeneous data acquisition mechanism; specifically, in the aforementioned unified scenario, edge nodes acquire two types of data streams with significantly different frequencies from the environmental side interface and the electrical side interface, respectively, and maintain the original asynchronous sampling state in the cache layer without performing time axis interpolation alignment;

[0049] Specifically, low-frequency environmental physical quantities can be accessed through RS485 interface, industrial serial port or wireless slave station to form an environmental observation buffer; high-frequency electrical physical quantities can be directly input by local analog-to-digital converter, Hall sampling module or isolated sampling front end to form a state sampling buffer.

[0050] The two types of buffers each have independent timestamps or frame numbers, but the system does not copy low-frequency data to high-frequency points, nor does it downsample high-frequency data to align with the low-frequency sampling clock. Instead, it maintains a parallel structure of a slow clock domain and a fast clock domain. The reason for adopting this structure is that if the conventional approach is used to interpolate the environmental data once every 10 seconds onto the state sampling axis once every 1 millisecond, then an environmental point needs to be expanded into tens of thousands of pseudo-alignment points. This not only causes the memory usage to exceed the preset static memory capacity limit of the edge node, but also these expanded points do not increase the amount of real information.

[0051] Under the condition of limited computing power and static random access memory at edge nodes, this approach is prone to cache crowding, which weakens the ability to retain truly critical high-frequency segments. Therefore, this embodiment retains the environmental flow as a low-frequency control clue and the state flow as a high-frequency fault observation clue.

[0052] In one specific embodiment, it is assumed that within a certain minute, environmental observations only collect 6 points: E1 to E6; while high-frequency state streams collect 60,000 points: H1 to H60,000. Conventional interpolation methods require constructing a pseudo-environment sequence of length 60,000. In this embodiment, only E1 to E6 themselves and their time sequence are saved, and the processing of local windows of H is triggered based on the changes between E in the subsequent gating stage. In other words, the system does not forcibly copy and extend low-frequency environmental data to all high-frequency time nodes, but uses low-frequency environmental data to determine the triggering timing for processing high-frequency state segments.

[0053] As an exception handling mechanism, if the communication of the environment-side interface is interrupted for a short time, the low-frequency buffer maintains the most recent valid sample and its validity period mark; once the validity period expires, the subsequent gating logic enters the environment unknown mode; if a sudden packet loss occurs in the high-frequency sampling, the packet sequence number is retained to facilitate the subsequent window to discard discontinuous segments; if there is drift between the two clock sources, there is no need to perform full resampling, it is only necessary to establish the observation relationship based on the most recent valid time mark when the event occurs.

[0054] Within the same power station edge node, the meteorological station uploads irradiance and temperature every 10 seconds, and the current and voltage of the target component are sampled every 1 millisecond. When the cloud shadows pass quickly in the afternoon, the system only records the fact that the current environmental point has changed rapidly and uses it as a gating trigger condition to extract adjacent local segments from the high-frequency buffer for analysis, instead of expanding the 10-second environmental point into 10,000 millisecond-level pseudo samples.

[0055] The purpose of this step is to replace the computationally complex time interpolation structure with the original asynchronous sampling structure, thereby achieving low memory storage and high response processing of heterogeneous data in edge devices.

[0056] Furthermore, the edge node equipment is a photovoltaic module equipment, and the low-frequency environmental physical quantities include irradiance data and temperature data, while the high-frequency electrical physical quantities include transient current data and transient voltage data.

[0057] This embodiment provides a data item configuration method for edge nodes of photovoltaic modules; specifically, in the same photovoltaic power station scenario, irradiance data and temperature data are selected for low-frequency environmental physical quantities, and transient current data and transient voltage data are selected for high-frequency electrical physical quantities.

[0058] Specifically, irradiance and temperature were chosen as environmental observations because their rates of change are lower than the preset environmental rate of change threshold, yet they can form external constraints on the reasonable fluctuation boundaries of current and voltage. Transient current and transient voltage were chosen as state sampling quantities because they are more sensitive to faults such as poor contact, microcrack propagation, sampling front-end distortion, and loop jitter. Through this configuration, the system forms a diagnostic structure that constrains fast variables with slow variables.

[0059] Furthermore, in practical implementation, irradiance can be output by an irradiance sensor installed on the array side, and temperature can be output by a backplane temperature probe or an ambient temperature probe; transient current and transient voltage can be directly obtained by the component cascade sampling board; the system does not need to exhaust all environmental quantities or all electrical quantities, as long as the selected low-frequency quantity can characterize the external boundary conditions and the selected high-frequency quantity can characterize the internal state fluctuations, it can support subsequent gating and covariance analysis.

[0060] In one specific embodiment, assuming that the irradiance drops from 700 to 520 at a certain moment while the temperature remains at 35 degrees Celsius, the environmental change is mainly driven by the irradiance. If a decrease in current and voltage fluctuations are observed simultaneously within the preset normal allowable fluctuation range, it can be considered a candidate for a normal disturbance caused by cloud cover. Conversely, if the irradiance and temperature remain basically unchanged, but irregular spikes appear in the current and voltage at the same time, it is more likely to be an internal sampling link or connection status problem.

[0061] As an anomaly handling mechanism, if the irradiance sensor fails, temperature data can be used temporarily for low-frequency constraints, but the confidence level will be reduced and an environmental dimension degradation flag will be added to the label. If the temperature data is missing but the irradiance is complete, gating can still be performed. If only one high-frequency quantity is available among the current and voltage, the single-channel state change can still be calculated, but the cross-domain matrix dimension will be reduced, and the corresponding dimension reduction model will be used for subsequent manifold comparison.

[0062] In continuous monitoring of the same target component, the edge node collects a set of irradiance and temperature every 10 seconds, and a set of current and voltage every 1 millisecond. When clouds appear in the afternoon, the irradiance drops rapidly, the current drops synchronously, and the voltage rises slowly. The system regards this as a state change dominated by the environment. When maintenance personnel accidentally touch the wiring terminal in the evening, the irradiance and temperature remain almost unchanged, but the current and voltage simultaneously show irregular jumps. The system then switches to the fault identification process.

[0063] The purpose of this step is to construct environmental boundaries using irradiance and temperature, and to construct equipment status characteristics using current and voltage, thereby establishing a data acquisition foundation with clear causal meaning at the edge.

[0064] Furthermore, the derivative-gated evaluation module calculates the environmental change rate of low-frequency environmental observation data, specifically for: extracting the time series of low-frequency environmental observation data; performing approximate differentiation processing on the time series using a first-order difference algorithm to obtain the difference result; and using the difference result as the environmental change rate.

[0065] This embodiment provides a step for calculating the rate of environmental change. Specifically, in the above-mentioned continuous monitoring scenario, the derivative gating evaluation module forms a time series of low-frequency environmental observation data according to the arrival order, and then uses the first-order difference approximation to obtain the rate of environmental change.

[0066] Specifically, although the previous solution has already used environmental data to gating the determination of high-frequency faults, if it only compares whether the current environmental value exceeds a certain absolute threshold, it is easy to generate false triggers under different baselines such as early morning, noon, and evening. For example, an absolute value of irradiance higher than the normal range baseline is not equivalent to a sudden change in the environment, and a differential change rate needs to be introduced as a determination condition. Therefore, this embodiment introduces a first-order differential mechanism to describe the environmental change rate by the difference between adjacent sampling points.

[0067] For equally spaced sampling, the current value can be directly subtracted from the previous value. For non-equally spaced sampling, normalization can be performed by combining the sampling interval. For example, if the irradiance values ​​of an environmental sequence at three low-frequency sampling times are 500, 505, and 560, and the sampling interval is 10 seconds, the two difference results are 5 and 55, corresponding to a stable first segment and a rapidly changing second segment. If the stability threshold is 20, the system will trigger the rapidly changing environmental branch in the second segment. If a temperature sequence is also introduced, the temperature difference can be calculated separately, and then the dominant rate of change can be selected according to preset rules, or the weighted sum of multiple differences can be used as the total environmental rate of change.

[0068] As an exception handling mechanism, if the time series has just started and there is no previous sampling point, the current point is only initialized and the rate of change is not output; if the time interval between two consecutive sampling points is greater than the preset maximum sampling interval threshold, the difference can be marked as a low confidence metric and not used for formal gating; if the difference result is fluctuating around the threshold, a hysteresis region can be set, for example, entering a branch with drastic changes requires a value greater than 20, while exiting requires a value lower than 15, in order to reduce state jitter.

[0069] During the daily operation of the same target component, the irradiance at 11:00:00, 11:00:10, and 11:00:20 is 780, 784, and 620, respectively. The system performs a first-order difference on the adjacent samples, obtaining 4 and -164, respectively. Since the absolute value of the second difference, 164, is much greater than the threshold, the edge node determines that the environment is changing rapidly and then sends a gating signal to the high-frequency domain, instead of continuing to use the instantaneous spike judgment logic under stable weather conditions.

[0070] The purpose of this step is to characterize environmental dynamics using the rate of change rather than absolute values, so that the gating conditions are closer to the real disturbance process, thereby enabling rapid identification of environmental abrupt changes.

[0071] Furthermore, the feature mask control module generates a feature mask signal and injects it into the high-frequency state sampling domain. Specifically, it is used to: trigger an internal hardware interrupt signal if the rate of change of the environment is greater than the stability threshold; use the internal hardware interrupt signal as the feature mask signal; and change the direct memory access transmission target address corresponding to the high-frequency state sampling data through the feature mask signal to complete the injection of the feature mask signal into the high-frequency state sampling domain.

[0072] This embodiment provides a feature mask injection mechanism; specifically, after the rate of environmental change has been determined to exceed the stability threshold, the system directly applies the determination result to the high-frequency sampling link through an internal hardware interrupt, and completes the mask injection by changing the transmission target address of direct memory access.

[0073] Specifically, while setting an environmental anomaly flag at the software level can express state changes, the response speed is slow and may miss critical time slices with high-frequency sampling tasks. In high-frequency sampling scenarios of edge devices, if the polling method is still used to update the judgment path, when the cloud layer quickly blocks the view and the high-frequency peak arrives, the conventional fault logic may be executed first and generate false alarms. Therefore, this embodiment further introduces a hardware interrupt-driven mask injection method, so that the low-frequency gating result can intervene in the high-frequency data processing chain with a shorter delay.

[0074] The internal hardware interrupt signal is a high-priority level transition signal triggered by the microcontroller's external interrupt pin or timer. This interrupt signal takes effect immediately in the next microcontroller machine cycle after the derivative gating evaluation logic determines that the rate of change of the environment has exceeded the limit, so as to ensure zero-latency forced intervention of the direct memory access controller.

[0075] Specifically, when the rate of change of the output environment of the derivative gating evaluation module is greater than the stationary threshold, the controller triggers an internal hardware interrupt. This interrupt sets the mask validity flag on the one hand, and switches the direct memory access target address of the high-frequency state sampling data on the other hand. For example, the high-frequency data was originally written to the regular diagnostic buffer B0. After the mask is triggered, the direct memory access transport target is rewritten to the accompanying analysis buffer B1. Since B1 is decoupled from the regular fault logic, during the mask validity period, the original alarm branch based on the instantaneous threshold does not directly make a fault conclusion on this batch of high-frequency data, but waits for the subsequent cross-domain covariance determination.

[0076] In one specific embodiment, assuming that the original configuration of direct memory access was from the analog-to-digital conversion result register to the buffer starting at address 0x1000, once the mask interrupt is triggered, the system rewrites it to the buffer starting at address 0x2000 from the analog-to-digital conversion result register; thus, high-frequency samples within the next few milliseconds no longer enter the regular alarm pipeline, but instead enter the accompanying analysis dedicated area; after the accompanying window ends, the direct memory access address is restored to the regular area; through this address-level switching, the mask signal does not need to be attached to the data content sample by sample, reducing the additional marking overhead of the high-frequency stream;

[0077] As an exception handling mechanism, if the direct memory access is at the end of a single transfer when the interrupt is triggered, the system can switch the address after the current transfer is completed to avoid partial packet writing; if the analysis buffer is full, the oldest environmental disturbance candidate segment is discarded first, while the newly triggered segment is retained, and a buffer squeeze event is registered; if the interrupt is triggered frequently, a minimum reentry interval can be set to prevent the mask state from bouncing back and forth.

[0078] Under the same photovoltaic array in the afternoon with rapid cloudy weather, the rate of change of irradiance exceeded the threshold twice in a row. After the first interruption was triggered by the edge node, the high-frequency current and voltage flow was switched from the regular buffer to the accompanying analysis buffer, and the original instant alarm logic was suspended. When the second environmental event arrived in a short time, the system determined that it was still within the same accompanying window. Therefore, instead of creating a new regular alarm, it continued to write the subsequent high-frequency segments into the accompanying analysis buffer until the environment returned to stability.

[0079] The purpose of this mechanism is to inject the low-frequency event of environmental change into the high-frequency processing channel in the form of hardware-level control signals, thereby achieving fault judgment suspension and data diversion with microsecond-level response.

[0080] Furthermore, the cross-domain covariance calculation module performs integral operations on the high-frequency state sampling data within the accompanying time window to obtain high-frequency integral values, and calculates the state change of the high-frequency integral values. Specifically, it is used to: perform moving average filtering on the high-frequency state sampling data to obtain denoised high-frequency data; perform integral operations on the denoised high-frequency data within the time range of the accompanying time window to obtain high-frequency integral values; and calculate the difference between the high-frequency integral values ​​and the preset historical stationary integral values, using the difference as the state change.

[0081] This embodiment provides a high-frequency state change extraction step; specifically, after the accompanying time window has been started, the system first performs a moving average filter on the high-frequency state sampling data, and then integrates the denoised high-frequency data, using the difference between the integration result and the historical stationary integral value as the state change.

[0082] Specifically, using only high-frequency instantaneous peak values ​​as state representation, while fast, is susceptible to thermal noise, sampling jitter, and single-point spikes during periods of drastic environmental change, leading to unstable features. Therefore, this embodiment introduces a filtering + integration processing chain to convert the high-frequency local sequence into a window-level cumulative response, enabling more robust correlation analysis with low-frequency environmental changes. For example, within a 10-millisecond accompanying window, the original high-frequency current samples are 8.0, 8.5, 7.9, 8.1, and 8.2. Directly using the second value of 8.5 may be affected by single-point spikes. The system first uses a moving average of length 3 to obtain an approximate denoised sequence, such as 8.13, 8.17, and 8.07.

[0083] Summing the sequence or integrating over the sampling interval yields a high-frequency integral value, such as 24.37. If the integral reference value for the corresponding window under historical stationary conditions is 24.00, then the state change is 0.37. This change reflects the overall state change within the window better than a single-point peak value.

[0084] To further explain the above calculation logic, under the discrete sampling architecture of the edge node microcontroller, the integration operation is essentially a discrete numerical integration. Specifically, the system uses the original sampling time interval of the high-frequency state sampling data as a fixed integration step size, and multiplies each denoised high-frequency data value within the accompanying time window with the integration step size to obtain the state area micro-element within a single sampling period.

[0085] The process involves linearly summing all the state area micro-elements generated within the accompanying time window to output a high-frequency integral value. This process avoids complex continuous domain calculus function calls and efficiently achieves energy dimension feature extraction in edge nodes with limited computing power.

[0086] Furthermore, historical stability integral values ​​can be derived from window statistics of multiple confirmed normal periods in the past, such as taking the mean, median, or establishing reference values ​​for the same time period and the same temperature range; this can avoid mixing normal operating conditions in the early morning and at noon into the same baseline.

[0087] As an anomaly handling mechanism, if the number of high-frequency samples within the accompanying window is insufficient to complete the moving average, it can degenerate into direct integration, but the confidence level of this judgment needs to be reduced; if historical stationary integral values ​​have not yet been established, only the current integral result is retained for subsequent learning, and the formal state change is not immediately formed; if obvious saturation or pruning signs appear after filtering, it indicates that the sampling chain may be distorted, and the system can generate sampling chain anomaly markers in advance.

[0088] On the same target component, after cloud obstruction occurs at 13:20, the system extracts the current segment of the next 5 milliseconds and performs a moving average to obtain a set of denoised samples, which are then accumulated to form an integral value.

[0089] Compared with the historical integral reference value of the same window when the component was running stably in the morning, only the value decreased within the preset tolerance range. Therefore, it was subsequently included in the cross-domain analysis along with the environmental change value. At 15:05, the external environment was stable, but the difference between the window integral value and the historical stable reference value was greater than the preset deviation from the baseline threshold. The system marked the change value of this state as an abnormal candidate for subsequent diagnosis.

[0090] The purpose of this step is to transform high-frequency local fluctuations from single-point spikes into window-level cumulative features, thereby improving the stability of state representation in noisy environments.

[0091] Furthermore, the cross-domain covariance calculation module combines low-frequency environmental observation data to calculate the cross-domain covariance matrix, specifically used for: calculating the fluctuation range of low-frequency environmental observation data within the accompanying time window, and using the fluctuation range as a low-frequency change quantity; extracting the state change quantity and the low-frequency change quantity to construct a cross-domain feature vector; and calculating the cross-domain covariance matrix based on the cross-domain feature vector.

[0092] This embodiment provides a step for constructing a cross-domain covariance matrix. Specifically, after obtaining the high-frequency state change, the system further calculates the fluctuation range of the low-frequency environmental observation data within the accompanying time window, uses it as the low-frequency change, and together with the state change, forms a cross-domain feature vector, and then calculates the cross-domain covariance matrix accordingly.

[0093] Specifically, although the previous step has obtained the high-frequency state change quantity, judging solely based on the magnitude of this change quantity still cannot answer the core question of whether the change is consistent with the environmental change. Therefore, this embodiment introduces the low-frequency change quantity and uses cross-domain covariance to characterize the joint change relationship between the two.

[0094] Specifically, within the accompanying time window, the system calculates the fluctuation range of low-frequency environmental data, which is the difference between the maximum and minimum values ​​within the window. The fluctuation range is suitable for implementation on edge devices because it is simple to calculate and does not require complex fitting. After obtaining the low-frequency change, it is combined with the high-frequency state change of the corresponding window to form a cross-domain feature vector. If the system considers both irradiance and temperature, a three-dimensional vector can be formed, such as [state change, irradiance range, temperature range]. If only one dominant environmental quantity is used, a two-dimensional vector can be formed, such as [state change, environmental range].

[0095] Since the cross-domain feature vector at a single moment cannot be used to directly solve for a meaningful covariance matrix, when the cross-domain covariance matrix is ​​calculated based on the cross-domain feature vector, the system pushes the cross-domain feature vector obtained in the current time window into a short-term observation queue pre-established in the edge node. This short-term observation queue stores the most recent preset number of cross-domain feature vectors, such as the most recent 3.

[0096] The system uses the set of feature vectors in the short-term observation queue as the sample space, calculates the covariance between each dimension, and then obtains the cross-domain covariance matrix that reflects the correlation between current and recent short-term fluctuations. The specific solution logic of the cross-domain covariance matrix is ​​further clarified: for multiple cross-domain feature vectors stored in the short-term observation queue, the system calculates the unbiased estimate of the product of the co-bias of any two feature dimensions in all queue samples, and directly uses it as the off-diagonal element value of the corresponding row and column position in the cross-domain covariance matrix.

[0097] Meanwhile, the system calculates the unbiased average of the squared deviations of each feature dimension from its mean, and uses it as the variance element value on the main diagonal of the covariance matrix. Through this standardized covariance business flow rule, the system can directly quantify the consistency of coordinated changes between high-frequency state evolution and low-frequency environment evolution using deterministic arithmetic operations without the need for implicit nonlinear model fitting.

[0098] In one specific embodiment, assuming that the system's short-term observation queue already contains two historically retained cross-domain feature vectors V1=[0.3, 20] and V2=[0.5, 35], when the current accompanying time window calculates V3=[0.4, 28], V3 is pushed into the queue and used as the current main feature vector;

[0099] The system calculates the covariance matrix based on the sample space composed of these three vectors. If the matrix obtained by using the sample covariance calculation in the two-dimensional case is approximately [[0.01, 0.75], [0.75, 56.33]], then the off-diagonal term 0.75 reflects the positive correlation between recent state changes and environmental changes. If V4=[-1.0, 40] is obtained in a subsequent window and updated into the queue, it may show a deviation from the preset consistency benchmark with the existing correlation pattern, resulting in the matrix element change rate being greater than the preset update threshold.

[0100] As an anomaly handling mechanism, if there is only one low-frequency data point in the accompanying window and a range cannot be formed, the range can be set to 0 and the window's environmental resolution can be marked as insufficient. If one of the multiple environmental quantities is missing, a dimension-reduced vector is constructed using the number of available quantities, and the covariance model of the corresponding dimension is called. If both the state change quantity and the low-frequency change quantity are close to 0, the window is considered a low-information window and can be left untouched for formal diagnosis, but only used for historical stationary sample accumulation.

[0101] During the operation of the same power station, from 13:20 to 13:20:05, the system obtained a current integral difference of 0.4, an irradiance range of 210, and a temperature range of 0.5. Therefore, a three-dimensional vector was constructed and put into the short-time queue for calculation. When the environment was stable but the sampling end was abnormal at 13:47, the system obtained a current integral difference of 1.6, an irradiance range of 5, and a temperature range of 0.1. After this vector was pushed in, it will form a significantly different covariance relationship with the previous environmental disturbance window, which will be used for subsequent manifold comparison and identification.

[0102] The purpose of this step is to transform the observation of high-frequency and low-frequency changes from independent observations to joint observations, thereby achieving a quantitative expression of causal consistency and anomalous deviations.

[0103] Furthermore, the fault identification output module compares the cross-domain covariance matrix with a preset healthy baseline manifold. Specifically, it is used to: extract the eigenvectors of the cross-domain covariance matrix; calculate the Mahalanobis distance between the eigenvectors and the healthy baseline manifold; compare the Mahalanobis distance with a preset manifold boundary threshold; if the Mahalanobis distance is greater than the manifold boundary threshold, it is determined that the cross-domain covariance matrix deviates from the healthy baseline manifold; if the Mahalanobis distance is less than or equal to the manifold boundary threshold, it is determined that the cross-domain covariance matrix does not deviate from the healthy baseline manifold.

[0104] This embodiment provides a fault identification step based on a healthy baseline manifold; specifically, the system extracts feature vectors from the cross-domain covariance matrix, calculates the Mahalanobis distance between the feature vectors and the distribution of the healthy baseline manifold, and outputs fault alarms or normal environmental disturbance labels based on the manifold boundary threshold.

[0105] Specifically, the cross-domain covariance matrix has been obtained in the previous step, but if we only observe whether the matrix elements increase or decrease, there is a lack of a unified standard for normal boundaries. Therefore, this embodiment further introduces a healthy baseline manifold to represent the joint distribution area formed by high-frequency state changes and low-frequency environmental changes when the target component is operating normally for a long time. The reason for using Mahalanobis distance is that it not only considers the location of the feature center, but also the correlation between the dimensions, which is suitable for handling covariance-related features.

[0106] In practice, several representative quantities can be extracted from the cross-domain covariance matrix as eigenvectors, such as the diagonal elements, off-diagonal elements, and their combinations. Assuming a two-dimensional case, a matrix can be simplified to F=[0.01, 0.45, 38.0]. The healthy baseline manifold can be estimated from such a set of eigenvectors of historical normal samples, with its center being, for example, μ=[0.02, 0.50, 37.0]. The distribution covariance is formed by statistics from normal samples.

[0107] Specifically, the healthy baseline manifold is constructed based on a multivariate Gaussian distribution model. By performing maximum likelihood estimation on the cross-domain feature vector set within the historical normal operation interval, the feature mean vector of the set is calculated as the center of the baseline manifold, and the covariance matrix of the set is calculated as the distribution covariance, thereby defining the probability density closed envelope representing the normal environmental disturbance conditions in the multidimensional feature space.

[0108] The system calculates the Mahalanobis distance from the current feature vector to the distribution. If the distance does not exceed a threshold, such as 3.0, the current relationship is determined to still be within the normal environmental disturbance range. If the distance exceeds the threshold, the current cross-domain relationship is determined to have deviated from the healthy mode.

[0109] In addition, when calculating Mahalanobis distance, if the covariance matrix of the healthy baseline manifold distribution becomes ill-conditioned or singular due to the linear correlation coefficient between feature dimensions being greater than the preset correlation threshold or the variance of normal samples being lower than the preset lower limit of variance, i.e., it is irreversible, the system will add a non-zero constant preset regularization factor, such as 1e-5, to the main diagonal elements of the distribution's covariance matrix to ensure that the matrix is ​​full rank and complete the inversion calculation, thereby avoiding the system crash caused by the division by zero under the limited computing power of edge nodes;

[0110] Furthermore, in this embodiment, the extraction of the eigenvector of the cross-domain covariance matrix specifically refers to the extraction of representative elements from the matrix data structure, such as extracting all diagonal elements and off-diagonal elements on a specific side, and concatenating them in a fixed order to form a one-dimensional attribute representation vector, which is used to mark the feature space position of the current cross-domain matrix in a multi-dimensional coordinate system; rather than referring to the matrix eigenvectors that satisfy the characteristic equation in the field of linear algebra.

[0111] By directly extracting and flattening matrix elements, the constructed eigenvectors can not only fully retain the variance and covariance information of cross-domain fluctuations, but also avoid the problems of dense iterative computing power consumption and floating-point truncation error that edge devices may face when solving matrix eigenvalues ​​and eigenvectors.

[0112] In one specific embodiment, it is assumed that under normal circumstances, when the environment increases, the state change usually increases synchronously, so most of the off-diagonal terms are positive; if the matrix features obtained by the current window show that the environment changes drastically while the state change is reversed abnormally or the environment remains almost unchanged while the state change is very drastic, then the Mahalanobis distance will increase significantly, thus falling outside the healthy baseline manifold.

[0113] As an anomaly handling mechanism, if the number of healthy baseline manifold samples is insufficient, such as in the early stages of system deployment, the Mahalanobis distance result is only used as a reference and does not directly trigger a strong alarm, but instead outputs an event pending confirmation; if the feature vector has a missing dimension, a reduced-order manifold model with the same dimension is called; if the Mahalanobis distance is exactly equal to the boundary threshold, it is treated as not deviating to reduce boundary sensitivity; if multiple consecutive windows exceed the threshold, the alarm level can be upgraded; if only a single window exceeds the boundary and recovers quickly afterward, a light alarm can be output first and a review mark can be retained.

[0114] In the monitoring of the same target component, the matrix feature vector generated by the cloud occlusion event at 13:20 is close to the center of the normal manifold, and the Mahalanobis distance is only 1.2. Therefore, the system outputs a normal environmental disturbance label. At 15:05, the external environment is stable, but the cross-domain feature vector shows that the change in state and the change in environment no longer maintain the historical consistency. The Mahalanobis distance reaches 5.8, which exceeds the boundary threshold. Based on this, the system generates a system fault alarm command and uploads it to the operation and maintenance platform.

[0115] The purpose of this step is to establish quantifiable normal boundaries for cross-domain relationships, thereby enabling stable differentiation between environmental disturbances and real faults.

[0116] Furthermore, the fault identification output module monitors the instantaneous fluctuation characteristics of high-frequency state sampling data and determines whether there is a data mutation. Specifically, it is used to: calculate the instantaneous volatility of high-frequency state sampling data; compare the instantaneous volatility with a preset high-frequency mutation threshold; if the instantaneous volatility is greater than the high-frequency mutation threshold, it is determined that there is a data mutation in the high-frequency state sampling data; if the instantaneous volatility is less than or equal to the high-frequency mutation threshold, it is determined that there is no data mutation in the high-frequency state sampling data.

[0117] This embodiment provides a high-frequency mutation detection step under a stable environment; specifically, when the system determines that the rate of environmental change does not exceed the stability threshold, the fault identification output module directly calculates the instantaneous fluctuation rate of the high-frequency state sampling data and compares it with the high-frequency mutation threshold to determine whether there is a data mutation.

[0118] Specifically, the aforementioned environmental gating mechanism is mainly used to handle periods of active external disturbances; however, the environment is stable during most daily periods. If the accompanying window and manifold comparison process is still forced at this time, it will increase unnecessary overhead. Therefore, this embodiment adds a fast channel for stable environments: when the environment is stable, as long as there is an abnormal change in the high-frequency state sampling data, it can be given priority as a candidate for suspected hardware failure.

[0119] Instantaneous volatility can be obtained from the difference between adjacent high-frequency samples, the absolute value of the sample difference, or the ratio of the maximum and minimum difference within a short window to the mean. For example, if a current sequence is 8.00, 8.02, 8.01, and 8.95 at four consecutive millisecond points, the difference between the third and fourth points is 0.94. If the high-frequency mutation threshold is set to 0.30, then this change significantly exceeds the threshold, and a data mutation can be identified. If another sequence is 8.00, 8.03, 8.01, and 8.04, then the maximum adjacent difference is only 0.03, which does not constitute a mutation.

[0120] As an anomaly handling mechanism, if the high-frequency sampling value is at the edge of the sensor's range, saturation shear may occur. In this case, the range status should be checked first before deciding whether to treat it as a valid mutation. If the high-frequency channel is briefly affected by electromagnetic interference, causing a single-point glitch, a minimum number of continuous samples can be set, for example, a suspected fault can only be output after two or three consecutive samples exceed the threshold, in order to suppress false alarms of isolated glitch. If the instantaneous volatility is exactly equal to the threshold, it should be treated as if there is no mutation.

[0121] During the operation of the same target component in a clear and stable period, both irradiance and temperature difference remained within a stable range. At 14:12, the high-frequency voltage sequence showed multiple millisecond-level transitions in succession, while the environmental data did not change synchronously. The system directly calculated the instantaneous fluctuation rate and compared it with the mutation threshold. Once the threshold was exceeded, a suspected hardware failure interruption signal was generated, prompting maintenance personnel to prioritize checking the sampling board solder joints, connectors, and front-end sensor status.

[0122] The purpose of this step is to provide a low-latency anomaly detection path under stable environmental conditions, thereby enabling a rapid response to real hardware failures.

[0123] Furthermore, this system also includes a baseline manifold update module, which is used to: obtain the cross-domain covariance matrix corresponding to all environmental disturbance normal labels generated during the historical normal operation cycle; form a historical covariance set from the cross-domain covariance matrices corresponding to all environmental disturbance normal labels; recalculate the baseline manifold parameters based on the historical covariance set; and adaptively update the healthy baseline manifold according to the baseline manifold parameters.

[0124] This embodiment provides an adaptive update mechanism for a healthy baseline manifold. Specifically, the system collects cross-domain covariance matrices that are identified as having normal environmental disturbance labels during historical normal operation cycles, forms a historical covariance set, and recalculates the baseline manifold parameters based on this set to complete the adaptive update of the healthy baseline manifold.

[0125] Specifically, if the health baseline manifold remains fixed for a long period of time, the original normal boundary may gradually become disconnected from the actual operating state due to seasonal changes, component aging, sensor calibration drift, and long-term migration of different lighting conditions, leading to an increase in false alarms. Therefore, this embodiment adds an adaptive update module on top of the aforementioned static discrimination framework, so that the normal boundary can be gradually adjusted with long-term stable operating data.

[0126] In practice, the system only absorbs window data that has been identified as normal environmental disturbances, and does not absorb data that has been alarmed or is pending confirmation, thus avoiding contaminating the normal model with abnormal samples; the historical covariance set can be maintained with a fixed number of samples, a fixed time span, or a rolling window method; for example, the matrix of the most recent 500 normal samples can be retained, and after extracting their feature vectors, the center and distribution parameters can be recalculated to obtain a new healthy baseline manifold; if some samples are too close to the boundary of the current manifold, a lower weight can be set to reduce the risk of the boundary drifting too quickly;

[0127] In one specific embodiment, assuming that four normal matrix eigenvectors, F1, F2, F3, and F4, are accumulated at a certain stage, the system calculates their mean as the new manifold center. If the distance between the new center and the old center after the update is less than the preset center drift tolerance threshold, it indicates that the long-term state of the equipment is stable and the manifold is only slightly modified. If the difference gradually shifts but no discrete burst occurs, it indicates that the overall normal operating condition changes slowly, and the system allows the boundary to be updated according to the preset smooth evolution growth. Conversely, if a batch of candidate samples causes the rate of change of distribution dispersion to be greater than the preset evolution tolerance threshold, it will not be updated immediately, but the batch of samples will be put aside to wait for more confirmation data.

[0128] As an anomaly handling mechanism, if there are insufficient normal environmental disturbance labels available for updating, the old manifold remains unchanged; if no stable normal labels are generated for several consecutive days, the characterization system may be in an abnormal state for a long time or the data may not meet the preset signal-to-noise ratio requirements. At this time, the update is paused and the manifold freeze status is reported to the host computer; if the change between the old and new baseline parameters exceeds the preset safety range, a step-by-step transition method is adopted for updating instead of a one-time replacement to avoid abrupt changes in the diagnostic boundary.

[0129] After the same photovoltaic power plant enters summer, the afternoon cloud shadow disturbance pattern is different from that in spring, but the target components are still operating normally. The edge nodes gradually incorporate these windows that have been repeatedly confirmed as environmental disturbances into the historical covariance set, recalculate the baseline manifold parameters, and make the new normal boundary closer to the summer operating conditions. Thereafter, when a real sampling link anomaly occurs, the system can still effectively identify the deviation based on the updated healthy manifold, and will not mistake seasonal normal differences for faults.

[0130] The purpose of this mechanism is to enable the normal baseline to evolve adaptively with long-term operating conditions, thereby achieving the continuous effectiveness and long-term stability of the diagnostic boundary.

[0131] 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 photovoltaic module fault prediction and health management system based on multi-source Internet of Things data, characterized in that, Applied to edge node devices, including: A heterogeneous data acquisition module is used to acquire low-frequency environmental observation data and high-frequency state sampling data of the edge node device, wherein the sampling rate of the low-frequency environmental observation data is lower than the sampling rate of the high-frequency state sampling data. The derivative-gated evaluation module is used to calculate the environmental change rate of the low-frequency environmental observation data and compare the environmental change rate with a preset stationarity threshold. The feature mask control module is used to generate a feature mask signal and inject it into the high-frequency state sampling domain that processes the high-frequency state sampling data if the environmental change rate is greater than the stability threshold. Based on the feature mask signal, the fault judgment logic based on a preset high-frequency transient threshold is suspended, and an accompanying time window is started. The cross-domain covariance calculation module is used to perform integral operations on the high-frequency state sampling data within the accompanying time window to obtain high-frequency integral values, calculate the state change of the high-frequency integral values ​​relative to the preset historical stationary integral values, construct a cross-domain feature vector by combining the fluctuation range of the low-frequency environmental observation data within the accompanying time window, and calculate the cross-domain covariance matrix based on the cross-domain feature vector. The fault identification output module is used to compare the cross-domain covariance matrix with a preset healthy baseline manifold representing the historical normal operating state. If the cross-domain covariance matrix deviates from the healthy baseline manifold, a system fault alarm command is generated; if the cross-domain covariance matrix does not deviate from the healthy baseline manifold, an environmental disturbance normal label is generated. It is also used to monitor the instantaneous volatility of the high-frequency state sampling data if the environmental change rate is less than or equal to the stability threshold. If a data mutation is detected, a suspected hardware fault interruption signal is generated; if no data mutation is detected, an environmental disturbance normal label is output.

2. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, The heterogeneous data acquisition module acquires low-frequency environmental observation data and high-frequency state sampling data of the edge node device, specifically for: acquiring low-frequency environmental physical quantities of the edge node device and using the low-frequency environmental physical quantities as the low-frequency environmental observation data; acquiring high-frequency electrical physical quantities of the edge node device and using the high-frequency electrical physical quantities as the high-frequency state sampling data; wherein the low-frequency environmental observation data and the high-frequency state sampling data maintain their original asynchronous sampling state and are not interpolated or aligned on the time axis.

3. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 2, characterized in that, The edge node device is a photovoltaic module device, the low-frequency environmental physical quantity includes irradiance data and temperature data, and the high-frequency electrical physical quantity includes transient current data and transient voltage data.

4. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, The derivative-gated evaluation module calculates the environmental change rate of the low-frequency environmental observation data, specifically by: extracting the time series of the low-frequency environmental observation data; performing approximate differentiation on the time series using a first-order difference algorithm to obtain the difference result; and using the difference result as the environmental change rate.

5. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, The feature mask control module generates a feature mask signal and injects it into the high-frequency state sampling domain. Specifically, it is used to: trigger an internal hardware interrupt signal if the environmental change rate is greater than the stability threshold; use the internal hardware interrupt signal as the feature mask signal; and change the direct memory access transmission target address corresponding to the high-frequency state sampling data through the feature mask signal to complete the injection of the feature mask signal into the high-frequency state sampling domain.

6. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, The cross-domain covariance calculation module performs integration on the high-frequency state sampling data within the accompanying time window to obtain a high-frequency integral value, and calculates the state change of the high-frequency integral value. Specifically, it performs moving average filtering on the high-frequency state sampling data to obtain denoised high-frequency data; performs integration on the denoised high-frequency data within the time range of the accompanying time window to obtain the high-frequency integral value; calculates the difference between the high-frequency integral value and a preset historical stationary integral value, and uses the difference as the state change.

7. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, The cross-domain covariance calculation module calculates the cross-domain covariance matrix in conjunction with the low-frequency environmental observation data, specifically used to: calculate the fluctuation range of the low-frequency environmental observation data within the accompanying time window, and use the fluctuation range as a low-frequency change quantity. Extract the state change and the low-frequency change to construct a cross-domain feature vector; calculate the cross-domain covariance matrix based on the cross-domain feature vector.

8. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, The fault identification output module compares the cross-domain covariance matrix with a preset healthy baseline manifold, specifically by: extracting the eigenvectors of the cross-domain covariance matrix; calculating the Mahalanobis distance between the eigenvectors and the healthy baseline manifold; comparing the Mahalanobis distance with a preset manifold boundary threshold; if the Mahalanobis distance is greater than the manifold boundary threshold, then the cross-domain covariance matrix is ​​determined to deviate from the healthy baseline manifold; if the Mahalanobis distance is less than or equal to the manifold boundary threshold, then the cross-domain covariance matrix is ​​determined not to deviate from the healthy baseline manifold.

9. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, The fault identification output module monitors the instantaneous fluctuation characteristics of the high-frequency state sampling data and determines whether there is a data mutation. Specifically, it is used to: calculate the instantaneous volatility of the high-frequency state sampling data; compare the instantaneous volatility with a preset high-frequency mutation threshold; if the instantaneous volatility is greater than the high-frequency mutation threshold, it is determined that there is a data mutation in the high-frequency state sampling data; if the instantaneous volatility is less than or equal to the high-frequency mutation threshold, it is determined that there is no data mutation in the high-frequency state sampling data.

10. The photovoltaic module fault prediction and health management system based on multi-source Internet of Things data according to claim 1, characterized in that, It also includes a baseline manifold update module, which is used to: obtain the cross-domain covariance matrix corresponding to all environmental disturbance normal labels generated during the historical normal operation cycle; form a historical covariance set from the cross-domain covariance matrices corresponding to all environmental disturbance normal labels; recalculate the baseline manifold parameters based on the historical covariance set; and adaptively update the healthy baseline manifold according to the baseline manifold parameters.

Images