LED light-emitting disc fault early warning method and system based on current harmonic spectrum analysis

Abstract

The present application relates to the technical field of electrical equipment state monitoring, and more particularly to a LED light-emitting disc fault early warning method and system based on current harmonic spectrum analysis, which comprises collecting a driving power supply current signal and performing frequency domain transformation on the frames, intercepting a high-frequency observation window to obtain a high-frequency energy subsequence of each frame; performing robust standardization on each frame, filtering in the frequency axis local neighborhood median and squaring to obtain microscopic divergence; accumulating the divergence of each frequency point to obtain a total distortion energy scalar and smoothing filtering to obtain a trend quantity; setting a fixed baseline threshold based on the health data at the initial operation stage, dynamically correcting the threshold combined with the working condition median ratio, and alarming when the trend quantity continuously exceeds the threshold for a preset number of times; the present application improves the timeliness and accuracy of LED light-emitting disc driving power supply fault early warning.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment condition monitoring technology. More specifically, this invention relates to a method and system for early warning of LED light-emitting panels based on current harmonic spectrum analysis. Background Technology

[0002] LED LED displays, with their high luminous efficacy, long lifespan, and fast response, have widely replaced traditional incandescent and halogen lamps, becoming the core light-emitting components for railway signal lights, industrial lighting fixtures, and large outdoor displays. In practical engineering applications, LED displays typically use high-frequency switching power supplies as the driving source, employing pulse width modulation (PWM) technology to achieve constant current control and ensure stable brightness. Under long-term high-frequency switching conditions, the core semiconductor switching devices within the driver power supply experience irreversible lattice defect accumulation and conduction characteristic degradation in their PN junction regions due to continuous electrothermal stress. Early signs of this degradation include a slow increase in on-resistance, a gradual extension of reverse recovery time, and increased switching edge jitter. These microscopic physical changes modulate the current waveform output by the driver power supply, causing subtle distortions in its high-frequency harmonic components. As degradation intensifies, it eventually leads to significant output current distortion, abnormal brightness of the LED display, and even complete driver power supply failure, causing signal light malfunctions and seriously endangering railway safety or production continuity.

[0003] Currently, fault warning technologies for LED light-emitting disk driver power supplies mainly include the following types of solutions. The first type is the optical parameter detection solution, which uses photoelectric sensors to monitor the luminous flux attenuation or color temperature shift of the light-emitting disk. An alarm is triggered when the light output drops below a preset threshold. This solution can directly reflect the final output state of the light-emitting disk. However, optical sensors are easily affected by ambient light pollution and dust obstruction, and anomalies can only be detected after significant light decay has already occurred in the light-emitting disk. It cannot proactively predict the gradual failure process of internal components in the driver power supply. The second type is the electrical overcurrent / overvoltage protection solution, which sets voltage and current comparators at the output or input of the driver power supply. When the electrical quantity exceeds a preset safety threshold, the power supply is immediately cut off or an alarm is issued. This solution has a fast response speed, but it is essentially a reactive protection against serious faults that have already occurred, and it lacks the ability to identify weak electrical signs that appear during the slow degradation of components. The third type is the total harmonic distortion (THD) monitoring scheme, which evaluates power quality by calculating the total harmonic distortion rate of the current or voltage waveform. When the THD exceeds the set value, it is judged as abnormal. However, the THD index mainly reflects the energy ratio of the low frequency band, while the high frequency distortion caused by PN junction degradation is mainly concentrated near the switching frequency and its harmonics. The energy in this frequency band is much lower than the fundamental frequency and is submerged in the THD calculation, making it difficult to effectively characterize the degree of degradation.

[0004] In practical applications such as railway signal lights, tunnel lighting, and outdoor displays, LED light panels operate under complex power supply environments characterized by frequent grid fluctuations and significant changes in load current. The energy of the fundamental and low-order harmonics of the power grid is several orders of magnitude higher than the high-frequency components near the switching frequency, causing the weak high-frequency characteristics generated in the early stages of PN junction degradation to be completely masked in the spectrum. Random fluctuations in supply voltage and changes in the amplitude of load current cause non-stationary drift in the high-frequency noise floor, and fixed thresholds cannot adapt to this dynamic change, easily leading to false alarms. As the equipment operates for longer periods, the high-frequency distortion characteristics caused by PN junction degradation exhibit a slow, long-term upward trend. If a historical sliding time window is used to dynamically update the threshold, the threshold will drift upward in sync with this aging trend, resulting in the inability to trigger alarms even when the degradation is already quite severe, leading to missed alarms. Summary of the Invention

[0005] To address the problems of existing LED light-emitting disk fault early warning technologies being unable to reliably extract weak high-frequency fault features from strong interference environments in the very early stages of PN junction degradation, and lacking an adaptive threshold mechanism to cope with operating condition fluctuations and threshold drift, resulting in high rates of missed and false alarms, this invention provides an LED light-emitting disk fault early warning method and system based on current harmonic spectrum analysis.

[0006] In a first aspect, the present invention provides a fault early warning method for LED light-emitting panels based on current harmonic spectrum analysis, employing the following technical solution: The current signal of the LED driver power supply is collected, and the current signal is subjected to time-domain frame segmentation and frequency-domain transformation. A high-frequency observation window is extracted to obtain the high-frequency energy subsequence of each frame. For each frame, the high-frequency energy subsequence is subjected to intra-frame robust normalization, and then local neighborhood median filtering is performed on each frequency point on the frequency axis and the square value is calculated to obtain the microstructure divergence of each frequency point. The microstructure divergence of all frequency points in each frame is accumulated to obtain the total distortion energy scalar, and the total distortion energy scalar is smoothed and filtered to obtain the smoothing trend quantity. A fixed baseline threshold is set based on the smoothing trend amount under the initial healthy state of the equipment during commissioning, and the fixed baseline threshold is dynamically corrected according to the current operating conditions to obtain a dynamic threshold. When the smoothing trend amount exceeds the dynamic threshold for a preset number of consecutive times, a fault is determined to have occurred and an early warning is issued.

[0007] Preferably, the step of performing time-domain frame truncation and frequency-domain transformation on the current signal, and extracting a high-frequency observation window, includes: The current signal is segmented into frames according to a preset time window length and sliding step size to obtain a time-domain discrete sequence for each frame. A window function is applied to the time-domain discrete sequence of each frame, and then a discrete Fourier transform is performed to obtain the original energy amplitude spectrum of each frame. The range from a preset start frequency to a preset end frequency in the original energy amplitude spectrum is extracted as the high-frequency observation window to obtain the high-frequency energy subsequence of each frame.

[0008] Preferably, the intra-frame robust normalization of the high-frequency energy sub-sequence for each frame includes: Calculate the global median and the absolute deviation of the global median for the high-frequency energy subsequence in the current frame; The standardization is achieved by subtracting the global median from the energy amplitude of each frequency point in the high-frequency energy subsequence of the current frame, and then dividing by the absolute deviation of the global median.

[0009] Preferably, the step of performing local neighborhood median filtering and calculating the square value for each frequency point on the frequency axis includes: For each effective frequency point, a predetermined number of adjacent frequency points centered on it are selected to form a local neighborhood; the median of the standardized residuals of all frequency points in the local neighborhood is calculated as a local reference; the standardized residual of the center frequency point is subtracted from the local reference and the squared value is obtained to obtain the microstructure divergence of the center frequency point.

[0010] Preferably, the step of accumulating the microstructure divergence of all frequency points within each frame includes: The total distortion energy scalar is obtained by summing the microstructure divergence of all effective frequency points within the high-frequency observation window in each frame.

[0011] Preferably, the smoothing filter applied to the total distortion energy scalar includes: The total distortion energy scalar is smoothed using an exponentially weighted moving average filter; The smoothing trend amount of the current frame is the weighted sum of the total distortion energy scalar of the current frame and the smoothing trend amount of the previous frame, wherein the weight of the total distortion energy scalar of the current frame is less than the weight of the smoothing trend amount of the previous frame.

[0012] Preferably, setting a fixed baseline threshold based on the smoothing trend amount during the initial healthy state of the equipment includes: The smoothed trend values ​​are obtained from multiple consecutive time windows during the initial stage of equipment commissioning, forming a historical trend value sequence; Calculate the lower quartile and upper quartile of the historical trend series; The fixed baseline threshold is obtained by multiplying the difference between the upper quartile and the lower quartile by a preset constant and then adding the upper quartile.

[0013] Preferably, the step of dynamically correcting the fixed baseline threshold according to the current operating conditions includes: Calculate the median of the smoothing trend amount for multiple consecutive time windows preceding the current frame, and use it as the real-time median; Calculate the median of the historical trend series as the baseline median; The ratio of the real-time median to the baseline median is used as the operating condition scaling factor; The dynamic threshold is obtained by multiplying the fixed baseline threshold by the larger value between the operating condition scaling factor and 1.

[0014] Preferably, the step of determining a fault and issuing an early warning when the smoothing trend quantity continuously exceeds the dynamic threshold for a preset number of times includes: The smoothing trend value of the current frame is compared with the dynamic threshold of the current frame frame by frame; if the smoothing trend value exceeds the dynamic threshold, the anomaly count is incremented by one; if the smoothing trend value does not exceed the dynamic threshold, the anomaly count is cleared to zero; when the anomaly count reaches a preset number of times threshold, a fault is determined to have occurred and an early warning is issued.

[0015] Secondly, the present invention provides an LED light-emitting disk fault early warning system based on current harmonic spectrum analysis, which adopts the following technical solution: The LED light-emitting disk fault early warning system based on current harmonic spectrum analysis includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned LED light-emitting disk fault early warning method based on current harmonic spectrum analysis is implemented.

[0016] By adopting the above technical solution, the above-mentioned LED light-emitting disk fault early warning method based on current harmonic spectrum analysis is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0017] The present invention has the following beneficial effects: This invention performs time-domain frame truncation and frequency-domain transformation on the acquired current signal, and extracts a high-frequency observation window with the switching frequency multiples as the core. This eliminates the huge energy interference from the grid fundamental wave and low-order harmonics at the source, and precisely locks the analysis focus on the switching transient characteristic frequency band modulated by PN junction degradation. By adopting intra-frame robust normalization and utilizing the median's immunity to outliers, the influence of grid voltage fluctuations and load current changes on the overall amplitude of high-frequency energy is effectively removed. This solves the problem of non-stationary drift of the spectral noise floor caused by changes in operating conditions, allowing subsequent analysis to focus on the relative spectral shape rather than the absolute energy level.

[0018] Based on this, a local neighborhood median filter is introduced: Since sudden broadband electromagnetic interference manifests in the frequency domain as a collective unidirectional rise in energy at multiple adjacent frequency points, and the intermodulation distortion caused by PN junction degradation is essentially a relative redistribution of energy between adjacent frequencies, by performing differential operations with the neighborhood median as a local reference, the deviation of the center frequency point relative to the local background is preserved, while the collective rise component is effectively deducted. Thus, while preserving the degradation-sensitive characteristics, the contamination of the judgment result by broadband interference is accurately eliminated.

[0019] Further squaring the difference results reveals a deeper causal logic: the relative deviation between frequency points caused by degradation is usually only a few times that of normal random fluctuations under a standardized scale, making it difficult to distinguish from background noise through linear processing; squaring produces a compression effect on values ​​less than 1, bringing random noise close to zero, and an exponential amplification effect on values ​​greater than 1, significantly enhancing the effective signal. This nonlinearly highlights the small amplitude deviation caused by PN junction degradation from strong background noise, greatly improving the detectability of very early and weak fault characteristics.

[0020] By accumulating discrete microscopic distortion energies in the frequency domain, a total distortion energy scalar reflecting the overall lattice health is formed. This scalar is then combined with an exponentially weighted moving average filter with a low forgetting factor to extract slow aging trends, effectively filtering out random spikes and interference from grid surges. A fixed baseline threshold is established using initial equipment health data and remains constant throughout the equipment's lifespan. This is complemented by a dynamic correction mechanism based on the median ratio of operating conditions' energy, ensuring that the threshold only tracks the overall amplitude changes of external grid voltage and load current, rather than drifting with the equipment's own aging characteristics. A continuous threshold exceeding count criterion distinguishes between persistent degradation and single transient impacts.

[0021] The aforementioned technical features, working in tandem, collectively solve the problem of extracting high-frequency, weak fault characteristics in complex industrial environments during the very early stages of PN junction degradation. They also achieve reliable early warning throughout the entire lifecycle, with fixed thresholds that do not drift, adaptability to changes in operating conditions, and resistance to transient interference, ultimately improving the timeliness and accuracy of fault warnings for LED light-emitting disk driver power supplies. Attached Figure Description

[0022] Figure 1 This is a flowchart of the steps of the LED light-emitting disk fault early warning method based on current harmonic spectrum analysis in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] Please see Figure 1The diagram illustrates a flowchart of a fault early warning method for LED light-emitting panels based on current harmonic spectrum analysis, provided by an embodiment of the present invention. The method includes the following steps: S001. Acquire the current signal of the LED driver power supply, perform time-domain frame segmentation and frequency-domain transformation on the current signal, and extract the high-frequency observation window to obtain the high-frequency energy subsequence of each frame.

[0025] It should be noted that in the raw current signal, the fundamental frequency (50Hz / 60Hz) and low-order harmonics of the power grid typically dominate, with amplitudes several orders of magnitude higher than those near the switching frequency. Direct analysis of the entire frequency band would completely mask the weak high-frequency characteristics. Therefore, a reasonable observation time window needs to be set in the frequency domain to accurately pinpoint the high-frequency range near the switching frequency and its harmonics. Energy fluctuations within this range directly reflect the transient behavior of the switching device during turn-on and turn-off, and are closely related to the lattice state of the PN junction. By extracting the high-frequency energy, low-frequency interference is eliminated, and the amount of data for subsequent processing is significantly reduced, creating conditions for embedded real-time processing. Furthermore, the combined preprocessing strategy of time-domain sliding window and frequency-domain windowing effectively suppresses spectral leakage and the picket-fence effect, ensuring the stability of frequency resolution.

[0026] Based on this, this step specifically involves the following operations: setting a sampling rate that satisfies the Nyquist criterion to convert the analog current signal into a digital sequence; then truncating the continuous sampled data into frames, ensuring the continuity of the features on the time axis through overlapping time windows; next, applying a window function to each time window and performing a discrete Fourier transform to map the time-domain signal to the frequency domain and calculate the energy amplitude of each frequency component; finally, truncating the high-frequency observation window in the frequency domain, discarding the low-frequency fundamental wave and each harmonic, to obtain the high-frequency energy subsequence.

[0027] Specifically, a 16-bit or higher resolution analog-to-digital converter is used to acquire the current signal on the LED driver power supply input side, i.e., the AC input side or DC bus side of the switching power supply, in a synchronous sampling manner.

[0028] Let the PWM modulation frequency of the switching power supply be... The fundamental frequency of the power grid is (Typically 50Hz or 60Hz), according to the Nyquist sampling theorem, the ADC sampling rate... It should be more than twice the highest frequency of attention.

[0029] In this embodiment, the sampling rate is set to the switching frequency. 4 to 6 times, that is ,in This ensures complete capture of the switching frequency and its harmonic components near its harmonics, while avoiding data redundancy caused by excessively high sampling rates.

[0030] in, The lower limit is set to 4 to ensure that the sampling rate is at least 4 times the switching frequency, thereby satisfying the Nyquist sampling theorem and fully capturing the harmonic components near the second and third harmonics of the switching frequency. These frequency bands reflect the energy distribution of the switching transient process and are the key frequency bands for capturing the early degradation of the PN junction. The upper limit of 6 is set to avoid excessively high sampling rates leading to excessively large data volumes, which would increase the real-time processing burden on the embedded system. Furthermore, the higher harmonics of the switching frequency are already very weak and insensitive to PN junction degradation. Therefore, It is the preferred range that balances complete acquisition of key frequency bands with real-time processing efficiency.

[0031] It should be noted that the switching frequency This is the core operating frequency of the drive power supply. Increased on-resistance and prolonged reverse recovery time caused by PN junction degradation directly modulate the steepness of the switching edge, thus producing characteristic distortion near the switching frequency and its integer multiples. The sampling rate is set to... It is 4 to 6 times that of the switching frequency, which is sufficient to cover the fundamental, second and third harmonic components of the switching frequency. These frequency bands reflect the energy distribution of the switching transient process and are key frequency bands for capturing the early degradation of the PN junction.

[0032] The continuously output sampled data stream from the ADC is truncated into frames according to fixed time windows. Let a single time window contain... There are sampling points, and the sliding step size between adjacent time windows is . Then the sampling point is the th sampling point. The time-domain discrete sequence of a time window is defined as: In the formula, Represented as a continuous sampling sequence, represents the time frame index; n represents the sampling point index within the time window, n = 0, 1, …, N-1, where N is the number of sampling points contained in a single time window; S represents the sliding step size between adjacent time windows.

[0033] A 50% overlap rate is used between time windows, that is... ( Even numbers are used to ensure continuous tracking of fault features along the time axis, preventing features from being truncated and missed if they happen to fall exactly at the time window boundary. The number of sampling points included in the time window. The selection of frequency resolution and real-time performance should be considered. In this embodiment, the preferred frequency resolution is... This ensures that the time window contains at least several complete power grid fundamental frequency cycles, guaranteeing that the frequency resolution of the Fourier transform is sufficient to distinguish between the fundamental frequency and the switching frequency components.

[0034] It should be noted that the degradation process of the drive power supply is slow and continuous, but grid fluctuations and load changes can cause the current signal to exhibit random fluctuations ranging from seconds to minutes. A sliding time window truncation method is used to divide the continuous signal into several time segments, allowing each frame to be independently analyzed in the frequency domain, reflecting the spectral state within that short time period. A 50% overlap rate ensures sufficient information redundancy between adjacent time windows. When a frame happens to span the moment a transient feature appears, that feature will inevitably fall within the effective range of the adjacent frame, avoiding the omission of critical information due to improper truncation.

[0035] For the The time-domain discrete sequence of the nth time window is processed using the Hanning window function. The time-domain discrete sequence of the nth time window is used to suppress spectral leakage caused by finite-length truncation, and a discrete Fourier transform is performed to obtain the complex frequency domain value, which is then squared to obtain the nth value. The original energy amplitude spectrum for each time window, which characterizes the relative power intensity of each frequency component on a unit resistance.

[0036] It should be noted that the current signal of the driving power supply is non-stationary; transient changes and load fluctuations can cause the spectral energy to change at different times. Windowing operation smooths and attenuates the data at the edges of the time window, weakening the spurious high-frequency components introduced by truncation, and making the transformed spectrum more accurately reflect the frequency composition of the signal itself. The energy amplitude is obtained by taking the square of the modulus of the complex result, and the phase information is discarded while the intensity information is retained, because PN junction degradation mainly affects the relative distribution of energy at each frequency point, rather than the phase relationship.

[0037] To eliminate the significant energy interference from the fundamental and low-order harmonics of the power grid and to focus on the high-frequency band affected by the degradation of the PN junction of the switching device, this embodiment extracts the first... The original energy amplitude spectrum of each time window is observed in high-frequency ranges, discarding low-frequency bands. The start and end indices of the effective frequency band are then: In the formula, This indicates rounding to the nearest integer. Indicates the switching frequency; Indicates the fundamental frequency of the power grid; This indicates the Nyquist frequency.

[0038] Starting Index The frequency is three times the switching frequency, which serves as the lower bound of the high-frequency observation window. This is because components below this frequency mainly include the fundamental frequency, the second and third harmonics, and low-frequency interference generated by power grid load fluctuations. Their energy is much higher than that of the switching ripple and is unrelated to PN junction degradation. Termination Index Using 0.8 times the Nyquist frequency as the cutoff boundary, on the one hand, this region is far away from the aliasing risk area, and on the other hand, the skin effect and changes in distributed parameters in the high-frequency band are most sensitive to PN junction degradation in this range. At the same time, it effectively avoids the unknown parasitic oscillation noise introduced by the PCB distributed capacitance in the ultra-high frequency band.

[0039] Through the above extraction, the first... The high-frequency energy subsequence of the frame.

[0040] It should be noted that the energy of the fundamental frequency and its second and third harmonics of the power grid mainly reflects the power transmission status of the power grid itself and the power frequency distortion introduced by nonlinear loads, and is unrelated to the health status of the semiconductor devices inside the LED driver power supply. If these are included in the analysis, their huge energy amplitude will interfere with the extraction of weak high-frequency signals through subsequent statistical calculations. Choosing 3 times the switching frequency as the starting boundary ensures both complete departure from the low-frequency interference region and entry into the main energy range of the switching ripple. Terminating at 0.8 times the Nyquist frequency avoids amplitude attenuation distortion introduced by the transition band of the anti-aliasing filter at the end of the frequency axis, and also avoids the parasitic oscillation frequency band caused by PCB layout and routing, thus ensuring that the selected high-frequency window accurately corresponds to the main energy range of the transient behavior of the switching devices.

[0041] At this point, the high-frequency energy subsequence of each frame is obtained.

[0042] S002. Perform intra-frame robust normalization on the high-frequency energy sub-sequence of each frame, and then perform local neighborhood median filtering on each frequency point on the frequency axis and calculate the square value to obtain the microstructure divergence of each frequency point.

[0043] It should be noted that, due to the strong skewness of the actual current harmonic energy distribution, and the significant energy surge caused by a fault, the mean is severely skewed, leading to distorted background estimation. Therefore, robust statistics immune to outliers need to be introduced. The median and median absolute deviation are used to construct a standardized benchmark to isolate the overall impact of macroscopic operating conditions such as grid voltage fluctuations and load changes on high-frequency energy. After frame-level standardization, considering that sudden broadband electromagnetic interference often simultaneously raises the energy of multiple adjacent local frequencies, directly analyzing a single frequency point can easily lead to misjudging these interferences as fault characteristics. Therefore, local neighborhood median filtering needs to be further applied on the frequency axis to eliminate the influence of collective unidirectional rise within the neighborhood. The intermodulation distortion caused by PN junction degradation is essentially a redistribution of energy between adjacent frequencies, manifested as a small deviation of the center frequency point relative to the neighborhood background. This deviation is extremely weak at the original scale, but through nonlinear squared mapping, it can be exponentially amplified, thereby highlighting the fault characteristics from the background noise.

[0044] Based on this, this step calculates the global median and the absolute deviation of the global median for each frame's high-frequency energy subsequence, performs frame-level robust normalization, and converts the absolute energy into a dimensionless relative deviation. At each effective frequency point on the frequency axis, the local neighborhood median is taken as a local reference, and the normalized value at the center point is subtracted from this local reference to offset the collective rise of broadband EMI. The difference is then squared to obtain the microstructure divergence.

[0045] Specifically, for the first The high-frequency energy subsequence of the frame, for the first The high-frequency energy subsequence of the frame is robustly normalized to obtain the standardized first... The high-frequency energy subsequence of the frame.

[0046] It should be noted that within each frame, the energy at each frequency point is affected by two factors: first, the overall energy level is determined by the grid voltage and load current, which rises and falls synchronously across all frequencies; second, energy distortion at individual frequencies is caused by PN junction degradation. The mean value captures the overall energy center, while the median is unaffected by a few abnormal frequencies. Even if some frequencies show signs of degradation, the mean value still accurately reflects the frame's baseline energy level. The median captures the normal energy differences between frequencies under healthy conditions. The standardized [value]... The high-frequency energy subsequence of the frame expresses the degree of deviation of each frequency point from the health baseline, and has a consistent interpretative scale under different current amplitudes.

[0047] To further eliminate the collective unidirectional rise in local frequency bands caused by sudden broadband electromagnetic interference, this embodiment employs local time window mid-range filtering on the frequency axis. For the first... A valid frequency point is defined, and a local neighborhood centered on it is defined. The length is 5; the expression for calculating the divergence of the microstructure is: In the formula, The first one after the standard The high-frequency energy subsequence of the frame Microstructural divergence at each effective frequency point; The first one after the standard The high-frequency energy subsequence of the frame The energy amplitude of each effective frequency point; Indicates the first The median of five standardized residuals within a local neighborhood of an effective frequency point; it represents the collective level of that local frequency band.

[0048] It should be noted that if a broadband interference simultaneously boosts all frequencies in the neighborhood, the difference between the center point and the median of the neighborhood is very small, and after squaring, it approaches zero, thus effectively canceling out the interference. However, if PN junction degradation causes a redistribution of energy between adjacent frequencies, the center point will have a significant deviation relative to the neighborhood, and this deviation is amplified exponentially after squaring. Meanwhile, the small deviations of normal random noise approach zero after squaring, thereby greatly improving the signal-to-noise ratio of weak fault characteristics.

[0049] The amplitude of cross-modulation distortion caused by PN junction degradation is typically on the order of one-thousandth of the main peak energy, making direct detection extremely difficult; squaring operations are used for amplitudes... The deviation is subjected to nonlinear transformation: The time value is compressed (normal noise tends to zero). The time value is amplified (effective signal enhancement).

[0050] The normal random fluctuation amplitude after standardization is usually between 0.3 and 0.5, which decreases to 0.09 to 0.25 after squaring; while the deviation caused by degradation can reach 1.2 to 2.0, which increases to 1.44 to 4.0 after squaring, and the dynamic range is magnified several times.

[0051] Thus, the microstructure divergence of each effective frequency point in the high-frequency energy subsequence of each frame is obtained.

[0052] S003. The microstructure divergence of all frequency points in each frame is accumulated to obtain the total distortion energy scalar, and the total distortion energy scalar is smoothed and filtered to obtain the smoothing trend quantity.

[0053] It should be noted that directly using the standardized high-frequency energy subsequence in the high-dimensional frequency domain as an early warning criterion would lead to excessive computational load and high memory consumption in embedded systems, and sporadic spikes at individual frequency points could easily cause misjudgments. Therefore, it is necessary to accumulate and fuse the microscopic distortion features scattered across hundreds of frequency points into a single scalar, significantly reducing the computational complexity of subsequent decisions. Subsequently, an EWMA filter is introduced to extract the stationary trend term during the physical aging process of the equipment, eliminating random spikes caused by instantaneous surges in the power grid.

[0054] Based on this, this step arithmetically sums the microscopic divergences of all effective frequency points within the high-frequency band to obtain a total distortion energy scalar characterizing the overall lattice health deterioration. An EWMA filter is then used to smooth this scalar, extracting the slow aging trend and filtering out instantaneous random fluctuations.

[0055] Specifically, the calculation expression for summing all micro-divergences within the high-frequency band is as follows: In the formula, Indicates the first The total distortion energy scalar of the frame; Indicates the starting index of the effective frequency band; Indicates the valid frequency band termination index; The first one after the standard The high-frequency energy subsequence of the frame Microstructure divergence at each effective frequency point.

[0056] It should be noted that the intermodulation distortion caused by PN junction degradation does not occur in isolation at a single frequency point, but is scattered across multiple frequency points throughout the high-frequency observation window. The distortion at a single frequency point may be very weak, but by accumulating these weak distortions in the frequency domain, the energy is concentrated and aggregated into a scalar reflecting the overall degree of degradation. This accumulation operation is equivalent to integrating the fragmented fault energy sparsely distributed in the frequency domain into a single indicator sensitive to the overall lattice health of the PN junction, while compressing the data dimension from hundreds of frequency points to a single scalar, significantly reducing the complexity of subsequent calculations.

[0057] The total distortion energy scalar of each frame is smoothed using an EWMA filter, and its calculation expression is as follows: In the formula, Indicates the first The smoothing trend of the frame after EWMA filtering; Indicates the first The total distortion energy scalar of the frame; Indicates the first The smoothing trend of the frame after EWMA filtering; This represents the preset forgetting factor, which is set to 0.2 in this embodiment, meaning that the current observation has a weight of 20% and the historical trend has a weight of 80%.

[0058] When m=1, the trend value after smoothing the first frame is directly equal to the total distortion energy of the current frame, without filtering.

[0059] It should be noted that the physical and chemical aging of equipment is a slow, long-term process, with its characteristic changes occurring on a timescale of hours, days, or even months. Meanwhile, occasional transient surges in the power grid can cause random spikes in the total distortion energy scalar on a second-level scale. The EWMA filter assigns higher weights to historical data through a low forgetting factor, thus smoothing out short-term spikes while preserving the slow upward trend, which is gradually reflected in the smoothed trend value as the number of frames increases.

[0060] At this point, the smoothing trend of each frame after filtering is obtained.

[0061] S004. Set a fixed baseline threshold based on the smoothing trend amount under the initial healthy state of the equipment, and dynamically correct the fixed baseline threshold according to the current operating conditions to obtain a dynamic threshold. When the smoothing trend amount exceeds the dynamic threshold for a preset number of consecutive times, a fault is determined to have occurred and an early warning is issued.

[0062] It should be noted that the output smoothed trend sequence does not follow a normal distribution. Using the traditional 3-Sigma method to construct dynamic thresholds will lead to statistical distortion. Furthermore, if historical sliding time windows are used to dynamically update the thresholds, the smoothed trend will show a slow upward trend when the PN junction enters a slow, accelerated aging phase, causing threshold tracking drift and resulting in missed detections. Therefore, it is necessary to establish a fixed baseline threshold using the initial healthy time window during commissioning, while introducing a short-term dynamic correction coefficient to adapt to grid load fluctuations, and using time continuity criteria to eliminate single, occasional shocks.

[0063] Based on this, this step utilizes health data from the initial stage of equipment commissioning to establish a fixed baseline threshold. This threshold remains unchanged throughout the equipment's entire lifespan, fundamentally eliminating the problem of threshold drift with aging. A dynamic operating condition correction mechanism is introduced. By comparing the ratio of the median energy of the recent time window to the initial baseline, it determines whether there has been an overall shift in the current grid voltage and load current, and accordingly scales the threshold proportionally, enabling it to adapt to complex and changing power supply environments. A continuous time window counting criterion is adopted, requiring abnormal conditions to persist for multiple time windows before triggering an alarm, thus effectively distinguishing between continuous degradation and single transient impacts.

[0064] Specifically, for the filtered smoothing trend sequence, the frame-filtered smoothing trend represents the smoothing trend value of the total distortion energy of high-frequency harmonics of the LED driver power supply within each time window. The larger the value, the more severe the degradation of the PN junction.

[0065] In the initial stage of commissioning, when the equipment is powered on and operating stably, i.e., when the equipment is still in a healthy state... A time window, The preferred value is between 500 and 1000. This embodiment uses 500 time windows and includes all... The filtered and smoothed trend values ​​for each time window are denoted as the historical trend value sequence; this sequence is sorted in ascending order of value, and its 25th percentile, i.e., the lower quartile, is determined. and the 75th percentile, i.e., the upper quartile. Based on the lower quartile and upper quartile, the calculation expression for the fixed baseline threshold is as follows: In the formula, Indicates a fixed baseline threshold; The robust outlier cutoff constant is determined statistically based on the maximum fluctuation amplitude of measured health data from industrial sites. The specific method is as follows: Collect smoothed trend data over 500 consecutive time windows during the initial operation of the equipment to form a health dataset, and calculate the maximum value of this dataset. Upper quartile and interquartile range And calculate the basic coverage factor. To ensure that no false alarms occur under extreme power grid disturbances, a reliability margin factor is introduced. (Determined based on the on-site signal-to-noise ratio, when the power grid fluctuates drastically) Take 1.5, when the power grid is stable. Take 1.2), then ; Typically, the value is 3.0 to 5.0, but in this embodiment, 4.5 is preferred.

[0066] It should be noted that in the initial stage of operation Each time window represents the normal energy fluctuation range of the device under factory health conditions. and The data sequence was extracted at the lower 25% and upper 75% of its range, and the difference between these two values ​​is called the interquartile range, reflecting the dispersion of the core data. The threshold was set far above the upper limit of normal fluctuations. Compared to the conventional 1.5x and 3x interquartile ranges in box plots, the 4.5x interquartile range provides ample tolerance for various normal fluctuations that may occur in high-noise industrial environments, ensuring that false alarms are never triggered by non-fault factors during the initial commissioning of the equipment. This fixed baseline threshold remains constant throughout the entire lifecycle of the equipment and does not change with the slow increase of the filtered smoothing trend.

[0067] Considering that when the grid voltage is consistently low or the overall load current increases, the noise floor at all high-frequency points will increase proportionally, resulting in an overall increase in the smoothed trend value after filtering, even when the equipment is in a healthy state. To address these changes in operating conditions, recent... A time window, this embodiment takes The specific method for calculating the dynamic correction coefficient is as follows: The first The closest frame before The median of the smoothed trend value after filtering within each time window is denoted as the real-time median; the initial... The median of the filtered smoothed trend value for each time window is denoted as the baseline median; the ratio between the real-time median and the baseline median is denoted as the _th_ time window. The scaling factor for the frame's operating conditions; Then the first The dynamic threshold of a frame is calculated using the following expression: In the formula, Indicates the first The dynamic threshold of a frame; Indicates a fixed baseline threshold; Indicates the first The scaling factor for the frame's operating conditions; This indicates taking the maximum value.

[0068] It should be noted that the real-time median reflects the overall steady-state level of high-frequency energy under the current real-time operating conditions, which is directly affected by the effective value of the grid voltage and the amplitude of the load current. The baseline median reflects the overall high-frequency energy level under standard operating conditions during the initial operation of the equipment. The operating condition scaling factor expresses the overall energy scaling ratio of the current operating conditions relative to the factory reference.

[0069] When the scaling factor is greater than 1, it indicates that the current operating condition causes the noise floor to increase proportionally. Therefore, the fixed baseline threshold is widened proportionally to avoid false alarms caused by changes in operating conditions. When the scaling factor is less than or equal to 1, it indicates that the energy level of the current operating condition is not higher than the baseline, and the threshold remains unchanged at the fixed baseline threshold.

[0070] This mechanism ensures that the threshold only tracks the amplitude changes of external operating conditions, and not the characteristic rise caused by the aging of the equipment itself. After separating the two, the increase in the smoothing trend caused by aging will not be tracked by the threshold synchronously, thus eliminating false negatives.

[0071] Specifically, when the first The smoothing trend after frame filtering is greater than that of the first frame. When the frame reaches the dynamic threshold, increment the exception counter by 1; when the frame reaches the threshold... The smoothing trend after frame filtering is less than or equal to the first frame. When the frame reaches the dynamic threshold, the counter is reset to zero.

[0072] The preset threshold for the number of occurrences is 3. If the counter continuously accumulates to 3, it is determined to be a real fault and an abnormal alarm is triggered immediately. If the counter continuously accumulates to less than 3, no alarm is triggered, and only the event of exceeding the threshold is recorded.

[0073] The preset threshold number is negatively correlated with the sliding step size of a single time window. When the sliding step size is small (i.e., the temporal resolution is high and the overlap rate of adjacent windows is greater than 75%), the over-threshold events caused by PN junction degradation will last for a longer time span of frames, so the preset threshold number can be set to a larger value (e.g., 5-7 times). When the sliding step size is large (i.e., the independence of adjacent frames is strong), to avoid missed detections, the preset threshold number is set to a smaller value (e.g., 2-3 times). In this embodiment, at a 50% overlap rate, a preset threshold number of 3 times is preferred.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault early warning method for LED light-emitting panels based on current harmonic spectrum analysis, characterized in that, include: The current signal of the LED driver power supply is collected, and the current signal is subjected to time-domain frame segmentation and frequency-domain transformation. A high-frequency observation window is extracted to obtain the high-frequency energy subsequence of each frame. For each frame, the high-frequency energy subsequence is subjected to intra-frame robust normalization, and then local neighborhood median filtering is performed on each frequency point on the frequency axis and the square value is calculated to obtain the microstructure divergence of each frequency point. The microstructure divergence of all frequency points in each frame is accumulated to obtain the total distortion energy scalar, and the total distortion energy scalar is smoothed and filtered to obtain the smoothing trend quantity. A fixed baseline threshold is set based on the smoothing trend amount under the initial healthy state of the equipment during commissioning, and the fixed baseline threshold is dynamically corrected according to the current operating conditions to obtain a dynamic threshold. When the smoothing trend amount exceeds the dynamic threshold for a preset number of consecutive times, a fault is determined to have occurred and an early warning is issued.

2. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 1, characterized in that, The step of performing time-domain frame truncation and frequency-domain transformation on the current signal, and extracting a high-frequency observation window, includes: The current signal is segmented into frames according to a preset time window length and sliding step size to obtain a time-domain discrete sequence for each frame. A window function is applied to the time-domain discrete sequence of each frame, and then a discrete Fourier transform is performed to obtain the original energy amplitude spectrum of each frame. The range from a preset start frequency to a preset end frequency in the original energy amplitude spectrum is extracted as the high-frequency observation window to obtain the high-frequency energy subsequence of each frame.

3. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 1, characterized in that, The intra-frame robust normalization of the high-frequency energy sub-sequence for each frame includes: Calculate the global median and the absolute deviation of the global median for the high-frequency energy subsequence in the current frame; The standardization is achieved by subtracting the global median from the energy amplitude of each frequency point in the high-frequency energy subsequence of the current frame, and then dividing by the absolute deviation of the global median.

4. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 1, characterized in that, The step of performing local neighborhood median filtering and calculating the square value for each frequency point on the frequency axis includes: For each effective frequency point, a predetermined number of adjacent frequency points centered on it are selected to form a local neighborhood; the median of the standardized residuals of all frequency points in the local neighborhood is calculated as a local reference; the standardized residual of the center frequency point is subtracted from the local reference and the squared value is obtained to obtain the microstructure divergence of the center frequency point.

5. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 1, characterized in that, The step of accumulating the microstructure divergence of all frequency points within each frame includes: The total distortion energy scalar is obtained by summing the microstructure divergence of all effective frequency points within the high-frequency observation window in each frame.

6. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 1, characterized in that, The smoothing filter applied to the total distortion energy scalar includes: The total distortion energy scalar is smoothed using an exponentially weighted moving average filter; The smoothing trend amount of the current frame is the weighted sum of the total distortion energy scalar of the current frame and the smoothing trend amount of the previous frame, wherein the weight of the total distortion energy scalar of the current frame is less than the weight of the smoothing trend amount of the previous frame.

7. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 1, characterized in that, The step of setting a fixed baseline threshold based on the smoothing trend amount during the initial healthy state of the equipment includes: The smoothed trend values ​​are obtained from multiple consecutive time windows during the initial stage of equipment commissioning, forming a historical trend value sequence; Calculate the lower quartile and upper quartile of the historical trend series; The fixed baseline threshold is obtained by multiplying the difference between the upper quartile and the lower quartile by a preset constant and then adding the upper quartile.

8. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 7, characterized in that, The step of dynamically correcting the fixed baseline threshold based on the current operating conditions includes: Calculate the median of the smoothing trend amount for multiple consecutive time windows preceding the current frame, and use it as the real-time median; Calculate the median of the historical trend series as the baseline median; The ratio of the real-time median to the baseline median is used as the operating condition scaling factor; The dynamic threshold is obtained by multiplying the fixed baseline threshold by the larger value between the operating condition scaling factor and 1.

9. The LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to claim 1, characterized in that, When the smoothing trend quantity continuously exceeds the dynamic threshold for a preset number of times, a fault is determined and an early warning is issued, including: The smoothing trend value of the current frame is compared with the dynamic threshold of the current frame frame by frame; if the smoothing trend value exceeds the dynamic threshold, the anomaly count is incremented by one; if the smoothing trend value does not exceed the dynamic threshold, the anomaly count is cleared to zero; when the anomaly count reaches a preset number of times threshold, a fault is determined to have occurred and an early warning is issued.

10. An LED light-emitting disk fault early warning system based on current harmonic spectrum analysis, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the LED light-emitting disk fault early warning method based on current harmonic spectrum analysis according to any one of claims 1-9.

Images