A generator set electric fault automatic detection alarm system

Abstract

This invention discloses an automatic detection and alarm system for electrical faults in generator sets, relating to the field of fault detection technology. By constructing a multi-module collaborative system consisting of signal acquisition, energy calculation, dynamic threshold generation, spectral feature extraction, waveform morphology recognition, comprehensive judgment, and alarm response, it integrates instantaneous energy, spectral disturbance characteristics, and non-periodic waveform morphology characteristics. Employing multi-dimensional feature fusion and intelligent discrimination algorithms, it effectively solves the problems in existing technologies where high-impedance faults are easily misjudged by low-frequency interference, leading to system passivation and missed fault detection. It achieves intelligent identification and real-time alarm for high-impedance electrical faults with high sensitivity, low false alarm rate, and graded response, significantly improving the operational safety and maintenance efficiency of generator sets in complex industrial environments.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and specifically to an automatic detection and alarm system for electrical faults in generator sets. Background Technology

[0002] Automatic detection and alarm for electrical faults in generator sets refers to a system that monitors the electrical operating status of generator sets in real time through sensors and control modules, such as parameters like voltage, current, and frequency. Once an abnormality or fault (such as short circuit, overload, or undervoltage) is detected, the system will automatically identify the fault type and trigger an alarm mechanism, reminding maintenance personnel to handle the issue promptly through sound, light, or information push notifications, thereby ensuring the safe and stable operation of the equipment and preventing the fault from escalating.

[0003] In wind farms, high-impedance faults (HIFs) can occur due to poor grounding or aging cables. These faults, due to their small current, are often difficult for traditional overcurrent protection devices to detect. Existing technology uses real-time sampling of the energy values ​​of phase current and neutral current, and introduces a dynamic factor of 110% to 300% to generate an adaptive threshold. This allows the system to accurately identify abnormal energy characteristics even under scenarios with frequent load changes. For example, when sudden changes in wind speed cause rapid adjustments in the system load, this method can effectively avoid the "threshold blind zone," promptly detect potential HIFs, and trigger an alarm, thereby improving system safety and fault response capabilities.

[0004] The existing technology has the following shortcomings:

[0005] In existing technologies for high impedance fault detection, the judgment method based on current energy and dynamic threshold is effective. However, in some industrial environments (such as docks or variable frequency drive systems), the spectrum of a large number of low-frequency interference signals is highly similar to that of high impedance fault signals, causing the detection system to misjudge the interference as a fault and frequently alarm. Ultimately, maintenance personnel artificially raise the threshold to mask the real fault signal, resulting in system passivation and missed fault detection. In severe cases, it can even lead to equipment damage or fire. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic detection and alarm system for electrical faults in generator sets, in order to address the shortcomings of the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an automatic detection and alarm system for electrical faults in generator sets, comprising a signal acquisition module, an energy calculation module, a dynamic threshold generation module, a spectrum feature extraction module, a waveform morphology recognition module, a comprehensive judgment module, and an alarm module;

[0008] The signal acquisition module is used to acquire the phase current and neutral current signals of the generator set in real time.

[0009] The energy calculation module is used to perform short-time energy integration on the collected current signal to obtain the current energy value;

[0010] The dynamic threshold generation module is used to generate an adaptive energy threshold based on statistical features within the historical sampling window.

[0011] The spectrum feature extraction module is used to perform short-time Fourier transform on the current signal and extract its spectrum distribution features;

[0012] The waveform morphology recognition module is used to analyze the non-periodic morphological characteristics of current signals to assist in identifying abnormal electrical behavior;

[0013] The comprehensive judgment module is used to perform multi-dimensional analysis and judgment by combining the current energy value, spectrum distribution characteristics and non-periodic morphological characteristics to determine whether a high impedance fault exists.

[0014] The alarm module is used to generate an alarm signal when the comprehensive judgment result indicates a fault state, and send the fault type information to the remote monitoring terminal.

[0015] Preferably, the signal acquisition module includes: sampling targets including three-phase phase current and neutral current; a current transformer or Hall current sensor is configured for current acquisition, the sensor having a wide bandwidth response capability from DC to 10kHz; the sampled signal is converted by a high-resolution analog-to-digital converter, and the sampling frequency and synchronization are controlled by an embedded controller; the acquired signal is written to a ring buffer in DMA mode and uploaded to the main control system through an SPI, I2C, CAN or UART interface.

[0016] Preferably, the energy calculation module includes: segmenting the acquired signal into segments according to a fixed-length sliding window; performing low-pass or band-pass filtering on each segment to remove high-frequency interference; and performing shaping and normalization processing on the signal to calculate the energy value. ;in, : represents the current value of the i-th sample, Δt is the sampling time interval, N is the number of sampling points in the current window, and E(t) is the instantaneous energy value in the time window.

[0017] Preferably, the spectral entropy variation factor is generated after analyzing the spectral entropy changes within a fixed time period. The generation method is as follows: the sampled signal within the time period T is segmented and analyzed to obtain the spectral entropy value of each segment, thus forming a spectral entropy time series. ; each of them For a fixed window of spectral entropy, calculate the standard deviation of the spectral entropy sequence over a time period T. The expression is: ;in: Let M be the spectral entropy value of the i-th time window, and M be the total number of sampling windows. The mean of the spectral entropy; obtain Standard deviation of the spectral entropy sequence over the time period K, calculate its standard deviation from the spectral entropy sequence within time period T. The absolute value of the difference between F is used as the spectral entropy variation factor.

[0018] Preferably, after analyzing the acquired morphological entropy, a morphological entropy anomaly factor is generated. The generation method is as follows: the sampled data is segmented, the morphological entropy of each waveform segment is calculated, and a time series is formed. ; each of them This is the morphological entropy calculated in the Z-th time window; a sliding time window of length W is defined, and the mean and standard deviation of the morphological entropy within the window are calculated: In the formula, The mean of morphological entropy. Let t be the standard deviation of morphological entropy and t be time; calculate the morphological entropy anomaly factor WEAF to determine the degree of deviation of morphological entropy at the current time point: .

[0019] Preferably, the calculated instantaneous energy value, spectral entropy variation factor, and morphological entropy anomaly factor are converted into a comprehensive feature vector. The comprehensive feature vector is used as the input to the machine learning model. The feature vectors labeled in the historical dataset and the corresponding high-impedance fault labels are used as training samples. The machine learning model uses the prediction of the high-impedance electrical fault assessment value label for each set of comprehensive feature vectors as the prediction objective and the minimization of the sum of prediction errors for all high-impedance electrical fault assessment value labels as the training objective. The machine learning model is trained until the sum of prediction errors converges, at which point the model training stops. The high-impedance electrical fault assessment value is determined based on the model output. The machine learning model is a multinomial regression model.

[0020] Preferably, the obtained high-impedance electrical fault assessment value is compared with a gradient risk threshold, which includes a first risk threshold and a second risk threshold, and the first risk threshold is less than the second risk threshold. The high-impedance electrical fault assessment value is compared with the first risk threshold and the second risk threshold respectively.

[0021] If the assessment value of a high-impedance electrical fault is greater than the second risk threshold, a first-level early warning signal is immediately generated, marking it as a high-risk high-impedance electrical fault and immediately triggering the protection action.

[0022] If the assessment value of a high-impedance electrical fault is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, a level two early warning signal is immediately generated, marking it as a medium-risk electrical anomaly, recording the event, and conducting manual re-inspection or entering the tracking state.

[0023] If the high-impedance electrical fault assessment value is less than the first risk threshold, a level three warning signal is immediately generated, marking it as a low-risk or non-fault state, and maintaining the normal monitoring status.

[0024] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0025] 1. This invention constructs a multi-module intelligent detection system comprising signal acquisition, energy calculation, dynamic threshold generation, spectral feature extraction, waveform morphology recognition, comprehensive judgment, and alarm response. It integrates multi-dimensional information such as instantaneous energy, spectral disturbance characteristics, and non-periodic waveform morphology, solving the problems of high-impedance fault signals being easily confused with industrial background interference, high false alarm rates, and significant missed alarm risks in existing technologies. The introduction of spectral entropy variation factors and morphological entropy anomaly factors effectively enhances the ability to identify concealed and unstructured electrical anomalies, improving the system's sensitivity and accuracy to high-impedance faults under complex operating conditions.

[0026] 2. This invention employs a multinomial regression model to learn and evaluate fused feature vectors, outputting quantified high-impedance electrical fault assessment values. Combined with a set gradient risk threshold, it provides tiered alarms, achieving an intelligent tiered response mechanism from routine monitoring to emergency protection. This system not only boasts advantages such as strong real-time performance, high accuracy, and good scalability, but also integrates with remote monitoring systems. It is suitable for power application scenarios such as ports, mines, and power plants, and has significant practical implications for improving generator unit operational safety and the level of intelligent operation and maintenance. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0028] Figure 1 This is a mind map of the system modules of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] For examples, please refer to Figure 1As shown in the figure, the automatic detection and alarm system for electrical faults in generator sets described in this embodiment includes a signal acquisition module, an energy calculation module, a dynamic threshold generation module, a spectrum feature extraction module, a waveform morphology recognition module, a comprehensive judgment module, and an alarm module.

[0031] The signal acquisition module is used to acquire the phase current and neutral current signals of the generator set in real time.

[0032] The energy calculation module is used to perform short-time energy integration on the collected current signal to obtain the current energy value;

[0033] The dynamic threshold generation module is used to generate an adaptive energy threshold based on statistical features within the historical sampling window.

[0034] The spectrum feature extraction module is used to perform short-time Fourier transform on the current signal and extract its spectrum distribution features;

[0035] The waveform morphology recognition module is used to analyze the non-periodic morphological characteristics of current signals to assist in identifying abnormal electrical behavior;

[0036] The comprehensive judgment module is used to perform multi-dimensional analysis and judgment by combining the current energy value, spectrum distribution characteristics and non-periodic morphological characteristics to determine whether a high impedance fault exists.

[0037] The alarm module is used to generate an alarm signal when the comprehensive judgment result indicates a fault state, and send the fault type information to the remote monitoring terminal.

[0038] The signal acquisition module is a fundamental component of the system of this invention. It is primarily used for high-precision, real-time acquisition of key current signals during generator operation, providing accurate raw data support for subsequent energy calculations, spectrum analysis, and waveform recognition. This module specifically includes the following functional units and implementation details:

[0039] The sampled target signals include:

[0040] Phase current signal: including A-phase, B-phase and C-phase current, used to reflect the distribution and changes of generator output power in the three-phase line;

[0041] Neutral current signal: the unbalanced component of the three-phase current, used to identify asymmetrical faults, grounding faults, and high-impedance abnormal paths.

[0042] The selection of sensing devices includes:

[0043] Current transformers (CTs): Precision current transformers (such as 0.2 class or 0.5 class accuracy) are selected and installed in the output lines and neutral line of each phase of the generator set;

[0044] Hall current sensor: used for passive data acquisition in DC or AC hybrid systems, with good isolation and dynamic response capabilities;

[0045] The sensor should have wide bandwidth characteristics (e.g., DC~10kHz) to support spectrum analysis requirements.

[0046] Signal conditioning and filtering circuits include:

[0047] Anti-interference design: including input RC filter and common-mode rejection inductor to suppress external high-frequency electromagnetic interference;

[0048] Automatic gain control (AGC) circuit: used to dynamically adjust the signal amplitude in scenarios with large load fluctuations to ensure ADC sampling accuracy;

[0049] Protection circuits include overvoltage protection and electrostatic discharge protection (such as TVS diodes) to ensure reliable equipment operation.

[0050] Analog-to-digital conversion (ADC):

[0051] Employ a high-resolution analog-to-digital converter (such as a 16-bit or 24-bit Σ-Δ ADC) with a sampling rate of 10kHz or higher to ensure capture of transient high-frequency characteristics;

[0052] Multi-channel synchronous sampling: Ensures the time synchronization of phase current and neutral current, which is beneficial for subsequent correlation analysis and waveform comparison.

[0053] Sampling control logic:

[0054] It can be scheduled and implemented by embedded controllers (such as STM32, TI DSP, FPGA);

[0055] It supports dual sampling mechanisms: timer interrupt and event-triggered sampling (such as current surge), improving response speed and resource efficiency.

[0056] The sampling results are written to a circular buffer via DMA; data interfaces such as SPI, I2C, CAN or UART are provided to upload to the main control unit or remote monitoring system; if the system has local edge processing capabilities, the most recent H seconds of full-wave data can be cached in on-chip memory for analysis.

[0057] The energy calculation module is mainly used to perform short-time window integration processing on the current signal acquired by the signal acquisition module in order to calculate the energy intensity of the signal in real time and capture potential high-impedance fault characteristics in the electrical system.

[0058] Set a fixed-length sliding time window (e.g., 10ms, 50ms, 100ms) for energy feature extraction in the local time domain; support adaptive adjustment of the window length according to the system operating status (e.g., frequency changes, load fluctuations) to improve anomaly detection sensitivity.

[0059] The original current signal is filtered (low-pass / band-pass) to remove noise interference; the signal is shaped to maintain a stable amplitude during integration; and the sampled signal is normalized or amplified to prevent calculator saturation or underflow.

[0060] Short-time energy calculation is performed based on the following mathematical model: ;in, : represents the current value of the i-th sample, Δt is the sampling time interval, N is the number of sampling points in the current window, and E(t) is the instantaneous energy value in the time window (the unit can be normalized to V²·s or A²·s). The energy value is updated in real time using the sliding window method to ensure sufficient sensitivity to small energy fluctuations caused by high impedance faults.

[0061] The calculated energy value is stored in a local buffer for subsequent use by the dynamic threshold module or the comprehensive judgment module; the output interface can support data transmission with the main control system or judgment module via a bus (such as internal register mapping, DMA, FIFO).

[0062] The dynamic threshold generation module is used to calculate an adaptive and robust current judgment threshold based on historical energy data of the current signal within a certain time window, so as to adapt to the dynamic changes in the operating status of the generator set and avoid misjudgment or missed detection by the static threshold in high interference or load fluctuation environments.

[0063] Storing energy calculation results over a recent period often employs a sliding window cache structure; the window length can range from several seconds to several minutes (e.g., the most recent 100 energy sampling periods) to support stability statistics; the implementation can be based on a circular array, FIFO queue, or ring cache to ensure efficient data updates.

[0064] The following key statistical characteristics were calculated from the cached historical energy data:

[0065] Average value: Reflects the current steady-state energy level of the system;

[0066] Standard deviation: indicates the range of fluctuation;

[0067] Maximum / Minimum values: used to help evaluate extreme values;

[0068] Median, percentile (P90, P95) (optional): used to mitigate the influence of outliers.

[0069] The algorithm with strong anti-interference ability (such as median filtering and moving weighted average) is used to filter noise and spikes.

[0070] Set a threshold adjustment factor α and dynamically select an appropriate coefficient based on the operating scenario (e.g., automatically change based on load rate, time period, and environment mode).

[0071] A typical threshold calculation model is as follows: ;in: The dynamic energy threshold at the current time point; The average energy value within the historical window. α is the standard deviation of the historical window; α is the dynamic weighting factor, which generally ranges from 1.0 to 3.0 and can be adaptively adjusted by the configuration strategy or training model.

[0072] The output results are sent to the comprehensive judgment module as threshold data; real-time updates and periodic releases are supported, and event callback mechanisms can also be triggered (such as threshold mutations causing the detection module to re-evaluate).

[0073] The spectrum feature extraction module is used to perform short-time Fourier transform or equivalent time-frequency analysis on the acquired current signal to extract the frequency components, spectral energy distribution, characteristic frequency band amplitude and other spectral features of the signal, so as to identify the hidden electrical features of high impedance faults such as frequency anomalies and harmonic distortion, thereby effectively improving the fault identification accuracy and anti-interference capability of the system.

[0074] It receives real-time sampled data (such as neutral current or phase current with a sampling rate of 10kHz) from the signal acquisition module; it uses a ring buffer or double buffer structure to buffer a complete time-domain window of data (such as 256 points or 512 points) to ensure the integrity of the FFT input data; it supports overlapping sliding window processing, such as 50% window overlap, to improve time-frequency accuracy.

[0075] For weighted window functions of time-domain sampled signals, Hamming, Hanning, or Blackman windows are often chosen to reduce spectral leakage. The window function length is matched with the sampling rate and the characteristic frequency of the signal; for example, a 20-100 ms window width is used in the analysis of a 50 Hz 1 kHz signal.

[0076] The frequency domain transformation of the windowed signal is achieved using the Fast Fourier Transform based on the Cooley-Tukey algorithm; it supports hardware acceleration (such as DSP or FPGA) or embedded MCU implementation; the output is a complex spectrum, and the power spectral density (PSD) of each frequency band can be obtained by squaring its magnitude.

[0077] The extracted features include: fundamental and harmonic amplitudes: such as the 50Hz fundamental wave, and even / odd harmonics such as 100Hz and 150Hz; total energy value of the frequency band: such as the energy proportion of different frequency bands at 0-300Hz, 300-1000Hz; spectral centroid: measuring the degree of frequency energy concentration; spectral entropy: reflecting the randomness of the spectrum, which helps to identify unstructured anomalies; abnormal frequency point identifiers: such as the presence of abnormally strong 400Hz or 900Hz points, which are characteristics of high impedance or arc faults. The extracted spectral features are transmitted to the comprehensive judgment module in the form of structured data (such as vectors).

[0078] Spectral entropy, based on the principle of information entropy, measures the degree of "order" or "disorder" in the spectrum distribution. It can effectively identify unstructured disturbances in fault signals, such as stray frequency components generated by electric arcs, intermittent contacts, or high-impedance paths.

[0079] After analyzing the changes in spectral entropy over a fixed time period, a spectral entropy variation factor is generated. The generation method is as follows:

[0080] The sampled signal within time period T is segmented and analyzed (short-time Fourier transform) to obtain the spectral entropy value of each segment, thus forming a spectral entropy time series: ; each of them The spectral entropy corresponds to a fixed window (e.g., 100ms or 256 sampling points). A total analysis period T is defined, and the standard deviation of the spectral entropy sequence within that period is calculated. The expression is: ;in: Let M be the spectral entropy value of the i-th time window, and M be the total number of sampling windows. The mean of the spectral entropy; similarly, obtain Standard deviation of the spectral entropy sequence over the time period K, calculate its standard deviation from the spectral entropy sequence within time period T. The absolute value of the difference between F is used as the spectral entropy variation factor.

[0081] A low spectral entropy variation factor indicates a stable signal spectrum distribution, which may be in normal operation. A sudden increase in the spectral entropy variation factor indicates a drastic change in the spectral morphology, which may be a fault signal such as poor high-impedance contact, arcing, or disturbance.

[0082] The waveform morphology recognition module is used to perform time-domain morphology analysis on generator current signals. By extracting the non-periodic features of the signal, it can identify non-periodic abnormal electrical behaviors that are difficult to detect by conventional spectrum analysis.

[0083] Real-time buffering of complete waveform segments of sampled signals (e.g., every 100ms segment, sampling rate 10kHz → 1000 points); supports sliding window update mechanism for easy online identification and continuous analysis.

[0084] The following morphological feature indicators are extracted to describe the geometric structure characteristics of the waveform:

[0085] Kurtosis: Characterizes the degree of sharp pulses in a waveform; arcs and intermittent discharges often exhibit high kurtosis. Crest Factor: The ratio of the waveform's peak value to its root mean square (RMS) value; high-impedance faults often increase this ratio. RMS Value: Represents signal strength; combined with other characteristics, it helps determine energy fluctuations. Waveform Periodicity Stability (Fundamental Period Matching): Measures whether a waveform possesses periodicity, helping to distinguish between normal and aperiodic disturbances. Zero Crossover Rate (ZCR): Indicates the frequency of waveform oscillations; ZCR often increases under unstructured disturbances. Morphological Entropy: Describes the complexity of the waveform; atypical fault signals have higher entropy values.

[0086] Morphological entropy reflects the complexity and uncertainty of waveforms in the time domain. Compared with traditional features (such as RMS or zero crossover rate), it can reveal whether the waveform has irregular fluctuations and whether it contains non-periodic, asymmetric, or random mutation components. This makes it particularly sensitive to complex unstructured faults such as arc discharge, intermittent contact, and high-frequency noise superposition.

[0087] After analyzing the obtained morphological entropy, a morphological entropy anomaly factor is generated. The generation method is as follows:

[0088] The sampled data is segmented, and the morphological entropy of each waveform segment is calculated to form a time series: ; each of them It is the morphological entropy calculated in the Zth time window (e.g., 100ms).

[0089] Define a sliding time window of length W (e.g., W=10), and calculate the mean and standard deviation based on the morphological entropy within the window: In the formula, The mean of morphological entropy. Let t be the standard deviation of morphological entropy and t be time; calculate the morphological entropy anomaly factor WEAF to determine the degree of deviation of morphological entropy at the current time point: .like <1: Morphological entropy fluctuations are normal, and waveform structure is stable; if 1≤ <2: There may be a slight abnormality; it is recommended to record and observe. ≥2: This is determined to be a structurally abnormal waveform.

[0090] The comprehensive judgment module is used to receive and integrate feature values ​​output from multiple sub-modules (such as the energy calculation module, spectrum feature extraction module, waveform morphology recognition module, etc.), perform multi-dimensional analysis and judgment logic, and finally determine whether the current generator set has a high impedance electrical fault.

[0091] The calculated instantaneous energy value, spectral entropy variation factor, and morphological entropy anomaly factor are converted into a comprehensive feature vector. This comprehensive feature vector is used as the input to the machine learning model. The feature vectors labeled in the historical dataset and their corresponding high-impedance fault labels are used as training samples. The machine learning model uses the prediction of the high-impedance electrical fault assessment value label for each set of comprehensive feature vectors as the prediction objective, and minimizes the sum of prediction errors for all high-impedance electrical fault assessment value labels as the training objective. The machine learning model is trained until the sum of prediction errors converges, at which point the model training stops. The high-impedance electrical fault assessment value is determined based on the model output. The machine learning model is a multinomial regression model.

[0092] The obtained high-impedance electrical fault assessment value is compared with the gradient risk threshold, which includes a first risk threshold and a second risk threshold, and the first risk threshold is less than the second risk threshold. The high-impedance electrical fault assessment value is compared with the first risk threshold and the second risk threshold respectively.

[0093] If the assessment value of a high-impedance electrical fault is greater than the second risk threshold, a first-level early warning signal is immediately generated, marking it as a high-risk high-impedance electrical fault and immediately triggering the protection action.

[0094] If the high-impedance electrical fault assessment value is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, a level two warning signal is immediately generated, marking it as a medium-risk electrical anomaly, recording the event, and recommending manual re-inspection or entering the tracking state.

[0095] If the high-impedance electrical fault assessment value is less than the first risk threshold, a level three warning signal is immediately generated, marking it as a low-risk or non-fault state, and maintaining the normal monitoring status.

[0096] It should be noted that Level 1 warning signals are more important than Level 2 warning signals, and Level 2 warning signals are more important than Level 3 warning signals.

[0097] Based on the fault risk level, local audible and visual alarms and relay outputs are triggered to respond to the fault. At the same time, information such as fault type, assessment value, and key characteristic data are sent to the remote monitoring terminal via wired or wireless communication to realize real-time early warning and remote linkage for electrical anomalies of the generator set, ensuring the safe and efficient operation of the system.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An automatic detection and alarm system for electrical faults in generator sets, characterized in that: It includes a signal acquisition module, an energy calculation module, a dynamic threshold generation module, a spectrum feature extraction module, a waveform morphology recognition module, a comprehensive judgment module, and an alarm module; The signal acquisition module is used to acquire the phase current and neutral current signals of the generator set in real time. The energy calculation module is used to perform short-time energy integration on the collected current signal to obtain the current energy value; The dynamic threshold generation module is used to generate an adaptive energy threshold based on statistical features within the historical sampling window. The spectrum feature extraction module is used to perform short-time Fourier transform on the current signal and extract its spectrum distribution features; The waveform morphology recognition module is used to analyze the non-periodic morphological characteristics of current signals to assist in identifying abnormal electrical behavior; The comprehensive judgment module is used to perform multi-dimensional analysis and judgment by combining the current energy value, spectrum distribution characteristics and non-periodic morphological characteristics to determine whether a high impedance fault exists. The alarm module is used to generate an alarm signal when the comprehensive judgment result indicates a fault state, and send the fault type information to the remote monitoring terminal.

2. The automatic detection and alarm system for electrical faults in generator sets according to claim 1, characterized in that: The signal acquisition module includes: sampling targets including three-phase phase current and neutral current; a current transformer or Hall current sensor is configured for current acquisition, and the sensor has a wide bandwidth response capability from DC to 10kHz; the sampled signal is converted by a high-resolution analog-to-digital converter, and the sampling frequency and synchronization are controlled by an embedded controller; the acquired signal is written to a ring buffer in DMA mode and uploaded to the main control system through an SPI, I2C, CAN or UART interface.

3. The automatic detection and alarm system for electrical faults in generator sets according to claim 1, characterized in that: The energy calculation module includes: segmenting the acquired signal into segments according to a fixed-length sliding window; performing low-pass or band-pass filtering on each segment to remove high-frequency interference; and performing shaping and normalization processing on the signal to calculate the energy value. ;in, : represents the current value of the i-th sample, Δt is the sampling time interval, N is the number of sampling points in the current window, and E(t) is the instantaneous energy value in the time window.

4. The automatic detection and alarm system for electrical faults in generator sets according to claim 3, characterized in that: After analyzing the changes in spectral entropy over a fixed time period, a spectral entropy variation factor is generated. The generation method is as follows: the sampled signal within the time period T is segmented and analyzed to obtain the spectral entropy value of each segment, thus forming a spectral entropy time series. ; each of them For a fixed window of spectral entropy, calculate the standard deviation of the spectral entropy sequence over a time period T. The expression is: ;in: Let M be the spectral entropy value of the i-th time window, and M be the total number of sampling windows. The mean of the spectral entropy; obtain Standard deviation of the spectral entropy sequence over the time period K, calculate its standard deviation from the spectral entropy sequence within time period T. The absolute value of the difference between F is used as the spectral entropy variation factor.

5. The automatic detection and alarm system for electrical faults in generator sets according to claim 4, characterized in that: After analyzing the acquired morphological entropy, a morphological entropy anomaly factor is generated. The generation method is as follows: the sampled data is segmented, the morphological entropy of each waveform segment is calculated, and a time series is formed. ; each of them This is the morphological entropy calculated in the Z-th time window; a sliding time window of length W is defined, and the mean and standard deviation of the morphological entropy within the window are calculated: In the formula, The mean of morphological entropy. Let t be the standard deviation of morphological entropy, and t be time. Calculate the morphological entropy anomaly factor WEAF to determine the degree of deviation of the morphological entropy at the current time point: .

6. The automatic detection and alarm system for electrical faults in a generator set according to claim 5, characterized in that: The calculated instantaneous energy value, spectral entropy variation factor, and morphological entropy anomaly factor are converted into a comprehensive feature vector. This comprehensive feature vector is used as the input to the machine learning model. The feature vectors labeled in the historical dataset and their corresponding high-impedance fault labels are used as training samples. The machine learning model uses the prediction of the high-impedance electrical fault assessment value label for each set of comprehensive feature vectors as the prediction objective, and minimizes the sum of prediction errors for all high-impedance electrical fault assessment value labels as the training objective. The machine learning model is trained until the sum of prediction errors converges, at which point the model training stops. The high-impedance electrical fault assessment value is determined based on the model output. The machine learning model is a multinomial regression model.

7. The automatic detection and alarm system for electrical faults in generator sets according to claim 6, characterized in that: The obtained high-impedance electrical fault assessment value is compared with the gradient risk threshold, which includes a first risk threshold and a second risk threshold, and the first risk threshold is less than the second risk threshold. The high-impedance electrical fault assessment value is compared with the first risk threshold and the second risk threshold respectively. If the assessment value of a high-impedance electrical fault is greater than the second risk threshold, a first-level early warning signal is immediately generated, marking it as a high-risk high-impedance electrical fault and immediately triggering the protection action. If the assessment value of a high-impedance electrical fault is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, a level two early warning signal is immediately generated, marking it as a medium-risk electrical anomaly, recording the event, and conducting manual re-inspection or entering the tracking state. If the high-impedance electrical fault assessment value is less than the first risk threshold, a level three warning signal is immediately generated, marking it as a low-risk or non-fault state, and maintaining the normal monitoring status.

Images